; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=amdgcn | FileCheck %s -check-prefixes=GCN
; RUN: llc < %s -mtriple=amdgcn -mcpu=hawaii | FileCheck %s -check-prefixes=GFX7
; RUN: llc < %s -mtriple=amdgcn -mcpu=tonga | FileCheck %s -check-prefixes=GFX8
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx900 | FileCheck %s -check-prefixes=GFX9
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 | FileCheck %s -check-prefixes=GFX10
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 | FileCheck %s -check-prefixes=GFX11,GFX11TRUE16
; RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 | FileCheck %s -check-prefixes=GFX11,GFX11FAKE16
define void @test_load_store(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_ushort v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_short v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_short v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_ushort v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_short v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_u16 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b16 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load bfloat, ptr addrspace(1) %in
store bfloat %val, ptr addrspace(1) %out
ret void
}
define <2 x bfloat> @v_load_global_v2bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dword v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dword v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dword v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <2 x bfloat>, ptr addrspace(1) %ptr
ret <2 x bfloat> %load
}
define <3 x bfloat> @v_load_global_v3bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v3bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <3 x bfloat>, ptr addrspace(1) %ptr
ret <3 x bfloat> %load
}
define <4 x bfloat> @v_load_global_v4bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <4 x bfloat>, ptr addrspace(1) %ptr
ret <4 x bfloat> %load
}
define <6 x bfloat> @v_load_global_v6bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v6bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v6bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx3 v[3:5], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v6bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx3 v[0:2], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v6bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx3 v[0:2], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v6bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx3 v[0:2], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v6bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b96 v[0:2], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <6 x bfloat>, ptr addrspace(1) %ptr
ret <6 x bfloat> %load
}
define <8 x bfloat> @v_load_global_v8bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[0:3], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <8 x bfloat>, ptr addrspace(1) %ptr
ret <8 x bfloat> %load
}
define <16 x bfloat> @v_load_global_v16bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v5, v1
; GFX8-NEXT: v_mov_b32_e32 v4, v0
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[4:5]
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 16, v4
; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[4:5]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v9, v1
; GFX9-NEXT: v_mov_b32_e32 v8, v0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[8:9], off
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[8:9], off offset:16
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v9, v1
; GFX10-NEXT: v_mov_b32_e32 v8, v0
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[8:9], off
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[8:9], off offset:16
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_dual_mov_b32 v5, v1 :: v_dual_mov_b32 v4, v0
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b128 v[0:3], v[4:5], off
; GFX11-NEXT: global_load_b128 v[4:7], v[4:5], off offset:16
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <16 x bfloat>, ptr addrspace(1) %ptr
ret <16 x bfloat> %load
}
define <32 x bfloat> @v_load_global_v32bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_load_dwordx4 v[20:23], v[0:1], s[4:7], 0 addr64 offset:32
; GCN-NEXT: buffer_load_dwordx4 v[28:31], v[0:1], s[4:7], 0 addr64 offset:48
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_load_dwordx4 v[20:23], v[0:1], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_load_dwordx4 v[28:31], v[0:1], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX7-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX7-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX7-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX7-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX7-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v12, v0
; GFX8-NEXT: v_mov_b32_e32 v13, v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 16, v12
; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v13, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 32, v12
; GFX8-NEXT: v_addc_u32_e32 v9, vcc, 0, v13, vcc
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[12:13]
; GFX8-NEXT: v_add_u32_e32 v12, vcc, 48, v12
; GFX8-NEXT: v_addc_u32_e32 v13, vcc, 0, v13, vcc
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[4:5]
; GFX8-NEXT: flat_load_dwordx4 v[8:11], v[8:9]
; GFX8-NEXT: flat_load_dwordx4 v[12:15], v[12:13]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v17, v1
; GFX9-NEXT: v_mov_b32_e32 v16, v0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[16:17], off
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[16:17], off offset:16
; GFX9-NEXT: global_load_dwordx4 v[8:11], v[16:17], off offset:32
; GFX9-NEXT: global_load_dwordx4 v[12:15], v[16:17], off offset:48
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v17, v1
; GFX10-NEXT: v_mov_b32_e32 v16, v0
; GFX10-NEXT: s_clause 0x3
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[16:17], off
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[16:17], off offset:16
; GFX10-NEXT: global_load_dwordx4 v[8:11], v[16:17], off offset:32
; GFX10-NEXT: global_load_dwordx4 v[12:15], v[16:17], off offset:48
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_dual_mov_b32 v13, v1 :: v_dual_mov_b32 v12, v0
; GFX11-NEXT: s_clause 0x3
; GFX11-NEXT: global_load_b128 v[0:3], v[12:13], off
; GFX11-NEXT: global_load_b128 v[4:7], v[12:13], off offset:16
; GFX11-NEXT: global_load_b128 v[8:11], v[12:13], off offset:32
; GFX11-NEXT: global_load_b128 v[12:15], v[12:13], off offset:48
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <32 x bfloat>, ptr addrspace(1) %ptr
ret <32 x bfloat> %load
}
define <64 x bfloat> @v_load_global_v64bf16(ptr addrspace(1) %ptr) {
; GCN-LABEL: v_load_global_v64bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x7c, v0
; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x78, v0
; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x74, v0
; GCN-NEXT: v_add_i32_e32 v10, vcc, 0x70, v0
; GCN-NEXT: v_add_i32_e32 v11, vcc, 0x6c, v0
; GCN-NEXT: v_add_i32_e32 v12, vcc, 0x68, v0
; GCN-NEXT: v_add_i32_e32 v13, vcc, 0x64, v0
; GCN-NEXT: v_add_i32_e32 v14, vcc, 0x60, v0
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:112
; GCN-NEXT: v_add_i32_e32 v15, vcc, 0x5c, v0
; GCN-NEXT: v_add_i32_e32 v16, vcc, 0x58, v0
; GCN-NEXT: v_add_i32_e32 v17, vcc, 0x54, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v9, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v10, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:96
; GCN-NEXT: v_add_i32_e32 v7, vcc, 0x50, v0
; GCN-NEXT: v_add_i32_e32 v8, vcc, 0x4c, v0
; GCN-NEXT: v_add_i32_e32 v9, vcc, 0x48, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v6, v11, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v12, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v13, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v14, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:80
; GCN-NEXT: v_add_i32_e32 v10, vcc, 0x44, v0
; GCN-NEXT: v_add_i32_e32 v11, vcc, 64, v0
; GCN-NEXT: v_add_i32_e32 v19, vcc, 60, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v6, v15, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v16, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v17, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v7, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:64
; GCN-NEXT: v_add_i32_e32 v20, vcc, 56, v0
; GCN-NEXT: v_add_i32_e32 v21, vcc, 52, v0
; GCN-NEXT: v_add_i32_e32 v22, vcc, 48, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v6, v8, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v9, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v10, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v11, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:48
; GCN-NEXT: buffer_load_dwordx4 v[7:10], v[1:2], s[4:7], 0 addr64 offset:32
; GCN-NEXT: buffer_load_dwordx4 v[11:14], v[1:2], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[15:18], v[1:2], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: buffer_store_dword v6, v19, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v1, vcc, 44, v0
; GCN-NEXT: buffer_store_dword v5, v20, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 40, v0
; GCN-NEXT: buffer_store_dword v4, v21, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v4, vcc, 36, v0
; GCN-NEXT: buffer_store_dword v3, v22, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v3, vcc, 32, v0
; GCN-NEXT: v_add_i32_e32 v5, vcc, 28, v0
; GCN-NEXT: v_add_i32_e32 v6, vcc, 24, v0
; GCN-NEXT: v_add_i32_e32 v19, vcc, 20, v0
; GCN-NEXT: v_add_i32_e32 v20, vcc, 16, v0
; GCN-NEXT: s_waitcnt vmcnt(6)
; GCN-NEXT: buffer_store_dword v10, v1, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v1, vcc, 12, v0
; GCN-NEXT: buffer_store_dword v9, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 8, v0
; GCN-NEXT: buffer_store_dword v8, v4, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v4, vcc, 4, v0
; GCN-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(8)
; GCN-NEXT: buffer_store_dword v18, v5, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v17, v6, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v16, v19, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v15, v20, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v14, v1, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v13, v2, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v12, v4, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_load_global_v64bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:112
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x7c, v0
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x78, v0
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x74, v0
; GFX7-NEXT: v_add_i32_e32 v10, vcc, 0x70, v0
; GFX7-NEXT: v_add_i32_e32 v19, vcc, 52, v0
; GFX7-NEXT: v_add_i32_e32 v20, vcc, 48, v0
; GFX7-NEXT: v_add_i32_e32 v21, vcc, 44, v0
; GFX7-NEXT: v_add_i32_e32 v22, vcc, 40, v0
; GFX7-NEXT: v_add_i32_e32 v23, vcc, 36, v0
; GFX7-NEXT: v_add_i32_e32 v24, vcc, 32, v0
; GFX7-NEXT: v_add_i32_e32 v25, vcc, 28, v0
; GFX7-NEXT: v_add_i32_e32 v26, vcc, 24, v0
; GFX7-NEXT: v_add_i32_e32 v27, vcc, 20, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v4, v9, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v3, v10, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:96
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x6c, v0
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x68, v0
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x64, v0
; GFX7-NEXT: v_add_i32_e32 v10, vcc, 0x60, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v4, v9, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v3, v10, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:80
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x5c, v0
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x58, v0
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x54, v0
; GFX7-NEXT: v_add_i32_e32 v10, vcc, 0x50, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v4, v9, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v3, v10, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:64
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 0x4c, v0
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x48, v0
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 0x44, v0
; GFX7-NEXT: v_add_i32_e32 v10, vcc, 64, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v4, v9, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v3, v10, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dwordx4 v[3:6], v[1:2], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: buffer_load_dwordx4 v[7:10], v[1:2], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_load_dwordx4 v[11:14], v[1:2], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_load_dwordx4 v[15:18], v[1:2], s[4:7], 0 addr64
; GFX7-NEXT: v_add_i32_e32 v1, vcc, 60, v0
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 56, v0
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: buffer_store_dword v6, v1, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v1, vcc, 16, v0
; GFX7-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 12, v0
; GFX7-NEXT: buffer_store_dword v4, v19, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v4, vcc, 8, v0
; GFX7-NEXT: buffer_store_dword v3, v20, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v0
; GFX7-NEXT: s_waitcnt vmcnt(6)
; GFX7-NEXT: buffer_store_dword v10, v21, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v9, v22, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v8, v23, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v7, v24, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(9)
; GFX7-NEXT: buffer_store_dword v14, v25, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v13, v26, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v12, v27, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v11, v1, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(12)
; GFX7-NEXT: buffer_store_dword v18, v2, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v17, v4, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v16, v3, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_load_global_v64bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v28, v0
; GFX8-NEXT: v_mov_b32_e32 v29, v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 16, v28
; GFX8-NEXT: v_addc_u32_e32 v5, vcc, 0, v29, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 32, v28
; GFX8-NEXT: v_addc_u32_e32 v9, vcc, 0, v29, vcc
; GFX8-NEXT: v_add_u32_e32 v12, vcc, 48, v28
; GFX8-NEXT: v_addc_u32_e32 v13, vcc, 0, v29, vcc
; GFX8-NEXT: v_add_u32_e32 v16, vcc, 64, v28
; GFX8-NEXT: v_addc_u32_e32 v17, vcc, 0, v29, vcc
; GFX8-NEXT: s_movk_i32 s4, 0x50
; GFX8-NEXT: v_add_u32_e32 v20, vcc, s4, v28
; GFX8-NEXT: v_addc_u32_e32 v21, vcc, 0, v29, vcc
; GFX8-NEXT: s_movk_i32 s4, 0x60
; GFX8-NEXT: v_add_u32_e32 v24, vcc, s4, v28
; GFX8-NEXT: v_addc_u32_e32 v25, vcc, 0, v29, vcc
; GFX8-NEXT: s_movk_i32 s4, 0x70
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[28:29]
; GFX8-NEXT: flat_load_dwordx4 v[12:15], v[12:13]
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT: v_addc_u32_e32 v29, vcc, 0, v29, vcc
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[4:5]
; GFX8-NEXT: flat_load_dwordx4 v[8:11], v[8:9]
; GFX8-NEXT: flat_load_dwordx4 v[16:19], v[16:17]
; GFX8-NEXT: flat_load_dwordx4 v[20:23], v[20:21]
; GFX8-NEXT: flat_load_dwordx4 v[24:27], v[24:25]
; GFX8-NEXT: flat_load_dwordx4 v[28:31], v[28:29]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_load_global_v64bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v29, v1
; GFX9-NEXT: v_mov_b32_e32 v28, v0
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[28:29], off
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[28:29], off offset:16
; GFX9-NEXT: global_load_dwordx4 v[8:11], v[28:29], off offset:32
; GFX9-NEXT: global_load_dwordx4 v[12:15], v[28:29], off offset:48
; GFX9-NEXT: global_load_dwordx4 v[16:19], v[28:29], off offset:64
; GFX9-NEXT: global_load_dwordx4 v[20:23], v[28:29], off offset:80
; GFX9-NEXT: global_load_dwordx4 v[24:27], v[28:29], off offset:96
; GFX9-NEXT: s_nop 0
; GFX9-NEXT: global_load_dwordx4 v[28:31], v[28:29], off offset:112
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_load_global_v64bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v33, v1
; GFX10-NEXT: v_mov_b32_e32 v32, v0
; GFX10-NEXT: s_clause 0x7
; GFX10-NEXT: global_load_dwordx4 v[0:3], v[32:33], off
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[32:33], off offset:16
; GFX10-NEXT: global_load_dwordx4 v[8:11], v[32:33], off offset:32
; GFX10-NEXT: global_load_dwordx4 v[12:15], v[32:33], off offset:48
; GFX10-NEXT: global_load_dwordx4 v[16:19], v[32:33], off offset:64
; GFX10-NEXT: global_load_dwordx4 v[20:23], v[32:33], off offset:80
; GFX10-NEXT: global_load_dwordx4 v[24:27], v[32:33], off offset:96
; GFX10-NEXT: global_load_dwordx4 v[28:31], v[32:33], off offset:112
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_load_global_v64bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_dual_mov_b32 v29, v1 :: v_dual_mov_b32 v28, v0
; GFX11-NEXT: s_clause 0x7
; GFX11-NEXT: global_load_b128 v[0:3], v[28:29], off
; GFX11-NEXT: global_load_b128 v[4:7], v[28:29], off offset:16
; GFX11-NEXT: global_load_b128 v[8:11], v[28:29], off offset:32
; GFX11-NEXT: global_load_b128 v[12:15], v[28:29], off offset:48
; GFX11-NEXT: global_load_b128 v[16:19], v[28:29], off offset:64
; GFX11-NEXT: global_load_b128 v[20:23], v[28:29], off offset:80
; GFX11-NEXT: global_load_b128 v[24:27], v[28:29], off offset:96
; GFX11-NEXT: global_load_b128 v[28:31], v[28:29], off offset:112
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <64 x bfloat>, ptr addrspace(1) %ptr
ret <64 x bfloat> %load
}
define void @v_store_global_v2bf16(<2 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dword v[1:2], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dword v[1:2], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dword v[1:2], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b32 v[1:2], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <2 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v3bf16(<3 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GCN-NEXT: buffer_store_short v2, v[3:4], s[4:7], 0 addr64 offset:4
; GCN-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v1, v[3:4], s[4:7], 0 addr64 offset:4
; GFX7-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dword v[2:3], v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v2
; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
; GFX8-NEXT: flat_store_short v[2:3], v1
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_short v[2:3], v1, off offset:4
; GFX9-NEXT: global_store_dword v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_short v[2:3], v1, off offset:4
; GFX10-NEXT: global_store_dword v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v3bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_store_b16 v[2:3], v1, off offset:4
; GFX11-NEXT: global_store_b32 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <3 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v4bf16(<4 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v1
; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
; GCN-NEXT: v_alignbit_b32 v0, v6, v0, 16
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_alignbit_b32 v2, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_store_dwordx2 v[1:2], v[4:5], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b64 v[2:3], v[0:1], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <4 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v8bf16(<8 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v1
; GCN-NEXT: v_alignbit_b32 v3, v2, v6, 16
; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
; GCN-NEXT: v_alignbit_b32 v1, v7, v10, 16
; GCN-NEXT: v_alignbit_b32 v0, v11, v0, 16
; GCN-NEXT: buffer_store_dwordx4 v[0:3], v[8:9], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[8:9], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b128 v[4:5], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <8 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v16bf16(<16 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v15
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v1
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_alignbit_b32 v3, v7, v6, 16
; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
; GCN-NEXT: v_alignbit_b32 v1, v15, v18, 16
; GCN-NEXT: v_alignbit_b32 v0, v19, v0, 16
; GCN-NEXT: v_alignbit_b32 v7, v20, v14, 16
; GCN-NEXT: v_alignbit_b32 v6, v13, v12, 16
; GCN-NEXT: v_alignbit_b32 v5, v11, v10, 16
; GCN-NEXT: v_alignbit_b32 v4, v9, v8, 16
; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[16:17], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_store_dwordx4 v[0:3], v[16:17], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v15
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v14
; GFX7-NEXT: v_alignbit_b32 v14, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v13
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v12
; GFX7-NEXT: v_alignbit_b32 v13, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v10
; GFX7-NEXT: v_alignbit_b32 v12, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_alignbit_b32 v11, v0, v1, 16
; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
; GFX7-NEXT: buffer_store_dwordx4 v[11:14], v[16:17], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[16:17], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v8
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v9, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
; GFX9-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
; GFX10-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_store_b128 v[8:9], v[4:7], off offset:16
; GFX11-NEXT: global_store_b128 v[8:9], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <16 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v32bf16(<32 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v5
; GCN-NEXT: v_alignbit_b32 v5, v7, v6, 16
; GCN-NEXT: v_alignbit_b32 v4, v31, v4, 16
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_alignbit_b32 v3, v3, v2, 16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_alignbit_b32 v2, v1, v0, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v15
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v12
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_alignbit_b32 v13, v0, v1, 16
; GCN-NEXT: v_alignbit_b32 v12, v6, v7, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v10
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v11, v0, v1, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v8
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v10, v0, v1, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v23
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v22
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v21
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v20
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_alignbit_b32 v9, v0, v1, 16
; GCN-NEXT: v_alignbit_b32 v8, v6, v7, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v18
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v7, v0, v1, 16
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v16
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v29
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v28
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v27
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v26
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GCN-NEXT: v_alignbit_b32 v6, v0, v1, 16
; GCN-NEXT: v_alignbit_b32 v16, v16, v14, 16
; GCN-NEXT: v_alignbit_b32 v15, v15, v17, 16
; GCN-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v25
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v24
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v14, v0, v14, 16
; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v30
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dwordx4 v[6:9], v[0:1], s[4:7], 0 addr64 offset:32
; GCN-NEXT: buffer_store_dwordx4 v[10:13], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt expcnt(1)
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v17
; GCN-NEXT: v_alignbit_b32 v17, v6, v18, 16
; GCN-NEXT: buffer_store_dwordx4 v[14:17], v[0:1], s[4:7], 0 addr64 offset:48
; GCN-NEXT: buffer_store_dwordx4 v[2:5], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v3, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v2, v1, v0, 16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v14
; GFX7-NEXT: buffer_load_dword v14, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v15
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v31, 16, v5
; GFX7-NEXT: v_alignbit_b32 v5, v7, v6, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v13
; GFX7-NEXT: v_alignbit_b32 v13, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v10
; GFX7-NEXT: v_alignbit_b32 v11, v0, v1, 16
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v12
; GFX7-NEXT: v_alignbit_b32 v12, v6, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v9
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v8
; GFX7-NEXT: v_alignbit_b32 v10, v6, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v23
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v22
; GFX7-NEXT: v_alignbit_b32 v9, v6, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v21
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v20
; GFX7-NEXT: v_alignbit_b32 v8, v6, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v19
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v18
; GFX7-NEXT: v_alignbit_b32 v7, v6, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v17
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v16
; GFX7-NEXT: v_alignbit_b32 v6, v6, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v30
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v24
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_alignbit_b32 v4, v31, v4, 16
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_alignbit_b32 v17, v14, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v29
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v28
; GFX7-NEXT: v_alignbit_b32 v16, v14, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v27
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v26
; GFX7-NEXT: v_alignbit_b32 v15, v14, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v25
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_alignbit_b32 v14, v14, v18, 16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dwordx4 v[14:17], v[0:1], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: buffer_store_dwordx4 v[6:9], v[0:1], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_store_dwordx4 v[10:13], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_store_dwordx4 v[2:5], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[16:17], v[0:3]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 48, v16
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v17, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[12:15]
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 32, v16
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v17, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[8:11]
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v16
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v17, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[16:17], v[12:15], off offset:48
; GFX9-NEXT: global_store_dwordx4 v[16:17], v[8:11], off offset:32
; GFX9-NEXT: global_store_dwordx4 v[16:17], v[4:7], off offset:16
; GFX9-NEXT: global_store_dwordx4 v[16:17], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[16:17], v[12:15], off offset:48
; GFX10-NEXT: global_store_dwordx4 v[16:17], v[8:11], off offset:32
; GFX10-NEXT: global_store_dwordx4 v[16:17], v[4:7], off offset:16
; GFX10-NEXT: global_store_dwordx4 v[16:17], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x3
; GFX11-NEXT: global_store_b128 v[16:17], v[12:15], off offset:48
; GFX11-NEXT: global_store_b128 v[16:17], v[8:11], off offset:32
; GFX11-NEXT: global_store_b128 v[16:17], v[4:7], off offset:16
; GFX11-NEXT: global_store_b128 v[16:17], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <32 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @v_store_global_v64bf16(<64 x bfloat> %val, ptr addrspace(1) %ptr) {
; GCN-LABEL: v_store_global_v64bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v21
; GCN-NEXT: v_alignbit_b32 v21, v23, v22, 16
; GCN-NEXT: v_alignbit_b32 v20, v31, v20, 16
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GCN-NEXT: v_alignbit_b32 v19, v19, v18, 16
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v16
; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:136
; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:132
; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GCN-NEXT: v_alignbit_b32 v18, v18, v22, 16
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dwordx4 v[18:21], v[16:17], s[4:7], 0 addr64 offset:32
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshrrev_b32_e32 v18, 16, v13
; GCN-NEXT: v_alignbit_b32 v13, v15, v14, 16
; GCN-NEXT: v_alignbit_b32 v12, v18, v12, 16
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GCN-NEXT: v_alignbit_b32 v11, v11, v10, 16
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_alignbit_b32 v10, v9, v8, 16
; GCN-NEXT: buffer_store_dwordx4 v[10:13], v[16:17], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:128
; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:124
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:120
; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:116
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v11
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:112
; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:108
; GCN-NEXT: v_alignbit_b32 v11, v8, v9, 16
; GCN-NEXT: v_alignbit_b32 v10, v10, v12, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v14
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:104
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:100
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v9, v8, v9, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v13
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v8, v8, v12, 16
; GCN-NEXT: buffer_store_dwordx4 v[8:11], v[16:17], s[4:7], 0 addr64 offset:112
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:96
; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:92
; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:88
; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:84
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v11
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:80
; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:76
; GCN-NEXT: v_alignbit_b32 v11, v8, v9, 16
; GCN-NEXT: v_alignbit_b32 v10, v10, v12, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v14
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:72
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:68
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v9, v8, v9, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v13
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v8, v8, v12, 16
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v30
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v29
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v28
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v27
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v26
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v25
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v24
; GCN-NEXT: buffer_store_dwordx4 v[8:11], v[16:17], s[4:7], 0 addr64 offset:96
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:32
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_alignbit_b32 v3, v1, v6, 16
; GCN-NEXT: v_alignbit_b32 v2, v2, v4, 16
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:28
; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:24
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v12
; GCN-NEXT: v_alignbit_b32 v1, v1, v13, 16
; GCN-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:20
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:16
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v14
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v18
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v20
; GCN-NEXT: v_alignbit_b32 v0, v4, v0, 16
; GCN-NEXT: v_alignbit_b32 v6, v5, v19, 16
; GCN-NEXT: v_alignbit_b32 v5, v13, v21, 16
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:12
; GCN-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:8
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v22
; GCN-NEXT: v_alignbit_b32 v4, v4, v23, 16
; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:4
; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:64
; GCN-NEXT: s_waitcnt vmcnt(9)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: s_waitcnt vmcnt(8)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: s_waitcnt vmcnt(7)
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v7
; GCN-NEXT: s_waitcnt vmcnt(6)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v10
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v11
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v7
; GCN-NEXT: v_alignbit_b32 v7, v8, v15, 16
; GCN-NEXT: v_alignbit_b32 v11, v9, v20, 16
; GCN-NEXT: v_alignbit_b32 v10, v21, v10, 16
; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:60
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:56
; GCN-NEXT: s_waitcnt vmcnt(6)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v12
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v13
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v9, v8, v9, 16
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:52
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:48
; GCN-NEXT: s_waitcnt vmcnt(6)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v14
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v18
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v8, v8, v14, 16
; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:44
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v19
; GCN-NEXT: s_waitcnt vmcnt(4)
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v20
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GCN-NEXT: v_alignbit_b32 v15, v14, v15, 16
; GCN-NEXT: v_alignbit_b32 v14, v19, v12, 16
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:40
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:36
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GCN-NEXT: v_alignbit_b32 v13, v13, v18, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v19
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: v_alignbit_b32 v12, v12, v18, 16
; GCN-NEXT: buffer_store_dwordx4 v[12:15], v[16:17], s[4:7], 0 addr64 offset:80
; GCN-NEXT: buffer_store_dwordx4 v[8:11], v[16:17], s[4:7], 0 addr64 offset:64
; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[16:17], s[4:7], 0 addr64 offset:48
; GCN-NEXT: buffer_store_dwordx4 v[0:3], v[16:17], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_store_global_v64bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:128
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:120
; GFX7-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:116
; GFX7-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:112
; GFX7-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:108
; GFX7-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:104
; GFX7-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:100
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_alignbit_b32 v3, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v2, v1, v0, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: s_waitcnt vmcnt(7)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(6)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GFX7-NEXT: s_waitcnt vmcnt(5)
; GFX7-NEXT: v_mul_f32_e32 v33, 1.0, v33
; GFX7-NEXT: v_alignbit_b32 v36, v31, v32, 16
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v37
; GFX7-NEXT: v_mul_f32_e32 v34, 1.0, v34
; GFX7-NEXT: v_lshrrev_b32_e32 v33, 16, v33
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v38
; GFX7-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GFX7-NEXT: v_alignbit_b32 v35, v33, v34, 16
; GFX7-NEXT: v_alignbit_b32 v34, v31, v32, 16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v39
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v48
; GFX7-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GFX7-NEXT: v_alignbit_b32 v33, v31, v32, 16
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:132
; GFX7-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:96
; GFX7-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:92
; GFX7-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:88
; GFX7-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:84
; GFX7-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:80
; GFX7-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:76
; GFX7-NEXT: s_waitcnt vmcnt(6)
; GFX7-NEXT: buffer_store_dwordx4 v[33:36], v[31:32], s[4:7], 0 addr64 offset:112
; GFX7-NEXT: s_nop 0
; GFX7-NEXT: v_lshrrev_b32_e32 v33, 16, v5
; GFX7-NEXT: v_alignbit_b32 v5, v7, v6, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v12
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_alignbit_b32 v12, v6, v7, 16
; GFX7-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:72
; GFX7-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:68
; GFX7-NEXT: buffer_load_dword v15, off, s[0:3], s32
; GFX7-NEXT: v_alignbit_b32 v13, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v10
; GFX7-NEXT: v_alignbit_b32 v11, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v9
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v8
; GFX7-NEXT: v_alignbit_b32 v10, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v23
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v22
; GFX7-NEXT: v_alignbit_b32 v9, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v21
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v20
; GFX7-NEXT: v_alignbit_b32 v8, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v19
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v18
; GFX7-NEXT: v_alignbit_b32 v7, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(9)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v37
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v28
; GFX7-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(9)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v38
; GFX7-NEXT: v_alignbit_b32 v4, v33, v4, 16
; GFX7-NEXT: s_waitcnt vmcnt(8)
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v39
; GFX7-NEXT: v_alignbit_b32 v36, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(6)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v49
; GFX7-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v48
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(5)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v50
; GFX7-NEXT: v_alignbit_b32 v35, v18, v19, 16
; GFX7-NEXT: v_alignbit_b32 v34, v0, v1, 16
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:32
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:28
; GFX7-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:24
; GFX7-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:20
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:16
; GFX7-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:12
; GFX7-NEXT: s_waitcnt vmcnt(8)
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_alignbit_b32 v33, v6, v14, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v17
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v16
; GFX7-NEXT: v_alignbit_b32 v6, v6, v14, 16
; GFX7-NEXT: s_waitcnt vmcnt(7)
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v15
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v30
; GFX7-NEXT: buffer_store_dwordx4 v[33:36], v[31:32], s[4:7], 0 addr64 offset:96
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v29
; GFX7-NEXT: v_alignbit_b32 v17, v14, v15, 16
; GFX7-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v27
; GFX7-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v26
; GFX7-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:44
; GFX7-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:8
; GFX7-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:4
; GFX7-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:60
; GFX7-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:56
; GFX7-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:40
; GFX7-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX7-NEXT: v_alignbit_b32 v15, v14, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v25
; GFX7-NEXT: v_alignbit_b32 v16, v16, v20, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v24
; GFX7-NEXT: v_alignbit_b32 v14, v14, v20, 16
; GFX7-NEXT: s_waitcnt vmcnt(14)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_alignbit_b32 v21, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(13)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v18
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(12)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v19
; GFX7-NEXT: v_alignbit_b32 v20, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(11)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v22
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(10)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v23
; GFX7-NEXT: v_alignbit_b32 v19, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(8)
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v35
; GFX7-NEXT: s_waitcnt vmcnt(5)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v29
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v30
; GFX7-NEXT: v_alignbit_b32 v18, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v28
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v33
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v34
; GFX7-NEXT: v_alignbit_b32 v25, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v27
; GFX7-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v26
; GFX7-NEXT: v_alignbit_b32 v24, v22, v23, 16
; GFX7-NEXT: v_alignbit_b32 v23, v0, v1, 16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v36
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v37
; GFX7-NEXT: v_alignbit_b32 v22, v0, v1, 16
; GFX7-NEXT: buffer_store_dwordx4 v[22:25], v[31:32], s[4:7], 0 addr64 offset:80
; GFX7-NEXT: buffer_store_dwordx4 v[18:21], v[31:32], s[4:7], 0 addr64 offset:64
; GFX7-NEXT: buffer_store_dwordx4 v[14:17], v[31:32], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: buffer_store_dwordx4 v[6:9], v[31:32], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_store_dwordx4 v[10:13], v[31:32], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_store_dwordx4 v[2:5], v[31:32], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_store_global_v64bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX8-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX8-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX8-NEXT: s_movk_i32 s4, 0x70
; GFX8-NEXT: s_movk_i32 s5, 0x50
; GFX8-NEXT: s_waitcnt vmcnt(2)
; GFX8-NEXT: v_add_u32_e32 v34, vcc, s4, v32
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_addc_u32_e32 v35, vcc, 0, v33, vcc
; GFX8-NEXT: s_movk_i32 s4, 0x60
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[34:35], v[28:31]
; GFX8-NEXT: flat_store_dwordx4 v[32:33], v[0:3]
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v32
; GFX8-NEXT: v_addc_u32_e32 v29, vcc, 0, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, s5, v32
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 64, v32
; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v33, vcc
; GFX8-NEXT: flat_store_dwordx4 v[28:29], v[24:27]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_add_u32_e32 v24, vcc, 48, v32
; GFX8-NEXT: v_addc_u32_e32 v25, vcc, 0, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v26, vcc, 32, v32
; GFX8-NEXT: v_addc_u32_e32 v27, vcc, 0, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v28, vcc, 16, v32
; GFX8-NEXT: v_addc_u32_e32 v29, vcc, 0, v33, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[20:23]
; GFX8-NEXT: flat_store_dwordx4 v[2:3], v[16:19]
; GFX8-NEXT: flat_store_dwordx4 v[24:25], v[12:15]
; GFX8-NEXT: flat_store_dwordx4 v[26:27], v[8:11]
; GFX8-NEXT: flat_store_dwordx4 v[28:29], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_store_global_v64bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[28:31], off offset:112
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[24:27], off offset:96
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[20:23], off offset:80
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[16:19], off offset:64
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[12:15], off offset:48
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[8:11], off offset:32
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[4:7], off offset:16
; GFX9-NEXT: global_store_dwordx4 v[32:33], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_store_global_v64bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x2
; GFX10-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[28:31], off offset:112
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[24:27], off offset:96
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[20:23], off offset:80
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[16:19], off offset:64
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[12:15], off offset:48
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[8:11], off offset:32
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[4:7], off offset:16
; GFX10-NEXT: global_store_dwordx4 v[32:33], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_store_global_v64bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x2
; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT: scratch_load_b32 v31, off, s32
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_clause 0x7
; GFX11-NEXT: global_store_b128 v[32:33], v[28:31], off offset:112
; GFX11-NEXT: global_store_b128 v[32:33], v[24:27], off offset:96
; GFX11-NEXT: global_store_b128 v[32:33], v[20:23], off offset:80
; GFX11-NEXT: global_store_b128 v[32:33], v[16:19], off offset:64
; GFX11-NEXT: global_store_b128 v[32:33], v[12:15], off offset:48
; GFX11-NEXT: global_store_b128 v[32:33], v[8:11], off offset:32
; GFX11-NEXT: global_store_b128 v[32:33], v[4:7], off offset:16
; GFX11-NEXT: global_store_b128 v[32:33], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <64 x bfloat> %val, ptr addrspace(1) %ptr
ret void
}
define void @test_store_fpimm(ptr addrspace(1) %ptr0, ptr addrspace(1) %ptr1) {
; GCN-LABEL: test_store_fpimm:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mov_b32_e32 v4, 0x3f80
; GCN-NEXT: v_mov_b32_e32 v5, 0x4228
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_store_short v4, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_store_short v5, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_store_fpimm:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_mov_b32_e32 v4, 0x3f80
; GFX7-NEXT: buffer_store_short v4, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: v_mov_b32_e32 v0, 0x4228
; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_store_fpimm:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v4, 0x3f80
; GFX8-NEXT: flat_store_short v[0:1], v4
; GFX8-NEXT: v_mov_b32_e32 v0, 0x4228
; GFX8-NEXT: flat_store_short v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_store_fpimm:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v4, 0x3f80
; GFX9-NEXT: global_store_short v[0:1], v4, off
; GFX9-NEXT: v_mov_b32_e32 v0, 0x4228
; GFX9-NEXT: global_store_short v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_store_fpimm:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v4, 0x3f80
; GFX10-NEXT: v_mov_b32_e32 v5, 0x4228
; GFX10-NEXT: global_store_short v[0:1], v4, off
; GFX10-NEXT: global_store_short v[2:3], v5, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_store_fpimm:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v4, 0x3f80
; GFX11-NEXT: v_mov_b32_e32 v5, 0x4228
; GFX11-NEXT: global_store_b16 v[0:1], v4, off
; GFX11-NEXT: global_store_b16 v[2:3], v5, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store bfloat 1.0, ptr addrspace(1) %ptr0
store bfloat 42.0, ptr addrspace(1) %ptr1
ret void
}
define void @test_load_store_f32_to_bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_f32_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_f32_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_f32_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dword v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: flat_store_short v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_f32_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dword v0, v[0:1], off
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc
; GFX9-NEXT: global_store_short_d16_hi v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_f32_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dword v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo
; GFX10-NEXT: global_store_short_d16_hi v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_f32_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc_lo
; GFX11-NEXT: global_store_d16_hi_b16 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load float, ptr addrspace(1) %in
%val.bf16 = fptrunc float %val to bfloat
store bfloat %val.bf16, ptr addrspace(1) %out
ret void
}
define void @test_load_store_f64_to_bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_f64_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_f64_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_f64_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
; GFX8-NEXT: v_and_b32_e32 v7, 0x80000000, v1
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX8-NEXT: v_and_b32_e32 v8, 1, v6
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v8
; GFX8-NEXT: v_cmp_gt_f64_e64 s[4:5], |v[0:1]|, v[4:5]
; GFX8-NEXT: v_cmp_nlg_f64_e64 s[6:7], |v[0:1]|, v[4:5]
; GFX8-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[4:5]
; GFX8-NEXT: v_add_u32_e64 v4, s[4:5], v6, v4
; GFX8-NEXT: s_or_b64 vcc, s[6:7], vcc
; GFX8-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
; GFX8-NEXT: v_or_b32_e32 v5, v4, v7
; GFX8-NEXT: v_bfe_u32 v4, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v5
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: flat_store_short v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_f64_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX9-NEXT: s_brev_b32 s8, 1
; GFX9-NEXT: s_movk_i32 s9, 0x7fff
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX9-NEXT: v_and_b32_e32 v7, 1, v6
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GFX9-NEXT: v_cmp_gt_f64_e64 s[6:7], |v[0:1]|, v[4:5]
; GFX9-NEXT: v_cmp_nlg_f64_e64 s[4:5], |v[0:1]|, v[4:5]
; GFX9-NEXT: v_cndmask_b32_e64 v4, -1, 1, s[6:7]
; GFX9-NEXT: v_add_u32_e32 v4, v6, v4
; GFX9-NEXT: s_or_b64 vcc, s[4:5], vcc
; GFX9-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc
; GFX9-NEXT: v_cmp_u_f64_e32 vcc, v[0:1], v[0:1]
; GFX9-NEXT: v_and_or_b32 v5, v1, s8, v4
; GFX9-NEXT: v_bfe_u32 v4, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v5, s9
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX9-NEXT: global_store_short_d16_hi v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_f64_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX10-NEXT: v_and_b32_e32 v7, 1, v6
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX10-NEXT: v_cmp_gt_f64_e64 s5, |v[0:1]|, v[4:5]
; GFX10-NEXT: v_cmp_nlg_f64_e64 s4, |v[0:1]|, v[4:5]
; GFX10-NEXT: v_cndmask_b32_e64 v4, -1, 1, s5
; GFX10-NEXT: s_or_b32 vcc_lo, s4, vcc_lo
; GFX10-NEXT: v_add_nc_u32_e32 v4, v6, v4
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc_lo
; GFX10-NEXT: v_cmp_u_f64_e32 vcc_lo, v[0:1], v[0:1]
; GFX10-NEXT: v_and_or_b32 v5, 0x80000000, v1, v4
; GFX10-NEXT: v_bfe_u32 v4, v4, 16, 1
; GFX10-NEXT: v_add3_u32 v4, v4, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
; GFX10-NEXT: global_store_short_d16_hi v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_f64_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_cvt_f32_f64_e64 v6, |v[0:1]|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX11-NEXT: v_and_b32_e32 v7, 1, v6
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_gt_f64_e64 s1, |v[0:1]|, v[4:5]
; GFX11-NEXT: v_cmp_nlg_f64_e64 s0, |v[0:1]|, v[4:5]
; GFX11-NEXT: v_cndmask_b32_e64 v4, -1, 1, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: s_or_b32 vcc_lo, s0, vcc_lo
; GFX11-NEXT: v_add_nc_u32_e32 v4, v6, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v4, v6, vcc_lo
; GFX11-NEXT: v_cmp_u_f64_e32 vcc_lo, v[0:1], v[0:1]
; GFX11-NEXT: v_and_or_b32 v5, 0x80000000, v1, v4
; GFX11-NEXT: v_bfe_u32 v4, v4, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v4, v4, v5, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX11-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
; GFX11-NEXT: global_store_d16_hi_b16 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load double, ptr addrspace(1) %in
%val.bf16 = fptrunc double %val to bfloat
store bfloat %val.bf16, ptr addrspace(1) %out
ret void
}
define void @test_load_store_bf16_to_f32(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_bf16_to_f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_bf16_to_f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_bf16_to_f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_ushort v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: flat_store_dword v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_bf16_to_f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: global_store_dword v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_bf16_to_f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_ushort v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: global_store_dword v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_bf16_to_f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_u16 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: global_store_b32 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load bfloat, ptr addrspace(1) %in
%val.f32 = fpext bfloat %val to float
store float %val.f32, ptr addrspace(1) %out
ret void
}
define void @test_load_store_bf16_to_f64(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_bf16_to_f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_bf16_to_f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX7-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_bf16_to_f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_ushort v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_bf16_to_f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_bf16_to_f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_ushort v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_bf16_to_f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_u16 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: global_store_b64 v[2:3], v[0:1], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load bfloat, ptr addrspace(1) %in
%val.f64 = fpext bfloat %val to double
store double %val.f64, ptr addrspace(1) %out
ret void
}
define void @test_load_store_v2bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dword v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_dword v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dword v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_dword v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dword v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dword v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b32 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load <2 x bfloat>, ptr addrspace(1) %in
store <2 x bfloat> %val, ptr addrspace(1) %out
ret void
}
define void @test_load_store_v4bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dwordx2 v[0:1], v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b64 v[2:3], v[0:1], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load <4 x bfloat>, ptr addrspace(1) %in
store <4 x bfloat> %val, ptr addrspace(1) %out
ret void
}
define void @test_load_store_v8bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[2:3], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[4:7], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b128 v[2:3], v[4:7], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load <8 x bfloat>, ptr addrspace(1) %in
store <8 x bfloat> %val, ptr addrspace(1) %out
ret void
}
define void @test_load_store_v16bf16(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_load_store_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_load_dwordx4 v[8:11], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: buffer_store_dwordx4 v[8:11], v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_load_store_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_load_dwordx4 v[8:11], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: buffer_store_dwordx4 v[4:7], v[2:3], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: buffer_store_dwordx4 v[8:11], v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_load_store_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 16, v0
; GFX8-NEXT: v_addc_u32_e32 v9, vcc, 0, v1, vcc
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; GFX8-NEXT: flat_load_dwordx4 v[8:11], v[8:9]
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v2
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v3, vcc
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: flat_store_dwordx4 v[2:3], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[8:11]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_load_store_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off offset:16
; GFX9-NEXT: global_load_dwordx4 v[8:11], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:16
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: global_store_dwordx4 v[2:3], v[8:11], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_load_store_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off offset:16
; GFX10-NEXT: global_load_dwordx4 v[8:11], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: global_store_dwordx4 v[2:3], v[4:7], off offset:16
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[2:3], v[8:11], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_load_store_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b128 v[4:7], v[0:1], off offset:16
; GFX11-NEXT: global_load_b128 v[8:11], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: global_store_b128 v[2:3], v[4:7], off offset:16
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b128 v[2:3], v[8:11], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load <16 x bfloat>, ptr addrspace(1) %in
store <16 x bfloat> %val, ptr addrspace(1) %out
ret void
}
define void @test_arg_store(bfloat %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_short v0, v[1:2], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v0, v[1:2], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_short v[1:2], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_short v[1:2], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_short v[1:2], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b16 v[1:2], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store bfloat %in, ptr addrspace(1) %out
ret void
}
define void @test_arg_store_v2bf16(<2 x bfloat> %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_store_dword v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dword v[1:2], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dword v[1:2], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dword v[1:2], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b32 v[1:2], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <2 x bfloat> %in, ptr addrspace(1) %out
ret void
}
define void @test_arg_store_v3bf16(<3 x bfloat> %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GCN-NEXT: buffer_store_short v2, v[3:4], s[4:7], 0 addr64 offset:4
; GCN-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v1, v[3:4], s[4:7], 0 addr64 offset:4
; GFX7-NEXT: buffer_store_dword v0, v[3:4], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dword v[2:3], v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v2
; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v3, vcc
; GFX8-NEXT: flat_store_short v[2:3], v1
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_short v[2:3], v1, off offset:4
; GFX9-NEXT: global_store_dword v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_short v[2:3], v1, off offset:4
; GFX10-NEXT: global_store_dword v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store_v3bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_store_b16 v[2:3], v1, off offset:4
; GFX11-NEXT: global_store_b32 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <3 x bfloat> %in, ptr addrspace(1) %out
ret void
}
define void @test_arg_store_v4bf16(<4 x bfloat> %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v1
; GCN-NEXT: v_alignbit_b32 v1, v3, v2, 16
; GCN-NEXT: v_alignbit_b32 v0, v6, v0, 16
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_store_dwordx2 v[0:1], v[4:5], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_alignbit_b32 v2, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v1, v1, v0, 16
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_store_dwordx2 v[1:2], v[4:5], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx2 v[2:3], v[0:1], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b64 v[2:3], v[0:1], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <4 x bfloat> %in, ptr addrspace(1) %out
ret void
}
define void @test_arg_store_v8bf16(<8 x bfloat> %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v1
; GCN-NEXT: v_alignbit_b32 v3, v2, v6, 16
; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
; GCN-NEXT: v_alignbit_b32 v1, v7, v10, 16
; GCN-NEXT: v_alignbit_b32 v0, v11, v0, 16
; GCN-NEXT: buffer_store_dwordx4 v[0:3], v[8:9], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[8:9], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[4:5], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_store_b128 v[4:5], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <8 x bfloat> %in, ptr addrspace(1) %out
ret void
}
define void @test_arg_store_v16bf16(<16 x bfloat> %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_arg_store_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v15
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v1
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_alignbit_b32 v3, v7, v6, 16
; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
; GCN-NEXT: v_alignbit_b32 v1, v15, v18, 16
; GCN-NEXT: v_alignbit_b32 v0, v19, v0, 16
; GCN-NEXT: v_alignbit_b32 v7, v20, v14, 16
; GCN-NEXT: v_alignbit_b32 v6, v13, v12, 16
; GCN-NEXT: v_alignbit_b32 v5, v11, v10, 16
; GCN-NEXT: v_alignbit_b32 v4, v9, v8, 16
; GCN-NEXT: buffer_store_dwordx4 v[4:7], v[16:17], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_store_dwordx4 v[0:3], v[16:17], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_arg_store_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v5, v5, v4, 16
; GFX7-NEXT: v_alignbit_b32 v4, v3, v2, 16
; GFX7-NEXT: v_alignbit_b32 v3, v1, v0, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v15
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v14
; GFX7-NEXT: v_alignbit_b32 v14, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v13
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v12
; GFX7-NEXT: v_alignbit_b32 v13, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v10
; GFX7-NEXT: v_alignbit_b32 v12, v0, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: v_alignbit_b32 v11, v0, v1, 16
; GFX7-NEXT: v_alignbit_b32 v6, v7, v6, 16
; GFX7-NEXT: buffer_store_dwordx4 v[11:14], v[16:17], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_store_dwordx4 v[3:6], v[16:17], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_arg_store_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
; GFX8-NEXT: s_nop 0
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v8
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v9, vcc
; GFX8-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_arg_store_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
; GFX9-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_arg_store_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_store_dwordx4 v[8:9], v[4:7], off offset:16
; GFX10-NEXT: global_store_dwordx4 v[8:9], v[0:3], off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_arg_store_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_store_b128 v[8:9], v[4:7], off offset:16
; GFX11-NEXT: global_store_b128 v[8:9], v[0:3], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store <16 x bfloat> %in, ptr addrspace(1) %out
ret void
}
define amdgpu_gfx void @test_inreg_arg_store(bfloat inreg %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_inreg_arg_store:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s39, 0xf000
; GCN-NEXT: s_mov_b32 s38, 0
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s4
; GCN-NEXT: s_mov_b32 s36, s38
; GCN-NEXT: s_mov_b32 s37, s38
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: buffer_store_short v2, v[0:1], s[36:39], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_inreg_arg_store:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s38, 0
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s4
; GFX7-NEXT: s_mov_b32 s39, 0xf000
; GFX7-NEXT: s_mov_b32 s36, s38
; GFX7-NEXT: s_mov_b32 s37, s38
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: buffer_store_short v2, v[0:1], s[36:39], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_inreg_arg_store:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v2, s4
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_inreg_arg_store:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v2, s4
; GFX9-NEXT: global_store_short v[0:1], v2, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_inreg_arg_store:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v2, s4
; GFX10-NEXT: global_store_short v[0:1], v2, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_inreg_arg_store:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v2, s4
; GFX11-NEXT: global_store_b16 v[0:1], v2, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store bfloat %in, ptr addrspace(1) %out
ret void
}
define bfloat @test_byval(ptr addrspace(5) byval(bfloat) %bv, bfloat %val) {
; GCN-LABEL: test_byval:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: buffer_store_short v1, off, s[0:3], s32
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_byval:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v1, off, s[0:3], s32
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_byval:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_byval:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_byval:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_byval:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_store_b16 off, v0, s32
; GFX11-NEXT: s_setpc_b64 s[30:31]
store bfloat %val, ptr addrspace(5) %bv
%retval = load bfloat, ptr addrspace(5) %bv
ret bfloat %retval
}
define void @test_sret(ptr addrspace(5) sret(bfloat) %sret, bfloat %val) {
; GCN-LABEL: test_sret:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_sret:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_sret:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_sret:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_sret:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_sret:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_store_b16 v0, v1, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
store bfloat %val, ptr addrspace(5) %sret
ret void
}
define void @test_bitcast_from_bfloat(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; GCN-LABEL: test_bitcast_from_bfloat:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_bitcast_from_bfloat:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v0, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_bitcast_from_bfloat:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_ushort v0, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_short v[2:3], v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_bitcast_from_bfloat:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v0, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_short v[2:3], v0, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_bitcast_from_bfloat:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_ushort v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_short v[2:3], v0, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_bitcast_from_bfloat:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_u16 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b16 v[2:3], v0, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load bfloat, ptr addrspace(1) %in
%val_int = bitcast bfloat %val to i16
store i16 %val_int, ptr addrspace(1) %out
ret void
}
define void @test_bitcast_to_bfloat(ptr addrspace(1) %out, ptr addrspace(1) %in) {
; GCN-LABEL: test_bitcast_to_bfloat:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v2, v[2:3], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v2, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_bitcast_to_bfloat:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v2, v[2:3], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v2, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_bitcast_to_bfloat:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_ushort v2, v[2:3]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: flat_store_short v[0:1], v2
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_bitcast_to_bfloat:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v2, v[2:3], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: global_store_short v[0:1], v2, off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_bitcast_to_bfloat:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_ushort v2, v[2:3], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: global_store_short v[0:1], v2, off
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_bitcast_to_bfloat:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_u16 v2, v[2:3], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: global_store_b16 v[0:1], v2, off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%val = load i16, ptr addrspace(1) %in
%val_fp = bitcast i16 %val to bfloat
store bfloat %val_fp, ptr addrspace(1) %out
ret void
}
define bfloat @test_ret(bfloat %in) {
; GCN-LABEL: test_ret:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret bfloat %in
}
define <2 x bfloat> @test_ret_v2bf16(<2 x bfloat> %in) {
; GCN-LABEL: test_ret_v2bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret_v2bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret_v2bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret_v2bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret_v2bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret_v2bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret <2 x bfloat> %in
}
define <3 x bfloat> @test_ret_v3bf16(<3 x bfloat> %in) {
; GCN-LABEL: test_ret_v3bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret_v3bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret_v3bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret_v3bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret_v3bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret_v3bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret <3 x bfloat> %in
}
define <4 x bfloat> @test_ret_v4bf16(<4 x bfloat> %in) {
; GCN-LABEL: test_ret_v4bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret_v4bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret_v4bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret_v4bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret_v4bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret_v4bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret <4 x bfloat> %in
}
define <8 x bfloat> @test_ret_v8bf16(<8 x bfloat> %in) {
; GCN-LABEL: test_ret_v8bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret_v8bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret_v8bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret_v8bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret_v8bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret_v8bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret <8 x bfloat> %in
}
define <16 x bfloat> @test_ret_v16bf16(<16 x bfloat> %in) {
; GCN-LABEL: test_ret_v16bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_ret_v16bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_ret_v16bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_ret_v16bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_ret_v16bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_ret_v16bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
ret <16 x bfloat> %in
}
define void @test_call(bfloat %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v2, s30, 0
; GCN-NEXT: v_writelane_b32 v2, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v2, 1
; GCN-NEXT: v_readlane_b32 s30, v2, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v2, s30, 0
; GFX7-NEXT: v_writelane_b32 v2, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v2, 1
; GFX7-NEXT: v_readlane_b32 s30, v2, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v2, s30, 0
; GFX8-NEXT: v_writelane_b32 v2, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v2, 1
; GFX8-NEXT: v_readlane_b32 s30, v2, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v2, s30, 0
; GFX9-NEXT: v_writelane_b32 v2, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v2, 1
; GFX9-NEXT: v_readlane_b32 s30, v2, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v2, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v2, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v2, 1
; GFX10-NEXT: v_readlane_b32 s30, v2, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v2, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v2, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v2, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b16 v1, v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v2, 1
; GFX11-NEXT: v_readlane_b32 s30, v2, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v2, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call bfloat @test_arg_store(bfloat %in)
store volatile bfloat %result, ptr addrspace(5) %out
ret void
}
define void @test_call_v2bf16(<2 x bfloat> %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call_v2bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v4, s30, 0
; GCN-NEXT: v_writelane_b32 v4, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_add_i32_e32 v3, vcc, 2, v2
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: buffer_store_short v1, v3, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v4, 1
; GCN-NEXT: v_readlane_b32 s30, v4, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v2bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v4, s30, 0
; GFX7-NEXT: v_writelane_b32 v4, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 2, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v1, v3, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v4, 1
; GFX7-NEXT: v_readlane_b32 s30, v4, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v2bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v2, s30, 0
; GFX8-NEXT: v_writelane_b32 v2, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v2, 1
; GFX8-NEXT: v_readlane_b32 s30, v2, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v2bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v2, s30, 0
; GFX9-NEXT: v_writelane_b32 v2, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v2, 1
; GFX9-NEXT: v_readlane_b32 s30, v2, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v2bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v2, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v2, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v2, 1
; GFX10-NEXT: v_readlane_b32 s30, v2, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call_v2bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v2, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v2, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v2, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b32 v1, v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v2, 1
; GFX11-NEXT: v_readlane_b32 s30, v2, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v2, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call <2 x bfloat> @test_arg_store_v2bf16(<2 x bfloat> %in)
store volatile <2 x bfloat> %result, ptr addrspace(5) %out
ret void
}
define void @test_call_v3bf16(<3 x bfloat> %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call_v3bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v5, s30, 0
; GCN-NEXT: v_writelane_b32 v5, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_add_i32_e32 v4, vcc, 4, v3
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GCN-NEXT: buffer_store_short v2, v4, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_dword v0, v3, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v5, 1
; GCN-NEXT: v_readlane_b32 s30, v5, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v3bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v4, s30, 0
; GFX7-NEXT: v_writelane_b32 v4, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 4, v3
; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v0, v3, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v4, 1
; GFX7-NEXT: v_readlane_b32 s30, v4, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v3bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v4, s30, 0
; GFX8-NEXT: v_writelane_b32 v4, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 4, v2
; GFX8-NEXT: buffer_store_short v1, v3, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v4, 1
; GFX8-NEXT: v_readlane_b32 s30, v4, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v3bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v3, s30, 0
; GFX9-NEXT: v_writelane_b32 v3, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v3, 1
; GFX9-NEXT: v_readlane_b32 s30, v3, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v3bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v3, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v3, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen offset:4
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v3, 1
; GFX10-NEXT: v_readlane_b32 s30, v3, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call_v3bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v3, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v3, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v3, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b16 v2, v1, off offset:4 dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: scratch_store_b32 v2, v0, off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v3, 1
; GFX11-NEXT: v_readlane_b32 s30, v3, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v3, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call <3 x bfloat> @test_arg_store_v2bf16(<3 x bfloat> %in)
store volatile <3 x bfloat> %result, ptr addrspace(5) %out
ret void
}
define void @test_call_v4bf16(<4 x bfloat> %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call_v4bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v8, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v8, s30, 0
; GCN-NEXT: v_writelane_b32 v8, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_add_i32_e32 v5, vcc, 6, v4
; GCN-NEXT: v_add_i32_e32 v6, vcc, 4, v4
; GCN-NEXT: v_add_i32_e32 v7, vcc, 2, v4
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: buffer_store_short v3, v5, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v2, v6, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v1, v7, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v8, 1
; GCN-NEXT: v_readlane_b32 s30, v8, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v4bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v6, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v6, s30, 0
; GFX7-NEXT: v_writelane_b32 v6, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 6, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: buffer_store_short v3, v5, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v4
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v4, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v6, 1
; GFX7-NEXT: v_readlane_b32 s30, v6, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v6, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v4bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v4, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v4, s30, 0
; GFX8-NEXT: v_writelane_b32 v4, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 4, v2
; GFX8-NEXT: buffer_store_dword v1, v3, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v4, 1
; GFX8-NEXT: v_readlane_b32 s30, v4, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v4, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v4bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v3, s30, 0
; GFX9-NEXT: v_writelane_b32 v3, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen offset:4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v3, 1
; GFX9-NEXT: v_readlane_b32 s30, v3, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v4bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v3, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v3, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v3, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen offset:4
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v0, v2, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v3, 1
; GFX10-NEXT: v_readlane_b32 s30, v3, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v3, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call_v4bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v3, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v3, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v3, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b64 v2, v[0:1], off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v3, 1
; GFX11-NEXT: v_readlane_b32 s30, v3, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v3, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call <4 x bfloat> @test_arg_store_v2bf16(<4 x bfloat> %in)
store volatile <4 x bfloat> %result, ptr addrspace(5) %out
ret void
}
define void @test_call_v8bf16(<8 x bfloat> %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call_v8bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v16, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v16, s30, 0
; GCN-NEXT: v_writelane_b32 v16, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_add_i32_e32 v9, vcc, 14, v8
; GCN-NEXT: v_add_i32_e32 v10, vcc, 12, v8
; GCN-NEXT: v_add_i32_e32 v11, vcc, 10, v8
; GCN-NEXT: v_add_i32_e32 v12, vcc, 8, v8
; GCN-NEXT: v_add_i32_e32 v13, vcc, 6, v8
; GCN-NEXT: v_add_i32_e32 v14, vcc, 4, v8
; GCN-NEXT: v_add_i32_e32 v15, vcc, 2, v8
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: buffer_store_short v7, v9, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v6, v10, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v5, v11, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v4, v12, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v3, v13, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v2, v14, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v1, v15, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v16, 1
; GCN-NEXT: v_readlane_b32 s30, v16, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v8bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v10, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v10, s30, 0
; GFX7-NEXT: v_writelane_b32 v10, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 14, v8
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: buffer_store_short v7, v9, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 12, v8
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: buffer_store_short v6, v7, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v6, vcc, 10, v8
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: buffer_store_short v5, v6, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 8, v8
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: buffer_store_short v4, v5, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v4, vcc, 6, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v8
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v8, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v10, 1
; GFX7-NEXT: v_readlane_b32 s30, v10, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v10, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v8bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v6, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v6, s30, 0
; GFX8-NEXT: v_writelane_b32 v6, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 12, v4
; GFX8-NEXT: buffer_store_dword v3, v5, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 8, v4
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v4
; GFX8-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v0, v4, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v6, 1
; GFX8-NEXT: v_readlane_b32 s30, v6, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v6, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v8bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v5, s30, 0
; GFX9-NEXT: v_writelane_b32 v5, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen offset:12
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen offset:8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen offset:4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v0, v4, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v5, 1
; GFX9-NEXT: v_readlane_b32 s30, v5, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v8bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v5, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v5, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v5, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen offset:12
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen offset:8
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen offset:4
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v0, v4, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v5, 1
; GFX10-NEXT: v_readlane_b32 s30, v5, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v5, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call_v8bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v5, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v5, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v5, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b128 v4, v[0:3], off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v5, 1
; GFX11-NEXT: v_readlane_b32 s30, v5, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v5, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call <8 x bfloat> @test_arg_store_v2bf16(<8 x bfloat> %in)
store volatile <8 x bfloat> %result, ptr addrspace(5) %out
ret void
}
define void @test_call_v16bf16(<16 x bfloat> %in, ptr addrspace(5) %out) {
; GCN-LABEL: test_call_v16bf16:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s18, s33
; GCN-NEXT: s_mov_b32 s33, s32
; GCN-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GCN-NEXT: buffer_store_dword v21, off, s[0:3], s33 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[16:17]
; GCN-NEXT: s_addk_i32 s32, 0x400
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v21, s30, 0
; GCN-NEXT: v_writelane_b32 v21, s31, 1
; GCN-NEXT: s_getpc_b64 s[16:17]
; GCN-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GCN-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GCN-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GCN-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_add_i32_e32 v17, vcc, 30, v16
; GCN-NEXT: v_add_i32_e32 v18, vcc, 28, v16
; GCN-NEXT: v_add_i32_e32 v19, vcc, 26, v16
; GCN-NEXT: v_add_i32_e32 v20, vcc, 24, v16
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GCN-NEXT: buffer_store_short v15, v17, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v15, vcc, 22, v16
; GCN-NEXT: v_add_i32_e32 v17, vcc, 20, v16
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: buffer_store_short v14, v18, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v14, vcc, 18, v16
; GCN-NEXT: v_add_i32_e32 v18, vcc, 16, v16
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GCN-NEXT: buffer_store_short v13, v19, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v13, vcc, 14, v16
; GCN-NEXT: v_add_i32_e32 v19, vcc, 12, v16
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: buffer_store_short v12, v20, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v12, vcc, 10, v16
; GCN-NEXT: v_add_i32_e32 v20, vcc, 8, v16
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GCN-NEXT: buffer_store_short v11, v15, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v11, vcc, 6, v16
; GCN-NEXT: v_add_i32_e32 v15, vcc, 4, v16
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: buffer_store_short v10, v17, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: v_add_i32_e32 v10, vcc, 2, v16
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: buffer_store_short v9, v14, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v8, v18, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v7, v13, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v6, v19, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v5, v12, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v4, v20, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v3, v11, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v2, v15, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v1, v10, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: buffer_store_short v0, v16, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_readlane_b32 s31, v21, 1
; GCN-NEXT: v_readlane_b32 s30, v21, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s33 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_addk_i32 s32, 0xfc00
; GCN-NEXT: s_mov_b32 s33, s18
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_call_v16bf16:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s18, s33
; GFX7-NEXT: s_mov_b32 s33, s32
; GFX7-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX7-NEXT: buffer_store_dword v18, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[16:17]
; GFX7-NEXT: s_addk_i32 s32, 0x400
; GFX7-NEXT: s_getpc_b64 s[16:17]
; GFX7-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX7-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX7-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX7-NEXT: v_writelane_b32 v18, s30, 0
; GFX7-NEXT: v_writelane_b32 v18, s31, 1
; GFX7-NEXT: s_waitcnt lgkmcnt(0)
; GFX7-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX7-NEXT: v_add_i32_e32 v17, vcc, 30, v16
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: buffer_store_short v15, v17, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v15, vcc, 28, v16
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX7-NEXT: buffer_store_short v14, v15, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v14, vcc, 26, v16
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX7-NEXT: buffer_store_short v13, v14, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v13, vcc, 24, v16
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX7-NEXT: buffer_store_short v12, v13, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v12, vcc, 22, v16
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX7-NEXT: buffer_store_short v11, v12, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v11, vcc, 20, v16
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX7-NEXT: buffer_store_short v10, v11, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v10, vcc, 18, v16
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: buffer_store_short v9, v10, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 16, v16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: buffer_store_short v8, v9, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 14, v16
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: buffer_store_short v7, v8, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 12, v16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: buffer_store_short v6, v7, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v6, vcc, 10, v16
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: buffer_store_short v5, v6, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 8, v16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: buffer_store_short v4, v5, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v4, vcc, 6, v16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: buffer_store_short v3, v4, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 4, v16
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_short v2, v3, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 2, v16
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v1, v2, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_short v0, v16, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_readlane_b32 s31, v18, 1
; GFX7-NEXT: v_readlane_b32 s30, v18, 0
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v18, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_addk_i32 s32, 0xfc00
; GFX7-NEXT: s_mov_b32 s33, s18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_call_v16bf16:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_mov_b32 s18, s33
; GFX8-NEXT: s_mov_b32 s33, s32
; GFX8-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX8-NEXT: buffer_store_dword v10, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[16:17]
; GFX8-NEXT: s_addk_i32 s32, 0x400
; GFX8-NEXT: s_getpc_b64 s[16:17]
; GFX8-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX8-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX8-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX8-NEXT: v_writelane_b32 v10, s30, 0
; GFX8-NEXT: v_writelane_b32 v10, s31, 1
; GFX8-NEXT: s_waitcnt lgkmcnt(0)
; GFX8-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 28, v8
; GFX8-NEXT: buffer_store_dword v7, v9, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 24, v8
; GFX8-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v6, vcc, 20, v8
; GFX8-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 16, v8
; GFX8-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 12, v8
; GFX8-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 8, v8
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v8
; GFX8-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v0, v8, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_readlane_b32 s31, v10, 1
; GFX8-NEXT: v_readlane_b32 s30, v10, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v10, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_addk_i32 s32, 0xfc00
; GFX8-NEXT: s_mov_b32 s33, s18
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_call_v16bf16:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_mov_b32 s18, s33
; GFX9-NEXT: s_mov_b32 s33, s32
; GFX9-NEXT: s_xor_saveexec_b64 s[16:17], -1
; GFX9-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[16:17]
; GFX9-NEXT: s_addk_i32 s32, 0x400
; GFX9-NEXT: s_getpc_b64 s[16:17]
; GFX9-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX9-NEXT: v_writelane_b32 v9, s30, 0
; GFX9-NEXT: v_writelane_b32 v9, s31, 1
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: buffer_store_dword v7, v8, s[0:3], 0 offen offset:28
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v6, v8, s[0:3], 0 offen offset:24
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen offset:20
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v4, v8, s[0:3], 0 offen offset:16
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v3, v8, s[0:3], 0 offen offset:12
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v2, v8, s[0:3], 0 offen offset:8
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v1, v8, s[0:3], 0 offen offset:4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_store_dword v0, v8, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_readlane_b32 s31, v9, 1
; GFX9-NEXT: v_readlane_b32 s30, v9, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_addk_i32 s32, 0xfc00
; GFX9-NEXT: s_mov_b32 s33, s18
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_call_v16bf16:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_mov_b32 s18, s33
; GFX10-NEXT: s_mov_b32 s33, s32
; GFX10-NEXT: s_xor_saveexec_b32 s16, -1
; GFX10-NEXT: buffer_store_dword v9, off, s[0:3], s33 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s16
; GFX10-NEXT: s_addk_i32 s32, 0x200
; GFX10-NEXT: s_getpc_b64 s[16:17]
; GFX10-NEXT: s_add_u32 s16, s16, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s17, s17, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX10-NEXT: v_writelane_b32 v9, s30, 0
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[16:17], 0x0
; GFX10-NEXT: v_writelane_b32 v9, s31, 1
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: buffer_store_dword v7, v8, s[0:3], 0 offen offset:28
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v6, v8, s[0:3], 0 offen offset:24
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v5, v8, s[0:3], 0 offen offset:20
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v4, v8, s[0:3], 0 offen offset:16
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v3, v8, s[0:3], 0 offen offset:12
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v2, v8, s[0:3], 0 offen offset:8
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v1, v8, s[0:3], 0 offen offset:4
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_store_dword v0, v8, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: v_readlane_b32 s31, v9, 1
; GFX10-NEXT: v_readlane_b32 s30, v9, 0
; GFX10-NEXT: s_xor_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v9, off, s[0:3], s33 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_addk_i32 s32, 0xfe00
; GFX10-NEXT: s_mov_b32 s33, s18
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_call_v16bf16:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_mov_b32 s2, s33
; GFX11-NEXT: s_mov_b32 s33, s32
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_store_b32 off, v9, s33 ; 4-byte Folded Spill
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, 16
; GFX11-NEXT: s_getpc_b64 s[0:1]
; GFX11-NEXT: s_add_u32 s0, s0, test_arg_store_v2bf16@gotpcrel32@lo+4
; GFX11-NEXT: s_addc_u32 s1, s1, test_arg_store_v2bf16@gotpcrel32@hi+12
; GFX11-NEXT: v_writelane_b32 v9, s30, 0
; GFX11-NEXT: s_load_b64 s[0:1], s[0:1], 0x0
; GFX11-NEXT: v_writelane_b32 v9, s31, 1
; GFX11-NEXT: s_waitcnt lgkmcnt(0)
; GFX11-NEXT: s_swappc_b64 s[30:31], s[0:1]
; GFX11-NEXT: scratch_store_b128 v8, v[4:7], off offset:16 dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: scratch_store_b128 v8, v[0:3], off dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: v_readlane_b32 s31, v9, 1
; GFX11-NEXT: v_readlane_b32 s30, v9, 0
; GFX11-NEXT: s_xor_saveexec_b32 s0, -1
; GFX11-NEXT: scratch_load_b32 v9, off, s33 ; 4-byte Folded Reload
; GFX11-NEXT: s_mov_b32 exec_lo, s0
; GFX11-NEXT: s_add_i32 s32, s32, -16
; GFX11-NEXT: s_mov_b32 s33, s2
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%result = call <16 x bfloat> @test_arg_store_v2bf16(<16 x bfloat> %in)
store volatile <16 x bfloat> %result, ptr addrspace(5) %out
ret void
}
define bfloat @test_alloca_load_store_ret(bfloat %in) {
; GCN-LABEL: test_alloca_load_store_ret:
; GCN: ; %bb.0: ; %entry
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: buffer_store_short v0, off, s[0:3], s32
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_alloca_load_store_ret:
; GFX7: ; %bb.0: ; %entry
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_alloca_load_store_ret:
; GFX8: ; %bb.0: ; %entry
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_alloca_load_store_ret:
; GFX9: ; %bb.0: ; %entry
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_alloca_load_store_ret:
; GFX10: ; %bb.0: ; %entry
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_store_short v0, off, s[0:3], s32
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; GFX10-NEXT: buffer_load_ushort v0, off, s[0:3], s32 glc dlc
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_alloca_load_store_ret:
; GFX11: ; %bb.0: ; %entry
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_store_b16 off, v0, s32 dlc
; GFX11-NEXT: s_waitcnt_vscnt null, 0x0
; GFX11-NEXT: scratch_load_u16 v0, off, s32 glc dlc
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_setpc_b64 s[30:31]
entry:
%in.addr = alloca bfloat, align 2, addrspace(5)
store volatile bfloat %in, ptr addrspace(5) %in.addr, align 2
%loaded = load volatile bfloat, ptr addrspace(5) %in.addr, align 2
ret bfloat %loaded
}
define { <32 x i32>, bfloat } @test_overflow_stack(bfloat %a, <32 x i32> %b) {
; GCN-LABEL: test_overflow_stack:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x7c, v0
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: buffer_store_dword v32, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x74, v0
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: buffer_store_dword v33, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x6c, v0
; GCN-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x64, v0
; GCN-NEXT: buffer_store_dword v29, v31, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x60, v0
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x5c, v0
; GCN-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x54, v0
; GCN-NEXT: buffer_store_dword v27, v30, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x50, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x4c, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: buffer_store_dword v26, v29, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x48, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x44, v0
; GCN-NEXT: buffer_store_dword v25, v31, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v25, vcc, 64, v0
; GCN-NEXT: v_add_i32_e32 v31, vcc, 60, v0
; GCN-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v2, vcc, 56, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v24, vcc, 52, v0
; GCN-NEXT: buffer_store_dword v23, v28, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v23, vcc, 48, v0
; GCN-NEXT: v_add_i32_e32 v28, vcc, 44, v0
; GCN-NEXT: buffer_store_dword v22, v27, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v22, vcc, 40, v0
; GCN-NEXT: v_add_i32_e32 v27, vcc, 36, v0
; GCN-NEXT: buffer_store_dword v21, v30, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v21, vcc, 32, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 28, v0
; GCN-NEXT: buffer_store_dword v20, v26, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v20, vcc, 24, v0
; GCN-NEXT: v_add_i32_e32 v26, vcc, 20, v0
; GCN-NEXT: buffer_store_dword v19, v29, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v19, vcc, 16, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 12, v0
; GCN-NEXT: buffer_store_dword v18, v25, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v18, vcc, 8, v0
; GCN-NEXT: v_add_i32_e32 v25, vcc, 4, v0
; GCN-NEXT: v_add_i32_e32 v0, vcc, 0x80, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: buffer_store_dword v17, v31, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v15, v24, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v14, v23, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v13, v28, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v12, v22, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v11, v27, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v10, v21, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v9, v30, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v8, v20, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v7, v26, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v6, v19, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v29, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v18, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v25, s[0:3], 0 offen
; GCN-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: test_overflow_stack:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x7c, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x78, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32
; GFX7-NEXT: v_add_i32_e32 v31, vcc, 0x74, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x70, v0
; GFX7-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x6c, v0
; GFX7-NEXT: buffer_store_dword v29, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x68, v0
; GFX7-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x64, v0
; GFX7-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x60, v0
; GFX7-NEXT: buffer_store_dword v26, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x5c, v0
; GFX7-NEXT: buffer_store_dword v25, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x58, v0
; GFX7-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x54, v0
; GFX7-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x50, v0
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x4c, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x48, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x44, v0
; GFX7-NEXT: buffer_store_dword v19, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 64, v0
; GFX7-NEXT: buffer_store_dword v18, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 60, v0
; GFX7-NEXT: buffer_store_dword v17, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 56, v0
; GFX7-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 52, v0
; GFX7-NEXT: buffer_store_dword v15, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 48, v0
; GFX7-NEXT: buffer_store_dword v14, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 44, v0
; GFX7-NEXT: buffer_store_dword v13, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 40, v0
; GFX7-NEXT: buffer_store_dword v12, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 36, v0
; GFX7-NEXT: buffer_store_dword v11, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 32, v0
; GFX7-NEXT: buffer_store_dword v10, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 28, v0
; GFX7-NEXT: buffer_store_dword v9, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 24, v0
; GFX7-NEXT: buffer_store_dword v8, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 20, v0
; GFX7-NEXT: buffer_store_dword v7, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 16, v0
; GFX7-NEXT: buffer_store_dword v6, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 12, v0
; GFX7-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 8, v0
; GFX7-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 4, v0
; GFX7-NEXT: v_add_i32_e32 v0, vcc, 0x80, v0
; GFX7-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: test_overflow_stack:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x7c, v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x78, v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32
; GFX8-NEXT: v_add_u32_e32 v31, vcc, 0x74, v0
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x70, v0
; GFX8-NEXT: buffer_store_dword v30, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x6c, v0
; GFX8-NEXT: buffer_store_dword v29, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x68, v0
; GFX8-NEXT: buffer_store_dword v28, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x64, v0
; GFX8-NEXT: buffer_store_dword v27, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x60, v0
; GFX8-NEXT: buffer_store_dword v26, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x5c, v0
; GFX8-NEXT: buffer_store_dword v25, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x58, v0
; GFX8-NEXT: buffer_store_dword v24, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x54, v0
; GFX8-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x50, v0
; GFX8-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x4c, v0
; GFX8-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x48, v0
; GFX8-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x44, v0
; GFX8-NEXT: buffer_store_dword v19, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 64, v0
; GFX8-NEXT: buffer_store_dword v18, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 60, v0
; GFX8-NEXT: buffer_store_dword v17, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 56, v0
; GFX8-NEXT: buffer_store_dword v16, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 52, v0
; GFX8-NEXT: buffer_store_dword v15, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 48, v0
; GFX8-NEXT: buffer_store_dword v14, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 44, v0
; GFX8-NEXT: buffer_store_dword v13, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 40, v0
; GFX8-NEXT: buffer_store_dword v12, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 36, v0
; GFX8-NEXT: buffer_store_dword v11, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 32, v0
; GFX8-NEXT: buffer_store_dword v10, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 28, v0
; GFX8-NEXT: buffer_store_dword v9, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 24, v0
; GFX8-NEXT: buffer_store_dword v8, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 20, v0
; GFX8-NEXT: buffer_store_dword v7, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 16, v0
; GFX8-NEXT: buffer_store_dword v6, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 12, v0
; GFX8-NEXT: buffer_store_dword v5, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 8, v0
; GFX8-NEXT: buffer_store_dword v4, v2, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 4, v0
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 0x80, v0
; GFX8-NEXT: buffer_store_dword v3, v2, s[0:3], 0 offen
; GFX8-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: test_overflow_stack:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:4
; GFX9-NEXT: s_nop 0
; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:8
; GFX9-NEXT: s_nop 0
; GFX9-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32
; GFX9-NEXT: s_nop 0
; GFX9-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(18)
; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT: s_waitcnt vmcnt(18)
; GFX9-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen offset:128
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: test_overflow_stack:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x2
; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX10-NEXT: buffer_load_dword v33, off, s[0:3], s32
; GFX10-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:112
; GFX10-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:108
; GFX10-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:104
; GFX10-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:100
; GFX10-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:96
; GFX10-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:92
; GFX10-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:88
; GFX10-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:84
; GFX10-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:80
; GFX10-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:76
; GFX10-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:72
; GFX10-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:68
; GFX10-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:64
; GFX10-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:60
; GFX10-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX10-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:52
; GFX10-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:48
; GFX10-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:44
; GFX10-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:40
; GFX10-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:36
; GFX10-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:32
; GFX10-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:28
; GFX10-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:24
; GFX10-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:20
; GFX10-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:16
; GFX10-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:12
; GFX10-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:8
; GFX10-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:4
; GFX10-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen
; GFX10-NEXT: s_waitcnt vmcnt(2)
; GFX10-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:124
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:120
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: buffer_store_dword v33, v0, s[0:3], 0 offen offset:116
; GFX10-NEXT: buffer_store_short v1, v0, s[0:3], 0 offen offset:128
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: test_overflow_stack:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x2
; GFX11-NEXT: scratch_load_b32 v33, off, s32 offset:8
; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT: scratch_load_b32 v31, off, s32
; GFX11-NEXT: s_clause 0x5
; GFX11-NEXT: scratch_store_b128 v0, v[22:25], off offset:80
; GFX11-NEXT: scratch_store_b128 v0, v[18:21], off offset:64
; GFX11-NEXT: scratch_store_b128 v0, v[14:17], off offset:48
; GFX11-NEXT: scratch_store_b128 v0, v[10:13], off offset:32
; GFX11-NEXT: scratch_store_b128 v0, v[6:9], off offset:16
; GFX11-NEXT: scratch_store_b128 v0, v[2:5], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: s_clause 0x2
; GFX11-NEXT: scratch_store_b128 v0, v[30:33], off offset:112
; GFX11-NEXT: scratch_store_b128 v0, v[26:29], off offset:96
; GFX11-NEXT: scratch_store_b16 v0, v1, off offset:128
; GFX11-NEXT: s_setpc_b64 s[30:31]
%ins.0 = insertvalue { <32 x i32>, bfloat } poison, <32 x i32> %b, 0
%ins.1 = insertvalue { <32 x i32>, bfloat } %ins.0 ,bfloat %a, 1
ret { <32 x i32>, bfloat } %ins.1
}
define <2 x float> @global_extload_v2bf16_to_v2f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v2bf16_to_v2f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v2bf16_to_v2f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dword v1, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v2bf16_to_v2f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dword v1, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v2bf16_to_v2f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dword v1, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v2bf16_to_v2f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dword v1, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v2bf16_to_v2f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b32 v1, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <2 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <2 x bfloat> %load to <2 x float>
ret <2 x float> %fpext
}
define <3 x float> @global_extload_v3bf16_to_v3f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v3bf16_to_v3f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v3bf16_to_v3f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v3bf16_to_v3f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[1:2], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v3bf16_to_v3f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[1:2], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v3bf16_to_v3f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[1:2], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v3bf16_to_v3f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[1:2], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <3 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <3 x bfloat> %load to <3 x float>
ret <3 x float> %fpext
}
define <4 x float> @global_extload_v4bf16_to_v4f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v4bf16_to_v4f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v4bf16_to_v4f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v4bf16_to_v4f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[2:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v4bf16_to_v4f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[2:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v4bf16_to_v4f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[2:3], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v4bf16_to_v4f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[2:3], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <4 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <4 x bfloat> %load to <4 x float>
ret <4 x float> %fpext
}
define <5 x float> @global_extload_v5bf16_to_v5f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v5bf16_to_v5f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v4, v[0:1], s[4:7], 0 addr64 offset:8
; GCN-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v5bf16_to_v5f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v4, v[0:1], s[4:7], 0 addr64 offset:8
; GFX7-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v5bf16_to_v5f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[2:5], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v5bf16_to_v5f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[2:5], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v5bf16_to_v5f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[2:5], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v5bf16_to_v5f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[2:5], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <5 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <5 x bfloat> %load to <5 x float>
ret <5 x float> %fpext
}
define <6 x float> @global_extload_v6bf16_to_v6f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v6bf16_to_v6f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[3:6], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v6bf16_to_v6f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx3 v[3:5], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v6bf16_to_v6f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx3 v[3:5], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v6bf16_to_v6f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx3 v[3:5], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v6bf16_to_v6f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx3 v[3:5], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v6bf16_to_v6f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b96 v[3:5], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v3
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v4
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <6 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <6 x bfloat> %load to <6 x float>
ret <6 x float> %fpext
}
define <8 x float> @global_extload_v8bf16_to_v8f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v8bf16_to_v8f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v8bf16_to_v8f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v8bf16_to_v8f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v8bf16_to_v8f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v8bf16_to_v8f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v8bf16_to_v8f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[4:7], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <8 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <8 x bfloat> %load to <8 x float>
ret <8 x float> %fpext
}
define <16 x float> @global_extload_v16bf16_to_v16f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v16bf16_to_v16f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v16bf16_to_v16f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v16bf16_to_v16f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
; GFX8-NEXT: flat_load_dwordx4 v[12:15], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v16bf16_to_v16f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX9-NEXT: global_load_dwordx4 v[12:15], v[0:1], off offset:16
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v16bf16_to_v16f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX10-NEXT: global_load_dwordx4 v[12:15], v[0:1], off offset:16
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX10-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v16bf16_to_v16f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b128 v[4:7], v[0:1], off
; GFX11-NEXT: global_load_b128 v[12:15], v[0:1], off offset:16
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <16 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <16 x bfloat> %load to <16 x float>
ret <16 x float> %fpext
}
define <32 x float> @global_extload_v32bf16_to_v32f32(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v32bf16_to_v32f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_load_dwordx4 v[20:23], v[0:1], s[4:7], 0 addr64 offset:32
; GCN-NEXT: buffer_load_dwordx4 v[28:31], v[0:1], s[4:7], 0 addr64 offset:48
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v32bf16_to_v32f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[4:7], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_dwordx4 v[12:15], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_load_dwordx4 v[20:23], v[0:1], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_load_dwordx4 v[28:31], v[0:1], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX7-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX7-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX7-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX7-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX7-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v32bf16_to_v32f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 16, v0
; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; GFX8-NEXT: flat_load_dwordx4 v[4:7], v[0:1]
; GFX8-NEXT: flat_load_dwordx4 v[12:15], v[2:3]
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 32, v0
; GFX8-NEXT: v_addc_u32_e32 v3, vcc, 0, v1, vcc
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 48, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
; GFX8-NEXT: flat_load_dwordx4 v[20:23], v[2:3]
; GFX8-NEXT: flat_load_dwordx4 v[28:31], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(3)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: s_waitcnt vmcnt(2)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX8-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX8-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX8-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX8-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX8-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX8-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX8-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX8-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX8-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX8-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX8-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX8-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v32bf16_to_v32f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX9-NEXT: global_load_dwordx4 v[12:15], v[0:1], off offset:16
; GFX9-NEXT: global_load_dwordx4 v[20:23], v[0:1], off offset:32
; GFX9-NEXT: global_load_dwordx4 v[28:31], v[0:1], off offset:48
; GFX9-NEXT: s_waitcnt vmcnt(3)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: s_waitcnt vmcnt(2)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX9-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v32bf16_to_v32f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x3
; GFX10-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
; GFX10-NEXT: global_load_dwordx4 v[12:15], v[0:1], off offset:16
; GFX10-NEXT: global_load_dwordx4 v[20:23], v[0:1], off offset:32
; GFX10-NEXT: global_load_dwordx4 v[28:31], v[0:1], off offset:48
; GFX10-NEXT: s_waitcnt vmcnt(3)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: s_waitcnt vmcnt(2)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX10-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX10-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX10-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX10-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX10-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX10-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX10-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX10-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX10-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v32bf16_to_v32f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x3
; GFX11-NEXT: global_load_b128 v[4:7], v[0:1], off
; GFX11-NEXT: global_load_b128 v[12:15], v[0:1], off offset:16
; GFX11-NEXT: global_load_b128 v[20:23], v[0:1], off offset:32
; GFX11-NEXT: global_load_b128 v[28:31], v[0:1], off offset:48
; GFX11-NEXT: s_waitcnt vmcnt(3)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v4
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_waitcnt vmcnt(2)
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v12
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v13
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v13
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v14
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v14
; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX11-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v20
; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v20
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v21
; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v21
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v22
; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v28
; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v28
; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v29
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v29
; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v30
; GFX11-NEXT: v_and_b32_e32 v29, 0xffff0000, v30
; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX11-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <32 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <32 x bfloat> %load to <32 x float>
ret <32 x float> %fpext
}
define <2 x double> @global_extload_v2bf16_to_v2f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v2bf16_to_v2f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v1
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v2bf16_to_v2f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dword v2, v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v2bf16_to_v2f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dword v2, v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v2bf16_to_v2f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dword v2, v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v2bf16_to_v2f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dword v0, v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v1
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v2bf16_to_v2f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b32 v0, v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v1
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <2 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <2 x bfloat> %load to <2 x double>
ret <2 x double> %fpext
}
define <3 x double> @global_extload_v3bf16_to_v3f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v3bf16_to_v3f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v3bf16_to_v3f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[1:2], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v3bf16_to_v3f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[1:2], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v3bf16_to_v3f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[1:2], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v3bf16_to_v3f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v3bf16_to_v3f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[0:1], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <3 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <3 x bfloat> %load to <3 x double>
ret <3 x double> %fpext
}
define <4 x double> @global_extload_v4bf16_to_v4f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v4bf16_to_v4f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GCN-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v4bf16_to_v4f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX7-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v4bf16_to_v4f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx2 v[0:1], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v4bf16_to_v4f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx2 v[0:1], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v4bf16_to_v4f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx2 v[1:2], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX10-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v4bf16_to_v4f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b64 v[1:2], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <4 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <4 x bfloat> %load to <4 x double>
ret <4 x double> %fpext
}
define <5 x double> @global_extload_v5bf16_to_v5f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v5bf16_to_v5f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v2, v[0:1], s[4:7], 0 addr64 offset:8
; GCN-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GCN-NEXT: v_cvt_f64_f32_e32 v[8:9], v2
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GCN-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v5bf16_to_v5f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v2, v[0:1], s[4:7], 0 addr64 offset:8
; GFX7-NEXT: buffer_load_dwordx2 v[0:1], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX7-NEXT: v_cvt_f64_f32_e32 v[8:9], v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX7-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v5bf16_to_v5f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX8-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v5bf16_to_v5f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v5bf16_to_v5f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[2:5], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v3
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v4
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX10-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX10-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v5bf16_to_v5f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[2:5], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v3
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v4
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX11-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <5 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <5 x bfloat> %load to <5 x double>
ret <5 x double> %fpext
}
define <6 x double> @global_extload_v6bf16_to_v6f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v6bf16_to_v6f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GCN-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GCN-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GCN-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GCN-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v6bf16_to_v6f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx3 v[0:2], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX7-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX7-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX7-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v6bf16_to_v6f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx3 v[0:2], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX8-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX8-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v6bf16_to_v6f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx3 v[0:2], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v3
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v4
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v5
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX9-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v6bf16_to_v6f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx3 v[4:6], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v6
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v6
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX10-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX10-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX10-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v6bf16_to_v6f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b96 v[4:6], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v6
; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v6
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX11-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX11-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <6 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <6 x bfloat> %load to <6 x double>
ret <6 x double> %fpext
}
define <8 x double> @global_extload_v8bf16_to_v8f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v8bf16_to_v8f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v1
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v3
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v3
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v5
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GCN-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GCN-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GCN-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GCN-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GCN-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v8bf16_to_v8f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[0:3], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v1
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v3
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v5
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX7-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX7-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX7-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX7-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX7-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v8bf16_to_v8f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v1
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v3
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v3
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v5
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX8-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX8-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX8-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX8-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v8bf16_to_v8f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[0:3], v[0:1], off
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v1
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v3
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v3
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v4
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v5
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX9-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX9-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX9-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v8bf16_to_v8f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: global_load_dwordx4 v[7:10], v[0:1], off
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v7
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v8
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v10
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v10
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX10-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX10-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX10-NEXT: v_cvt_f64_f32_e32 v[10:11], v11
; GFX10-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX10-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v8bf16_to_v8f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: global_load_b128 v[7:10], v[0:1], off
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v8
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v10
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v10
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX11-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX11-NEXT: v_cvt_f64_f32_e32 v[10:11], v11
; GFX11-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX11-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <8 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <8 x bfloat> %load to <8 x double>
ret <8 x double> %fpext
}
define <16 x double> @global_extload_v16bf16_to_v16f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v16bf16_to_v16f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_dwordx4 v[2:5], v[0:1], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_dwordx4 v[6:9], v[0:1], s[4:7], 0 addr64 offset:16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v3
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v4
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v4
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v5
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v6
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v7
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v7
; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v8
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v8
; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v9
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v9
; GCN-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GCN-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GCN-NEXT: v_cvt_f64_f32_e32 v[4:5], v10
; GCN-NEXT: v_cvt_f64_f32_e32 v[6:7], v11
; GCN-NEXT: v_cvt_f64_f32_e32 v[8:9], v12
; GCN-NEXT: v_cvt_f64_f32_e32 v[10:11], v13
; GCN-NEXT: v_cvt_f64_f32_e32 v[12:13], v14
; GCN-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GCN-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GCN-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GCN-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GCN-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GCN-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GCN-NEXT: v_cvt_f64_f32_e32 v[26:27], v26
; GCN-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GCN-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v16bf16_to_v16f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_dwordx4 v[2:5], v[0:1], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_dwordx4 v[6:9], v[0:1], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v3
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v4
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_lshlrev_b32_e32 v16, 16, v6
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v7
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v7
; GFX7-NEXT: v_lshlrev_b32_e32 v24, 16, v8
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v8
; GFX7-NEXT: v_lshlrev_b32_e32 v28, 16, v9
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v9
; GFX7-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v10
; GFX7-NEXT: v_cvt_f64_f32_e32 v[6:7], v11
; GFX7-NEXT: v_cvt_f64_f32_e32 v[8:9], v12
; GFX7-NEXT: v_cvt_f64_f32_e32 v[10:11], v13
; GFX7-NEXT: v_cvt_f64_f32_e32 v[12:13], v14
; GFX7-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX7-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX7-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX7-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX7-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GFX7-NEXT: v_cvt_f64_f32_e32 v[26:27], v26
; GFX7-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GFX7-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v16bf16_to_v16f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: flat_load_dwordx4 v[2:5], v[0:1]
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 16, v0
; GFX8-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc
; GFX8-NEXT: flat_load_dwordx4 v[6:9], v[0:1]
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v3
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v4
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v5
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v6
; GFX8-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v7
; GFX8-NEXT: v_and_b32_e32 v22, 0xffff0000, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v8
; GFX8-NEXT: v_and_b32_e32 v26, 0xffff0000, v8
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v9
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v9
; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX8-NEXT: v_cvt_f64_f32_e32 v[4:5], v10
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v11
; GFX8-NEXT: v_cvt_f64_f32_e32 v[8:9], v12
; GFX8-NEXT: v_cvt_f64_f32_e32 v[10:11], v13
; GFX8-NEXT: v_cvt_f64_f32_e32 v[12:13], v14
; GFX8-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX8-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX8-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX8-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX8-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX8-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GFX8-NEXT: v_cvt_f64_f32_e32 v[26:27], v26
; GFX8-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GFX8-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v16bf16_to_v16f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_dwordx4 v[2:5], v[0:1], off
; GFX9-NEXT: global_load_dwordx4 v[6:9], v[0:1], off offset:16
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v3
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v4
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v5
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v6
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v7
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v8
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v8
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v9
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v9
; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX9-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[4:5], v10
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v11
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v12
; GFX9-NEXT: v_cvt_f64_f32_e32 v[10:11], v13
; GFX9-NEXT: v_cvt_f64_f32_e32 v[12:13], v14
; GFX9-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX9-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX9-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX9-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX9-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX9-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GFX9-NEXT: v_cvt_f64_f32_e32 v[26:27], v26
; GFX9-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GFX9-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v16bf16_to_v16f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: global_load_dwordx4 v[2:5], v[0:1], off
; GFX10-NEXT: global_load_dwordx4 v[9:12], v[0:1], off offset:16
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v3
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v4
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v5
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v9
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v10
; GFX10-NEXT: v_and_b32_e32 v22, 0xffff0000, v10
; GFX10-NEXT: v_lshlrev_b32_e32 v24, 16, v11
; GFX10-NEXT: v_and_b32_e32 v26, 0xffff0000, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v12
; GFX10-NEXT: v_and_b32_e32 v30, 0xffff0000, v12
; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX10-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX10-NEXT: v_cvt_f64_f32_e32 v[4:5], v6
; GFX10-NEXT: v_cvt_f64_f32_e32 v[6:7], v7
; GFX10-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX10-NEXT: v_cvt_f64_f32_e32 v[10:11], v13
; GFX10-NEXT: v_cvt_f64_f32_e32 v[12:13], v14
; GFX10-NEXT: v_cvt_f64_f32_e32 v[14:15], v15
; GFX10-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX10-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX10-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX10-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX10-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GFX10-NEXT: v_cvt_f64_f32_e32 v[26:27], v26
; GFX10-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GFX10-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v16bf16_to_v16f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: global_load_b128 v[7:10], v[0:1], off
; GFX11-NEXT: global_load_b128 v[23:26], v[0:1], off offset:16
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v8
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v10
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v10
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v23
; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v23
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v24
; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v24
; GFX11-NEXT: v_lshlrev_b32_e32 v24, 16, v25
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v25
; GFX11-NEXT: v_lshlrev_b32_e32 v28, 16, v26
; GFX11-NEXT: v_and_b32_e32 v30, 0xffff0000, v26
; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v0
; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX11-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX11-NEXT: v_cvt_f64_f32_e32 v[10:11], v11
; GFX11-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX11-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX11-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX11-NEXT: v_cvt_f64_f32_e32 v[18:19], v18
; GFX11-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX11-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX11-NEXT: v_cvt_f64_f32_e32 v[24:25], v24
; GFX11-NEXT: v_cvt_f64_f32_e32 v[26:27], v27
; GFX11-NEXT: v_cvt_f64_f32_e32 v[28:29], v28
; GFX11-NEXT: v_cvt_f64_f32_e32 v[30:31], v30
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <16 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <16 x bfloat> %load to <16 x double>
ret <16 x double> %fpext
}
define <32 x double> @global_extload_v32bf16_to_v32f64(ptr addrspace(1) %ptr) {
; GCN-LABEL: global_extload_v32bf16_to_v32f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_mov_b32 s6, 0
; GCN-NEXT: s_mov_b32 s7, 0xf000
; GCN-NEXT: s_mov_b32 s4, s6
; GCN-NEXT: s_mov_b32 s5, s6
; GCN-NEXT: buffer_load_ushort v3, v[1:2], s[4:7], 0 addr64
; GCN-NEXT: buffer_load_ushort v4, v[1:2], s[4:7], 0 addr64 offset:2
; GCN-NEXT: buffer_load_ushort v5, v[1:2], s[4:7], 0 addr64 offset:4
; GCN-NEXT: buffer_load_ushort v6, v[1:2], s[4:7], 0 addr64 offset:6
; GCN-NEXT: buffer_load_ushort v7, v[1:2], s[4:7], 0 addr64 offset:8
; GCN-NEXT: buffer_load_ushort v8, v[1:2], s[4:7], 0 addr64 offset:10
; GCN-NEXT: buffer_load_ushort v9, v[1:2], s[4:7], 0 addr64 offset:12
; GCN-NEXT: buffer_load_ushort v10, v[1:2], s[4:7], 0 addr64 offset:14
; GCN-NEXT: buffer_load_ushort v11, v[1:2], s[4:7], 0 addr64 offset:16
; GCN-NEXT: buffer_load_ushort v12, v[1:2], s[4:7], 0 addr64 offset:18
; GCN-NEXT: buffer_load_ushort v13, v[1:2], s[4:7], 0 addr64 offset:20
; GCN-NEXT: buffer_load_ushort v14, v[1:2], s[4:7], 0 addr64 offset:22
; GCN-NEXT: buffer_load_ushort v15, v[1:2], s[4:7], 0 addr64 offset:24
; GCN-NEXT: buffer_load_ushort v16, v[1:2], s[4:7], 0 addr64 offset:26
; GCN-NEXT: buffer_load_ushort v17, v[1:2], s[4:7], 0 addr64 offset:28
; GCN-NEXT: buffer_load_ushort v18, v[1:2], s[4:7], 0 addr64 offset:30
; GCN-NEXT: buffer_load_ushort v25, v[1:2], s[4:7], 0 addr64 offset:48
; GCN-NEXT: buffer_load_ushort v26, v[1:2], s[4:7], 0 addr64 offset:50
; GCN-NEXT: buffer_load_ushort v27, v[1:2], s[4:7], 0 addr64 offset:52
; GCN-NEXT: buffer_load_ushort v28, v[1:2], s[4:7], 0 addr64 offset:54
; GCN-NEXT: buffer_load_ushort v29, v[1:2], s[4:7], 0 addr64 offset:56
; GCN-NEXT: buffer_load_ushort v30, v[1:2], s[4:7], 0 addr64 offset:58
; GCN-NEXT: buffer_load_ushort v31, v[1:2], s[4:7], 0 addr64 offset:60
; GCN-NEXT: buffer_load_ushort v32, v[1:2], s[4:7], 0 addr64 offset:62
; GCN-NEXT: buffer_load_ushort v19, v[1:2], s[4:7], 0 addr64 offset:32
; GCN-NEXT: buffer_load_ushort v20, v[1:2], s[4:7], 0 addr64 offset:34
; GCN-NEXT: buffer_load_ushort v21, v[1:2], s[4:7], 0 addr64 offset:36
; GCN-NEXT: buffer_load_ushort v22, v[1:2], s[4:7], 0 addr64 offset:38
; GCN-NEXT: buffer_load_ushort v23, v[1:2], s[4:7], 0 addr64 offset:40
; GCN-NEXT: buffer_load_ushort v24, v[1:2], s[4:7], 0 addr64 offset:42
; GCN-NEXT: buffer_load_ushort v33, v[1:2], s[4:7], 0 addr64 offset:44
; GCN-NEXT: buffer_load_ushort v34, v[1:2], s[4:7], 0 addr64 offset:46
; GCN-NEXT: s_waitcnt vmcnt(8)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v32
; GCN-NEXT: v_add_i32_e32 v32, vcc, 0xfc, v0
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v32, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0xf8, v0
; GCN-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v31
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0xf4, v0
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0xf0, v0
; GCN-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0xec, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v30
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0xe8, v0
; GCN-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0xe4, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v29
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v30, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_add_i32_e32 v2, vcc, 0xe0, v0
; GCN-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0xdc, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0xd8, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v28
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v29, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0xd4, v0
; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0xd0, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0xcc, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v27
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v28, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v27, vcc, 0xc8, v0
; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0xc4, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0xc0, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v26
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v30, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v26, vcc, 0xbc, v0
; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v27, vcc, 0xb8, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0xb4, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v25
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v28, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v25, vcc, 0xb0, v0
; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0xac, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0xa8, v0
; GCN-NEXT: s_waitcnt vmcnt(14) expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v34
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v26, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v26, vcc, 0xa4, v0
; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v27, vcc, 0xa0, v0
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x9c, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v33
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v30, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x98, v0
; GCN-NEXT: buffer_store_dword v1, v25, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x94, v0
; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x90, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v24
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v28, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v24, vcc, 0x8c, v0
; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x88, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x84, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v23
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v26, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v23, vcc, 0x80, v0
; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v26, vcc, 0x7c, v0
; GCN-NEXT: v_add_i32_e32 v27, vcc, 0x78, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v22
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v22, vcc, 0x74, v0
; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v30, vcc, 0x70, v0
; GCN-NEXT: v_add_i32_e32 v31, vcc, 0x6c, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v21
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v25, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v21, vcc, 0x68, v0
; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v25, vcc, 0x64, v0
; GCN-NEXT: v_add_i32_e32 v32, vcc, 0x60, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v20
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v24, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v20, vcc, 0x5c, v0
; GCN-NEXT: buffer_store_dword v1, v28, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v24, vcc, 0x58, v0
; GCN-NEXT: v_add_i32_e32 v28, vcc, 0x54, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v19
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v29, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v19, vcc, 0x50, v0
; GCN-NEXT: buffer_store_dword v1, v23, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v23, vcc, 0x4c, v0
; GCN-NEXT: v_add_i32_e32 v29, vcc, 0x48, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v18
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v26, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v18, vcc, 0x44, v0
; GCN-NEXT: buffer_store_dword v1, v27, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v26, vcc, 64, v0
; GCN-NEXT: v_add_i32_e32 v27, vcc, 60, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v17
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v22, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v17, vcc, 56, v0
; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v22, vcc, 52, v0
; GCN-NEXT: v_add_i32_e32 v30, vcc, 48, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v31, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v31, vcc, 44, v0
; GCN-NEXT: buffer_store_dword v1, v21, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v21, vcc, 40, v0
; GCN-NEXT: v_add_i32_e32 v33, vcc, 36, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v15
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v25, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v25, vcc, 32, v0
; GCN-NEXT: buffer_store_dword v1, v32, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v32, vcc, 28, v0
; GCN-NEXT: v_add_i32_e32 v34, vcc, 24, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v14
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v20, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v20, vcc, 20, v0
; GCN-NEXT: buffer_store_dword v1, v24, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v24, vcc, 16, v0
; GCN-NEXT: v_add_i32_e32 v35, vcc, 12, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v13
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: buffer_store_dword v2, v28, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v28, vcc, 8, v0
; GCN-NEXT: buffer_store_dword v1, v19, s[0:3], 0 offen
; GCN-NEXT: v_add_i32_e32 v19, vcc, 4, v0
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v12
; GCN-NEXT: v_lshlrev_b32_e32 v11, 16, v11
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v10
; GCN-NEXT: v_lshlrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v3
; GCN-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v5
; GCN-NEXT: v_lshlrev_b32_e32 v15, 16, v6
; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v7
; GCN-NEXT: v_lshlrev_b32_e32 v36, 16, v8
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GCN-NEXT: v_cvt_f64_f32_e32 v[3:4], v11
; GCN-NEXT: buffer_store_dword v2, v23, s[0:3], 0 offen
; GCN-NEXT: v_cvt_f64_f32_e32 v[5:6], v10
; GCN-NEXT: buffer_store_dword v1, v29, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_cvt_f64_f32_e32 v[1:2], v9
; GCN-NEXT: v_cvt_f64_f32_e32 v[7:8], v12
; GCN-NEXT: v_cvt_f64_f32_e32 v[9:10], v13
; GCN-NEXT: buffer_store_dword v4, v18, s[0:3], 0 offen
; GCN-NEXT: v_cvt_f64_f32_e32 v[11:12], v36
; GCN-NEXT: buffer_store_dword v3, v26, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_cvt_f64_f32_e32 v[3:4], v14
; GCN-NEXT: v_cvt_f64_f32_e32 v[13:14], v15
; GCN-NEXT: v_cvt_f64_f32_e32 v[15:16], v16
; GCN-NEXT: buffer_store_dword v6, v27, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v5, v17, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v2, v22, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v1, v30, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v12, v31, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v11, v21, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v16, v33, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v15, v25, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v14, v32, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v13, v34, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v4, v20, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v3, v24, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v10, v35, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v9, v28, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v8, v19, s[0:3], 0 offen
; GCN-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: global_extload_v32bf16_to_v32f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_mov_b32 s6, 0
; GFX7-NEXT: s_mov_b32 s7, 0xf000
; GFX7-NEXT: s_mov_b32 s4, s6
; GFX7-NEXT: s_mov_b32 s5, s6
; GFX7-NEXT: buffer_load_ushort v20, v[1:2], s[4:7], 0 addr64 offset:62
; GFX7-NEXT: buffer_load_ushort v22, v[1:2], s[4:7], 0 addr64 offset:60
; GFX7-NEXT: buffer_load_ushort v23, v[1:2], s[4:7], 0 addr64 offset:58
; GFX7-NEXT: buffer_load_ushort v24, v[1:2], s[4:7], 0 addr64 offset:56
; GFX7-NEXT: buffer_load_ushort v25, v[1:2], s[4:7], 0 addr64 offset:54
; GFX7-NEXT: buffer_load_ushort v26, v[1:2], s[4:7], 0 addr64 offset:52
; GFX7-NEXT: buffer_load_ushort v27, v[1:2], s[4:7], 0 addr64 offset:50
; GFX7-NEXT: buffer_load_ushort v28, v[1:2], s[4:7], 0 addr64 offset:48
; GFX7-NEXT: buffer_load_ushort v15, v[1:2], s[4:7], 0 addr64 offset:32
; GFX7-NEXT: buffer_load_ushort v18, v[1:2], s[4:7], 0 addr64 offset:34
; GFX7-NEXT: buffer_load_ushort v29, v[1:2], s[4:7], 0 addr64 offset:36
; GFX7-NEXT: buffer_load_ushort v30, v[1:2], s[4:7], 0 addr64 offset:38
; GFX7-NEXT: buffer_load_ushort v31, v[1:2], s[4:7], 0 addr64 offset:40
; GFX7-NEXT: buffer_load_ushort v32, v[1:2], s[4:7], 0 addr64 offset:42
; GFX7-NEXT: buffer_load_ushort v33, v[1:2], s[4:7], 0 addr64 offset:44
; GFX7-NEXT: buffer_load_ushort v34, v[1:2], s[4:7], 0 addr64 offset:46
; GFX7-NEXT: buffer_load_ushort v19, v[1:2], s[4:7], 0 addr64
; GFX7-NEXT: buffer_load_ushort v17, v[1:2], s[4:7], 0 addr64 offset:2
; GFX7-NEXT: buffer_load_ushort v14, v[1:2], s[4:7], 0 addr64 offset:4
; GFX7-NEXT: buffer_load_ushort v12, v[1:2], s[4:7], 0 addr64 offset:6
; GFX7-NEXT: buffer_load_ushort v10, v[1:2], s[4:7], 0 addr64 offset:8
; GFX7-NEXT: buffer_load_ushort v9, v[1:2], s[4:7], 0 addr64 offset:10
; GFX7-NEXT: buffer_load_ushort v7, v[1:2], s[4:7], 0 addr64 offset:12
; GFX7-NEXT: buffer_load_ushort v4, v[1:2], s[4:7], 0 addr64 offset:14
; GFX7-NEXT: buffer_load_ushort v5, v[1:2], s[4:7], 0 addr64 offset:16
; GFX7-NEXT: buffer_load_ushort v3, v[1:2], s[4:7], 0 addr64 offset:18
; GFX7-NEXT: buffer_load_ushort v6, v[1:2], s[4:7], 0 addr64 offset:20
; GFX7-NEXT: buffer_load_ushort v8, v[1:2], s[4:7], 0 addr64 offset:22
; GFX7-NEXT: buffer_load_ushort v11, v[1:2], s[4:7], 0 addr64 offset:24
; GFX7-NEXT: buffer_load_ushort v13, v[1:2], s[4:7], 0 addr64 offset:26
; GFX7-NEXT: buffer_load_ushort v16, v[1:2], s[4:7], 0 addr64 offset:28
; GFX7-NEXT: buffer_load_ushort v1, v[1:2], s[4:7], 0 addr64 offset:30
; GFX7-NEXT: s_waitcnt vmcnt(14)
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v20
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xfc, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xf8, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v22
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xf4, v0
; GFX7-NEXT: v_add_i32_e32 v22, vcc, 0xd8, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xf0, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v23
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xec, v0
; GFX7-NEXT: s_waitcnt vmcnt(14)
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v14
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xe8, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v24
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xe4, v0
; GFX7-NEXT: v_add_i32_e32 v24, vcc, 0xd0, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xe0, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v21, 16, v25
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v21
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xdc, v0
; GFX7-NEXT: s_waitcnt vmcnt(8)
; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v21, 16, v26
; GFX7-NEXT: buffer_store_dword v20, v22, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v21
; GFX7-NEXT: v_lshlrev_b32_e32 v22, 16, v27
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xd4, v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[22:23], v22
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v20, v24, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v28
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xcc, v0
; GFX7-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xc8, v0
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xc4, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v21, 16, v34
; GFX7-NEXT: v_cvt_f64_f32_e32 v[21:22], v21
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xc0, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xbc, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v33
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[22:23], v20
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xb8, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v32
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xb4, v0
; GFX7-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xb0, v0
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xac, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v21, 16, v31
; GFX7-NEXT: v_cvt_f64_f32_e32 v[21:22], v21
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xa8, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xa4, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v30
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[22:23], v20
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0xa0, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v29
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x9c, v0
; GFX7-NEXT: buffer_store_dword v23, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x98, v0
; GFX7-NEXT: buffer_store_dword v22, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x94, v0
; GFX7-NEXT: buffer_store_dword v21, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x90, v0
; GFX7-NEXT: buffer_store_dword v20, v2, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v18
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v18, vcc, 0x8c, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v19
; GFX7-NEXT: buffer_store_dword v21, v18, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v18, vcc, 0x88, v0
; GFX7-NEXT: buffer_store_dword v20, v18, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[18:19], v2
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v15
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x84, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v17
; GFX7-NEXT: buffer_store_dword v21, v15, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x80, v0
; GFX7-NEXT: buffer_store_dword v20, v15, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[20:21], v2
; GFX7-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x7c, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v12
; GFX7-NEXT: buffer_store_dword v2, v15, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 0x78, v0
; GFX7-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[1:2], v14
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v16
; GFX7-NEXT: v_cvt_f64_f32_e32 v[14:15], v14
; GFX7-NEXT: v_add_i32_e32 v16, vcc, 0x74, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v10
; GFX7-NEXT: buffer_store_dword v15, v16, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v15, vcc, 0x70, v0
; GFX7-NEXT: buffer_store_dword v14, v15, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[14:15], v12
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v13
; GFX7-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX7-NEXT: v_add_i32_e32 v16, vcc, 0x6c, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; GFX7-NEXT: buffer_store_dword v13, v16, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v13, vcc, 0x68, v0
; GFX7-NEXT: buffer_store_dword v12, v13, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[12:13], v10
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GFX7-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX7-NEXT: v_add_i32_e32 v16, vcc, 0x64, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT: buffer_store_dword v11, v16, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[16:17], v8
; GFX7-NEXT: v_add_i32_e32 v11, vcc, 0x60, v0
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x5c, v0
; GFX7-NEXT: buffer_store_dword v10, v11, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v17, v8, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 0x58, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v11, 16, v4
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v6
; GFX7-NEXT: buffer_store_dword v16, v8, s[0:3], 0 offen
; GFX7-NEXT: v_lshlrev_b32_e32 v16, 16, v5
; GFX7-NEXT: v_cvt_f64_f32_e32 v[4:5], v4
; GFX7-NEXT: v_add_i32_e32 v6, vcc, 0x54, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v7, 16, v7
; GFX7-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 0x50, v0
; GFX7-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX7-NEXT: v_cvt_f64_f32_e32 v[5:6], v16
; GFX7-NEXT: v_add_i32_e32 v16, vcc, 0x4c, v0
; GFX7-NEXT: buffer_store_dword v4, v16, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v4, vcc, 0x48, v0
; GFX7-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX7-NEXT: v_cvt_f64_f32_e32 v[3:4], v11
; GFX7-NEXT: v_add_i32_e32 v11, vcc, 0x44, v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[7:8], v7
; GFX7-NEXT: buffer_store_dword v6, v11, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v6, vcc, 64, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v9, 16, v9
; GFX7-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 60, v0
; GFX7-NEXT: v_cvt_f64_f32_e32 v[9:10], v9
; GFX7-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v4, vcc, 56, v0
; GFX7-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 52, v0
; GFX7-NEXT: buffer_store_dword v8, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 48, v0
; GFX7-NEXT: buffer_store_dword v7, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 44, v0
; GFX7-NEXT: buffer_store_dword v10, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 40, v0
; GFX7-NEXT: buffer_store_dword v9, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 36, v0
; GFX7-NEXT: buffer_store_dword v13, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 32, v0
; GFX7-NEXT: buffer_store_dword v12, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 28, v0
; GFX7-NEXT: buffer_store_dword v15, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 24, v0
; GFX7-NEXT: buffer_store_dword v14, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v3, vcc, 20, v0
; GFX7-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 16, v0
; GFX7-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v1, vcc, 12, v0
; GFX7-NEXT: buffer_store_dword v21, v1, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v1, vcc, 8, v0
; GFX7-NEXT: buffer_store_dword v20, v1, s[0:3], 0 offen
; GFX7-NEXT: v_add_i32_e32 v1, vcc, 4, v0
; GFX7-NEXT: buffer_store_dword v19, v1, s[0:3], 0 offen
; GFX7-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: global_extload_v32bf16_to_v32f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 2, v1
; GFX8-NEXT: v_addc_u32_e32 v10, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 4, v1
; GFX8-NEXT: v_addc_u32_e32 v4, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 6, v1
; GFX8-NEXT: v_addc_u32_e32 v6, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 8, v1
; GFX8-NEXT: v_addc_u32_e32 v8, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v11, vcc, 10, v1
; GFX8-NEXT: v_addc_u32_e32 v12, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v13, vcc, 12, v1
; GFX8-NEXT: v_addc_u32_e32 v14, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v17, vcc, 14, v1
; GFX8-NEXT: v_addc_u32_e32 v18, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v21, vcc, 16, v1
; GFX8-NEXT: v_addc_u32_e32 v22, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v15, vcc, 18, v1
; GFX8-NEXT: v_addc_u32_e32 v16, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v19, vcc, 20, v1
; GFX8-NEXT: v_addc_u32_e32 v20, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v23, vcc, 22, v1
; GFX8-NEXT: v_addc_u32_e32 v24, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v25, vcc, 24, v1
; GFX8-NEXT: v_addc_u32_e32 v26, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v27, vcc, 26, v1
; GFX8-NEXT: v_addc_u32_e32 v28, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v29, vcc, 28, v1
; GFX8-NEXT: v_addc_u32_e32 v30, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v31, vcc, 30, v1
; GFX8-NEXT: v_addc_u32_e32 v32, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v33, vcc, 32, v1
; GFX8-NEXT: v_addc_u32_e32 v34, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v35, vcc, 34, v1
; GFX8-NEXT: v_addc_u32_e32 v36, vcc, 0, v2, vcc
; GFX8-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX8-NEXT: buffer_store_dword v57, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX8-NEXT: v_add_u32_e32 v37, vcc, 36, v1
; GFX8-NEXT: flat_load_ushort v43, v[1:2]
; GFX8-NEXT: v_addc_u32_e32 v38, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v48, vcc, 38, v1
; GFX8-NEXT: v_addc_u32_e32 v49, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v50, vcc, 62, v1
; GFX8-NEXT: v_addc_u32_e32 v51, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v44, v[50:51]
; GFX8-NEXT: v_add_u32_e32 v50, vcc, 60, v1
; GFX8-NEXT: v_addc_u32_e32 v51, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v45, v[50:51]
; GFX8-NEXT: v_add_u32_e32 v50, vcc, 40, v1
; GFX8-NEXT: v_addc_u32_e32 v51, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v52, vcc, 58, v1
; GFX8-NEXT: v_addc_u32_e32 v53, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v46, v[52:53]
; GFX8-NEXT: v_add_u32_e32 v52, vcc, 42, v1
; GFX8-NEXT: v_addc_u32_e32 v53, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v54, vcc, 56, v1
; GFX8-NEXT: v_addc_u32_e32 v55, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v47, v[54:55]
; GFX8-NEXT: v_add_u32_e32 v54, vcc, 44, v1
; GFX8-NEXT: v_addc_u32_e32 v55, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v39, vcc, 54, v1
; GFX8-NEXT: v_addc_u32_e32 v40, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v56, v[39:40]
; GFX8-NEXT: v_add_u32_e32 v39, vcc, 52, v1
; GFX8-NEXT: v_addc_u32_e32 v40, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v57, v[39:40]
; GFX8-NEXT: v_add_u32_e32 v39, vcc, 46, v1
; GFX8-NEXT: v_addc_u32_e32 v40, vcc, 0, v2, vcc
; GFX8-NEXT: v_add_u32_e32 v41, vcc, 50, v1
; GFX8-NEXT: v_addc_u32_e32 v42, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v41, v[41:42]
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 48, v1
; GFX8-NEXT: v_addc_u32_e32 v2, vcc, 0, v2, vcc
; GFX8-NEXT: flat_load_ushort v42, v[9:10]
; GFX8-NEXT: flat_load_ushort v9, v[35:36]
; GFX8-NEXT: flat_load_ushort v10, v[37:38]
; GFX8-NEXT: flat_load_ushort v35, v[48:49]
; GFX8-NEXT: flat_load_ushort v36, v[50:51]
; GFX8-NEXT: flat_load_ushort v37, v[52:53]
; GFX8-NEXT: flat_load_ushort v48, v[54:55]
; GFX8-NEXT: flat_load_ushort v39, v[39:40]
; GFX8-NEXT: flat_load_ushort v49, v[1:2]
; GFX8-NEXT: flat_load_ushort v50, v[3:4]
; GFX8-NEXT: flat_load_ushort v51, v[5:6]
; GFX8-NEXT: flat_load_ushort v52, v[7:8]
; GFX8-NEXT: flat_load_ushort v53, v[11:12]
; GFX8-NEXT: flat_load_ushort v38, v[13:14]
; GFX8-NEXT: flat_load_ushort v14, v[17:18]
; GFX8-NEXT: flat_load_ushort v11, v[21:22]
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 4, v0
; GFX8-NEXT: flat_load_ushort v15, v[15:16]
; GFX8-NEXT: flat_load_ushort v13, v[19:20]
; GFX8-NEXT: flat_load_ushort v8, v[23:24]
; GFX8-NEXT: flat_load_ushort v6, v[25:26]
; GFX8-NEXT: flat_load_ushort v5, v[27:28]
; GFX8-NEXT: flat_load_ushort v7, v[29:30]
; GFX8-NEXT: flat_load_ushort v12, v[31:32]
; GFX8-NEXT: flat_load_ushort v16, v[33:34]
; GFX8-NEXT: v_add_u32_e32 v18, vcc, 0xc4, v0
; GFX8-NEXT: v_add_u32_e32 v20, vcc, 0xbc, v0
; GFX8-NEXT: v_add_u32_e32 v22, vcc, 0xb4, v0
; GFX8-NEXT: v_add_u32_e32 v24, vcc, 0xac, v0
; GFX8-NEXT: v_add_u32_e32 v26, vcc, 0xa4, v0
; GFX8-NEXT: v_add_u32_e32 v27, vcc, 0x9c, v0
; GFX8-NEXT: s_waitcnt vmcnt(14)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v43
; GFX8-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xfc, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v44
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX8-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v45
; GFX8-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xf8, v0
; GFX8-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xf4, v0
; GFX8-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v46
; GFX8-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xf0, v0
; GFX8-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xec, v0
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xe8, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v47
; GFX8-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
; GFX8-NEXT: buffer_store_dword v1, v4, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xe4, v0
; GFX8-NEXT: buffer_store_dword v3, v1, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xe0, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v56
; GFX8-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX8-NEXT: buffer_store_dword v2, v1, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0xdc, v0
; GFX8-NEXT: buffer_store_dword v4, v1, s[0:3], 0 offen
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v57
; GFX8-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xd8, v0
; GFX8-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0xd4, v0
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0xd0, v0
; GFX8-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v41
; GFX8-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xcc, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v42
; GFX8-NEXT: buffer_store_dword v2, v4, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0xc8, v0
; GFX8-NEXT: buffer_store_dword v1, v2, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[1:2], v3
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v49
; GFX8-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v50
; GFX8-NEXT: s_waitcnt vmcnt(14)
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v51
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v52
; GFX8-NEXT: buffer_store_dword v4, v18, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0xc0, v0
; GFX8-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[3:4], v17
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v39
; GFX8-NEXT: v_cvt_f64_f32_e32 v[17:18], v17
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v53
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v38
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v10
; GFX8-NEXT: buffer_store_dword v18, v20, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v18, vcc, 0xb8, v0
; GFX8-NEXT: buffer_store_dword v17, v18, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[17:18], v19
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v48
; GFX8-NEXT: v_cvt_f64_f32_e32 v[19:20], v19
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v9
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT: buffer_store_dword v20, v22, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v20, vcc, 0xb0, v0
; GFX8-NEXT: buffer_store_dword v19, v20, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[19:20], v21
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v37
; GFX8-NEXT: v_cvt_f64_f32_e32 v[21:22], v21
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; GFX8-NEXT: buffer_store_dword v22, v24, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v22, vcc, 0xa8, v0
; GFX8-NEXT: buffer_store_dword v21, v22, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[21:22], v23
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v36
; GFX8-NEXT: v_cvt_f64_f32_e32 v[23:24], v23
; GFX8-NEXT: buffer_store_dword v24, v26, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v24, vcc, 0xa0, v0
; GFX8-NEXT: buffer_store_dword v23, v24, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[23:24], v25
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v35
; GFX8-NEXT: v_cvt_f64_f32_e32 v[25:26], v25
; GFX8-NEXT: buffer_store_dword v26, v27, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[27:28], v10
; GFX8-NEXT: v_add_u32_e32 v26, vcc, 0x98, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GFX8-NEXT: v_add_u32_e32 v11, vcc, 0x94, v0
; GFX8-NEXT: buffer_store_dword v25, v26, s[0:3], 0 offen
; GFX8-NEXT: buffer_store_dword v28, v11, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v11, vcc, 0x90, v0
; GFX8-NEXT: buffer_store_dword v27, v11, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[27:28], v9
; GFX8-NEXT: v_cvt_f64_f32_e32 v[25:26], v14
; GFX8-NEXT: v_add_u32_e32 v14, vcc, 0x8c, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v15
; GFX8-NEXT: buffer_store_dword v28, v14, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v14, vcc, 0x88, v0
; GFX8-NEXT: buffer_store_dword v27, v14, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[14:15], v9
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v16
; GFX8-NEXT: v_cvt_f64_f32_e32 v[27:28], v9
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v13
; GFX8-NEXT: v_add_u32_e32 v13, vcc, 0x84, v0
; GFX8-NEXT: buffer_store_dword v28, v13, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v13, vcc, 0x80, v0
; GFX8-NEXT: buffer_store_dword v27, v13, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[27:28], v9
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v12
; GFX8-NEXT: v_cvt_f64_f32_e32 v[12:13], v9
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7c, v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[10:11], v10
; GFX8-NEXT: buffer_store_dword v13, v9, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x78, v0
; GFX8-NEXT: buffer_store_dword v12, v9, s[0:3], 0 offen
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX8-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX8-NEXT: v_add_u32_e32 v13, vcc, 0x74, v0
; GFX8-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX8-NEXT: buffer_store_dword v7, v13, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x70, v0
; GFX8-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX8-NEXT: v_cvt_f64_f32_e32 v[5:6], v5
; GFX8-NEXT: v_cvt_f64_f32_e32 v[12:13], v12
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x6c, v0
; GFX8-NEXT: buffer_store_dword v6, v7, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x68, v0
; GFX8-NEXT: buffer_store_dword v5, v6, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x64, v0
; GFX8-NEXT: buffer_store_dword v13, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x60, v0
; GFX8-NEXT: buffer_store_dword v12, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x5c, v0
; GFX8-NEXT: buffer_store_dword v9, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x58, v0
; GFX8-NEXT: buffer_store_dword v8, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x54, v0
; GFX8-NEXT: buffer_store_dword v28, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x50, v0
; GFX8-NEXT: buffer_store_dword v27, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x4c, v0
; GFX8-NEXT: buffer_store_dword v15, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x48, v0
; GFX8-NEXT: buffer_store_dword v14, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x44, v0
; GFX8-NEXT: buffer_store_dword v11, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 64, v0
; GFX8-NEXT: buffer_store_dword v10, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 60, v0
; GFX8-NEXT: buffer_store_dword v26, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 56, v0
; GFX8-NEXT: buffer_store_dword v25, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 52, v0
; GFX8-NEXT: buffer_store_dword v24, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 48, v0
; GFX8-NEXT: buffer_store_dword v23, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 44, v0
; GFX8-NEXT: buffer_store_dword v22, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 40, v0
; GFX8-NEXT: buffer_store_dword v21, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 36, v0
; GFX8-NEXT: buffer_store_dword v20, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 32, v0
; GFX8-NEXT: buffer_store_dword v19, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 28, v0
; GFX8-NEXT: buffer_store_dword v18, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 24, v0
; GFX8-NEXT: buffer_store_dword v17, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 20, v0
; GFX8-NEXT: buffer_store_dword v4, v5, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 16, v0
; GFX8-NEXT: buffer_store_dword v3, v4, s[0:3], 0 offen
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 12, v0
; GFX8-NEXT: v_add_u32_e32 v0, vcc, 8, v0
; GFX8-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
; GFX8-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX8-NEXT: buffer_load_dword v57, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX8-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: global_extload_v32bf16_to_v32f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: global_load_ushort v21, v[1:2], off offset:62
; GFX9-NEXT: global_load_ushort v23, v[1:2], off offset:60
; GFX9-NEXT: global_load_ushort v24, v[1:2], off offset:58
; GFX9-NEXT: global_load_ushort v25, v[1:2], off offset:56
; GFX9-NEXT: global_load_ushort v26, v[1:2], off offset:54
; GFX9-NEXT: global_load_ushort v27, v[1:2], off offset:52
; GFX9-NEXT: global_load_ushort v28, v[1:2], off offset:50
; GFX9-NEXT: global_load_ushort v29, v[1:2], off offset:48
; GFX9-NEXT: global_load_ushort v30, v[1:2], off offset:46
; GFX9-NEXT: global_load_ushort v31, v[1:2], off offset:44
; GFX9-NEXT: global_load_ushort v32, v[1:2], off offset:42
; GFX9-NEXT: global_load_ushort v33, v[1:2], off offset:40
; GFX9-NEXT: global_load_ushort v34, v[1:2], off offset:38
; GFX9-NEXT: global_load_ushort v19, v[1:2], off
; GFX9-NEXT: global_load_ushort v20, v[1:2], off offset:36
; GFX9-NEXT: global_load_ushort v17, v[1:2], off offset:2
; GFX9-NEXT: global_load_ushort v18, v[1:2], off offset:4
; GFX9-NEXT: global_load_ushort v16, v[1:2], off offset:34
; GFX9-NEXT: global_load_ushort v11, v[1:2], off offset:32
; GFX9-NEXT: global_load_ushort v13, v[1:2], off offset:6
; GFX9-NEXT: global_load_ushort v14, v[1:2], off offset:8
; GFX9-NEXT: global_load_ushort v15, v[1:2], off offset:30
; GFX9-NEXT: global_load_ushort v3, v[1:2], off offset:16
; GFX9-NEXT: global_load_ushort v4, v[1:2], off offset:18
; GFX9-NEXT: global_load_ushort v5, v[1:2], off offset:20
; GFX9-NEXT: global_load_ushort v6, v[1:2], off offset:22
; GFX9-NEXT: global_load_ushort v8, v[1:2], off offset:24
; GFX9-NEXT: global_load_ushort v10, v[1:2], off offset:26
; GFX9-NEXT: global_load_ushort v12, v[1:2], off offset:28
; GFX9-NEXT: global_load_ushort v9, v[1:2], off offset:10
; GFX9-NEXT: global_load_ushort v7, v[1:2], off offset:12
; GFX9-NEXT: s_nop 0
; GFX9-NEXT: global_load_ushort v1, v[1:2], off offset:14
; GFX9-NEXT: s_waitcnt vmcnt(31)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v21
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v2
; GFX9-NEXT: s_waitcnt vmcnt(30)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v23
; GFX9-NEXT: s_waitcnt vmcnt(28)
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v25
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:252
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:248
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v2
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v24
; GFX9-NEXT: s_waitcnt vmcnt(29)
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v26
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:244
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:240
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v2
; GFX9-NEXT: s_waitcnt vmcnt(30)
; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v27
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:236
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:232
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v23
; GFX9-NEXT: v_cvt_f64_f32_e32 v[23:24], v24
; GFX9-NEXT: s_waitcnt vmcnt(31)
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v28
; GFX9-NEXT: s_waitcnt vmcnt(30)
; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v29
; GFX9-NEXT: s_waitcnt vmcnt(29)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v30
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:228
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:224
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v25
; GFX9-NEXT: v_cvt_f64_f32_e32 v[25:26], v26
; GFX9-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:220
; GFX9-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:216
; GFX9-NEXT: v_cvt_f64_f32_e32 v[23:24], v27
; GFX9-NEXT: v_cvt_f64_f32_e32 v[27:28], v2
; GFX9-NEXT: s_waitcnt vmcnt(28)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v19
; GFX9-NEXT: s_waitcnt vmcnt(27)
; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v20
; GFX9-NEXT: v_cvt_f64_f32_e32 v[19:20], v19
; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v31
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v32
; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v33
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v34
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:212
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:208
; GFX9-NEXT: v_cvt_f64_f32_e32 v[21:22], v29
; GFX9-NEXT: s_waitcnt vmcnt(26)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT: v_cvt_f64_f32_e32 v[29:30], v30
; GFX9-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:204
; GFX9-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:200
; GFX9-NEXT: v_cvt_f64_f32_e32 v[25:26], v31
; GFX9-NEXT: v_cvt_f64_f32_e32 v[31:32], v32
; GFX9-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:196
; GFX9-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:192
; GFX9-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:188
; GFX9-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:184
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:180
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:176
; GFX9-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:172
; GFX9-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:168
; GFX9-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:164
; GFX9-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:160
; GFX9-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:156
; GFX9-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:152
; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v17
; GFX9-NEXT: v_cvt_f64_f32_e32 v[16:17], v16
; GFX9-NEXT: s_waitcnt vmcnt(39)
; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v11
; GFX9-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:148
; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:144
; GFX9-NEXT: s_waitcnt vmcnt(40)
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v13
; GFX9-NEXT: s_waitcnt vmcnt(39)
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v14
; GFX9-NEXT: v_cvt_f64_f32_e32 v[13:14], v11
; GFX9-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:140
; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:136
; GFX9-NEXT: v_cvt_f64_f32_e32 v[16:17], v2
; GFX9-NEXT: s_waitcnt vmcnt(40)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v15
; GFX9-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:132
; GFX9-NEXT: v_cvt_f64_f32_e32 v[14:15], v2
; GFX9-NEXT: s_waitcnt vmcnt(34)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v12
; GFX9-NEXT: v_cvt_f64_f32_e32 v[11:12], v2
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v10
; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:128
; GFX9-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT: v_cvt_f64_f32_e32 v[10:11], v2
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT: s_waitcnt vmcnt(38)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v9
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v8
; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT: v_cvt_f64_f32_e32 v[10:11], v2
; GFX9-NEXT: s_waitcnt vmcnt(39)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT: s_waitcnt vmcnt(38)
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v5
; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT: v_cvt_f64_f32_e32 v[6:7], v6
; GFX9-NEXT: v_cvt_f64_f32_e32 v[8:9], v2
; GFX9-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v4
; GFX9-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX9-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT: v_cvt_f64_f32_e32 v[1:2], v5
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v18
; GFX9-NEXT: v_cvt_f64_f32_e32 v[18:19], v21
; GFX9-NEXT: v_cvt_f64_f32_e32 v[13:14], v22
; GFX9-NEXT: v_cvt_f64_f32_e32 v[20:21], v20
; GFX9-NEXT: v_cvt_f64_f32_e32 v[22:23], v23
; GFX9-NEXT: v_cvt_f64_f32_e32 v[5:6], v12
; GFX9-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: global_extload_v32bf16_to_v32f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1f
; GFX10-NEXT: global_load_ushort v3, v[1:2], off
; GFX10-NEXT: global_load_ushort v4, v[1:2], off offset:2
; GFX10-NEXT: global_load_ushort v5, v[1:2], off offset:4
; GFX10-NEXT: global_load_ushort v6, v[1:2], off offset:6
; GFX10-NEXT: global_load_ushort v7, v[1:2], off offset:8
; GFX10-NEXT: global_load_ushort v8, v[1:2], off offset:10
; GFX10-NEXT: global_load_ushort v9, v[1:2], off offset:12
; GFX10-NEXT: global_load_ushort v10, v[1:2], off offset:14
; GFX10-NEXT: global_load_ushort v11, v[1:2], off offset:16
; GFX10-NEXT: global_load_ushort v12, v[1:2], off offset:18
; GFX10-NEXT: global_load_ushort v13, v[1:2], off offset:20
; GFX10-NEXT: global_load_ushort v14, v[1:2], off offset:22
; GFX10-NEXT: global_load_ushort v15, v[1:2], off offset:24
; GFX10-NEXT: global_load_ushort v16, v[1:2], off offset:26
; GFX10-NEXT: global_load_ushort v17, v[1:2], off offset:28
; GFX10-NEXT: global_load_ushort v18, v[1:2], off offset:30
; GFX10-NEXT: global_load_ushort v19, v[1:2], off offset:32
; GFX10-NEXT: global_load_ushort v20, v[1:2], off offset:34
; GFX10-NEXT: global_load_ushort v21, v[1:2], off offset:36
; GFX10-NEXT: global_load_ushort v22, v[1:2], off offset:38
; GFX10-NEXT: global_load_ushort v23, v[1:2], off offset:40
; GFX10-NEXT: global_load_ushort v24, v[1:2], off offset:42
; GFX10-NEXT: global_load_ushort v25, v[1:2], off offset:44
; GFX10-NEXT: global_load_ushort v26, v[1:2], off offset:46
; GFX10-NEXT: global_load_ushort v27, v[1:2], off offset:48
; GFX10-NEXT: global_load_ushort v28, v[1:2], off offset:62
; GFX10-NEXT: global_load_ushort v29, v[1:2], off offset:50
; GFX10-NEXT: global_load_ushort v30, v[1:2], off offset:52
; GFX10-NEXT: global_load_ushort v31, v[1:2], off offset:54
; GFX10-NEXT: global_load_ushort v32, v[1:2], off offset:60
; GFX10-NEXT: global_load_ushort v33, v[1:2], off offset:56
; GFX10-NEXT: global_load_ushort v34, v[1:2], off offset:58
; GFX10-NEXT: s_waitcnt vmcnt(31)
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: s_waitcnt vmcnt(30)
; GFX10-NEXT: v_lshlrev_b32_e32 v35, 16, v4
; GFX10-NEXT: s_waitcnt vmcnt(29)
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v5
; GFX10-NEXT: s_waitcnt vmcnt(28)
; GFX10-NEXT: v_lshlrev_b32_e32 v37, 16, v6
; GFX10-NEXT: s_waitcnt vmcnt(27)
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v7
; GFX10-NEXT: s_waitcnt vmcnt(26)
; GFX10-NEXT: v_lshlrev_b32_e32 v39, 16, v8
; GFX10-NEXT: s_waitcnt vmcnt(25)
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v9
; GFX10-NEXT: s_waitcnt vmcnt(24)
; GFX10-NEXT: v_lshlrev_b32_e32 v49, 16, v10
; GFX10-NEXT: s_waitcnt vmcnt(23)
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v11
; GFX10-NEXT: s_waitcnt vmcnt(22)
; GFX10-NEXT: v_lshlrev_b32_e32 v51, 16, v12
; GFX10-NEXT: s_waitcnt vmcnt(21)
; GFX10-NEXT: v_lshlrev_b32_e32 v52, 16, v13
; GFX10-NEXT: s_waitcnt vmcnt(20)
; GFX10-NEXT: v_lshlrev_b32_e32 v53, 16, v14
; GFX10-NEXT: s_waitcnt vmcnt(19)
; GFX10-NEXT: v_lshlrev_b32_e32 v54, 16, v15
; GFX10-NEXT: s_waitcnt vmcnt(18)
; GFX10-NEXT: v_lshlrev_b32_e32 v55, 16, v16
; GFX10-NEXT: v_cvt_f64_f32_e32 v[11:12], v37
; GFX10-NEXT: v_cvt_f64_f32_e32 v[15:16], v38
; GFX10-NEXT: s_waitcnt vmcnt(15)
; GFX10-NEXT: v_lshlrev_b32_e32 v66, 16, v19
; GFX10-NEXT: s_waitcnt vmcnt(14)
; GFX10-NEXT: v_lshlrev_b32_e32 v67, 16, v20
; GFX10-NEXT: s_waitcnt vmcnt(13)
; GFX10-NEXT: v_lshlrev_b32_e32 v64, 16, v21
; GFX10-NEXT: s_waitcnt vmcnt(12)
; GFX10-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX10-NEXT: s_waitcnt vmcnt(11)
; GFX10-NEXT: v_lshlrev_b32_e32 v70, 16, v23
; GFX10-NEXT: s_waitcnt vmcnt(10)
; GFX10-NEXT: v_lshlrev_b32_e32 v71, 16, v24
; GFX10-NEXT: s_waitcnt vmcnt(9)
; GFX10-NEXT: v_lshlrev_b32_e32 v80, 16, v25
; GFX10-NEXT: s_waitcnt vmcnt(8)
; GFX10-NEXT: v_lshlrev_b32_e32 v81, 16, v26
; GFX10-NEXT: s_waitcnt vmcnt(7)
; GFX10-NEXT: v_lshlrev_b32_e32 v82, 16, v27
; GFX10-NEXT: s_waitcnt vmcnt(6)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v28
; GFX10-NEXT: s_waitcnt vmcnt(5)
; GFX10-NEXT: v_lshlrev_b32_e32 v83, 16, v29
; GFX10-NEXT: s_waitcnt vmcnt(4)
; GFX10-NEXT: v_lshlrev_b32_e32 v84, 16, v30
; GFX10-NEXT: s_waitcnt vmcnt(3)
; GFX10-NEXT: v_lshlrev_b32_e32 v29, 16, v31
; GFX10-NEXT: s_waitcnt vmcnt(2)
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v32
; GFX10-NEXT: v_cvt_f64_f32_e32 v[1:2], v1
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v13, 16, v34
; GFX10-NEXT: v_lshlrev_b32_e32 v21, 16, v33
; GFX10-NEXT: v_cvt_f64_f32_e32 v[29:30], v29
; GFX10-NEXT: v_cvt_f64_f32_e32 v[5:6], v5
; GFX10-NEXT: v_cvt_f64_f32_e32 v[37:38], v84
; GFX10-NEXT: v_cvt_f64_f32_e32 v[13:14], v13
; GFX10-NEXT: v_cvt_f64_f32_e32 v[21:22], v21
; GFX10-NEXT: v_cvt_f64_f32_e32 v[25:26], v50
; GFX10-NEXT: v_cvt_f64_f32_e32 v[27:28], v51
; GFX10-NEXT: v_cvt_f64_f32_e32 v[50:51], v82
; GFX10-NEXT: v_cvt_f64_f32_e32 v[31:32], v52
; GFX10-NEXT: v_cvt_f64_f32_e32 v[33:34], v53
; GFX10-NEXT: v_cvt_f64_f32_e32 v[52:53], v80
; GFX10-NEXT: v_cvt_f64_f32_e32 v[7:8], v35
; GFX10-NEXT: v_cvt_f64_f32_e32 v[9:10], v36
; GFX10-NEXT: v_cvt_f64_f32_e32 v[19:20], v48
; GFX10-NEXT: v_cvt_f64_f32_e32 v[23:24], v49
; GFX10-NEXT: v_cvt_f64_f32_e32 v[35:36], v54
; GFX10-NEXT: v_cvt_f64_f32_e32 v[48:49], v55
; GFX10-NEXT: v_cvt_f64_f32_e32 v[54:55], v70
; GFX10-NEXT: v_lshlrev_b32_e32 v69, 16, v18
; GFX10-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:252
; GFX10-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:248
; GFX10-NEXT: v_cvt_f64_f32_e32 v[1:2], v83
; GFX10-NEXT: v_lshlrev_b32_e32 v68, 16, v17
; GFX10-NEXT: v_cvt_f64_f32_e32 v[3:4], v3
; GFX10-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:244
; GFX10-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:240
; GFX10-NEXT: v_cvt_f64_f32_e32 v[5:6], v81
; GFX10-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:236
; GFX10-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:232
; GFX10-NEXT: v_cvt_f64_f32_e32 v[13:14], v71
; GFX10-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:228
; GFX10-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:224
; GFX10-NEXT: v_cvt_f64_f32_e32 v[21:22], v65
; GFX10-NEXT: v_cvt_f64_f32_e32 v[64:65], v64
; GFX10-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:220
; GFX10-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:216
; GFX10-NEXT: v_cvt_f64_f32_e32 v[29:30], v67
; GFX10-NEXT: v_cvt_f64_f32_e32 v[66:67], v66
; GFX10-NEXT: buffer_store_dword v38, v0, s[0:3], 0 offen offset:212
; GFX10-NEXT: buffer_store_dword v37, v0, s[0:3], 0 offen offset:208
; GFX10-NEXT: v_cvt_f64_f32_e32 v[37:38], v69
; GFX10-NEXT: v_cvt_f64_f32_e32 v[17:18], v39
; GFX10-NEXT: v_cvt_f64_f32_e32 v[68:69], v68
; GFX10-NEXT: buffer_store_dword v2, v0, s[0:3], 0 offen offset:204
; GFX10-NEXT: buffer_store_dword v1, v0, s[0:3], 0 offen offset:200
; GFX10-NEXT: buffer_store_dword v51, v0, s[0:3], 0 offen offset:196
; GFX10-NEXT: buffer_store_dword v50, v0, s[0:3], 0 offen offset:192
; GFX10-NEXT: buffer_store_dword v6, v0, s[0:3], 0 offen offset:188
; GFX10-NEXT: buffer_store_dword v5, v0, s[0:3], 0 offen offset:184
; GFX10-NEXT: buffer_store_dword v53, v0, s[0:3], 0 offen offset:180
; GFX10-NEXT: buffer_store_dword v52, v0, s[0:3], 0 offen offset:176
; GFX10-NEXT: buffer_store_dword v14, v0, s[0:3], 0 offen offset:172
; GFX10-NEXT: buffer_store_dword v13, v0, s[0:3], 0 offen offset:168
; GFX10-NEXT: buffer_store_dword v55, v0, s[0:3], 0 offen offset:164
; GFX10-NEXT: buffer_store_dword v54, v0, s[0:3], 0 offen offset:160
; GFX10-NEXT: buffer_store_dword v22, v0, s[0:3], 0 offen offset:156
; GFX10-NEXT: buffer_store_dword v21, v0, s[0:3], 0 offen offset:152
; GFX10-NEXT: buffer_store_dword v65, v0, s[0:3], 0 offen offset:148
; GFX10-NEXT: buffer_store_dword v64, v0, s[0:3], 0 offen offset:144
; GFX10-NEXT: buffer_store_dword v30, v0, s[0:3], 0 offen offset:140
; GFX10-NEXT: buffer_store_dword v29, v0, s[0:3], 0 offen offset:136
; GFX10-NEXT: buffer_store_dword v67, v0, s[0:3], 0 offen offset:132
; GFX10-NEXT: buffer_store_dword v66, v0, s[0:3], 0 offen offset:128
; GFX10-NEXT: buffer_store_dword v38, v0, s[0:3], 0 offen offset:124
; GFX10-NEXT: buffer_store_dword v37, v0, s[0:3], 0 offen offset:120
; GFX10-NEXT: buffer_store_dword v69, v0, s[0:3], 0 offen offset:116
; GFX10-NEXT: buffer_store_dword v68, v0, s[0:3], 0 offen offset:112
; GFX10-NEXT: buffer_store_dword v49, v0, s[0:3], 0 offen offset:108
; GFX10-NEXT: buffer_store_dword v48, v0, s[0:3], 0 offen offset:104
; GFX10-NEXT: buffer_store_dword v36, v0, s[0:3], 0 offen offset:100
; GFX10-NEXT: buffer_store_dword v35, v0, s[0:3], 0 offen offset:96
; GFX10-NEXT: buffer_store_dword v34, v0, s[0:3], 0 offen offset:92
; GFX10-NEXT: buffer_store_dword v33, v0, s[0:3], 0 offen offset:88
; GFX10-NEXT: buffer_store_dword v32, v0, s[0:3], 0 offen offset:84
; GFX10-NEXT: buffer_store_dword v31, v0, s[0:3], 0 offen offset:80
; GFX10-NEXT: buffer_store_dword v28, v0, s[0:3], 0 offen offset:76
; GFX10-NEXT: buffer_store_dword v27, v0, s[0:3], 0 offen offset:72
; GFX10-NEXT: buffer_store_dword v26, v0, s[0:3], 0 offen offset:68
; GFX10-NEXT: buffer_store_dword v25, v0, s[0:3], 0 offen offset:64
; GFX10-NEXT: buffer_store_dword v24, v0, s[0:3], 0 offen offset:60
; GFX10-NEXT: buffer_store_dword v23, v0, s[0:3], 0 offen offset:56
; GFX10-NEXT: buffer_store_dword v20, v0, s[0:3], 0 offen offset:52
; GFX10-NEXT: buffer_store_dword v19, v0, s[0:3], 0 offen offset:48
; GFX10-NEXT: buffer_store_dword v18, v0, s[0:3], 0 offen offset:44
; GFX10-NEXT: buffer_store_dword v17, v0, s[0:3], 0 offen offset:40
; GFX10-NEXT: buffer_store_dword v16, v0, s[0:3], 0 offen offset:36
; GFX10-NEXT: buffer_store_dword v15, v0, s[0:3], 0 offen offset:32
; GFX10-NEXT: buffer_store_dword v12, v0, s[0:3], 0 offen offset:28
; GFX10-NEXT: buffer_store_dword v11, v0, s[0:3], 0 offen offset:24
; GFX10-NEXT: buffer_store_dword v10, v0, s[0:3], 0 offen offset:20
; GFX10-NEXT: buffer_store_dword v9, v0, s[0:3], 0 offen offset:16
; GFX10-NEXT: buffer_store_dword v8, v0, s[0:3], 0 offen offset:12
; GFX10-NEXT: buffer_store_dword v7, v0, s[0:3], 0 offen offset:8
; GFX10-NEXT: buffer_store_dword v4, v0, s[0:3], 0 offen offset:4
; GFX10-NEXT: buffer_store_dword v3, v0, s[0:3], 0 offen
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: global_extload_v32bf16_to_v32f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1f
; GFX11-NEXT: global_load_u16 v3, v[1:2], off offset:12
; GFX11-NEXT: global_load_u16 v4, v[1:2], off offset:8
; GFX11-NEXT: global_load_u16 v5, v[1:2], off offset:4
; GFX11-NEXT: global_load_u16 v6, v[1:2], off offset:2
; GFX11-NEXT: global_load_u16 v7, v[1:2], off
; GFX11-NEXT: global_load_u16 v8, v[1:2], off offset:6
; GFX11-NEXT: global_load_u16 v9, v[1:2], off offset:10
; GFX11-NEXT: global_load_u16 v10, v[1:2], off offset:14
; GFX11-NEXT: global_load_u16 v11, v[1:2], off offset:28
; GFX11-NEXT: global_load_u16 v12, v[1:2], off offset:24
; GFX11-NEXT: global_load_u16 v13, v[1:2], off offset:20
; GFX11-NEXT: global_load_u16 v14, v[1:2], off offset:18
; GFX11-NEXT: global_load_u16 v15, v[1:2], off offset:16
; GFX11-NEXT: global_load_u16 v16, v[1:2], off offset:22
; GFX11-NEXT: global_load_u16 v17, v[1:2], off offset:26
; GFX11-NEXT: global_load_u16 v18, v[1:2], off offset:30
; GFX11-NEXT: global_load_u16 v19, v[1:2], off offset:44
; GFX11-NEXT: global_load_u16 v20, v[1:2], off offset:40
; GFX11-NEXT: global_load_u16 v21, v[1:2], off offset:36
; GFX11-NEXT: global_load_u16 v22, v[1:2], off offset:34
; GFX11-NEXT: global_load_u16 v23, v[1:2], off offset:32
; GFX11-NEXT: global_load_u16 v24, v[1:2], off offset:38
; GFX11-NEXT: global_load_u16 v25, v[1:2], off offset:42
; GFX11-NEXT: global_load_u16 v26, v[1:2], off offset:46
; GFX11-NEXT: global_load_u16 v27, v[1:2], off offset:60
; GFX11-NEXT: global_load_u16 v28, v[1:2], off offset:56
; GFX11-NEXT: global_load_u16 v29, v[1:2], off offset:52
; GFX11-NEXT: global_load_u16 v30, v[1:2], off offset:50
; GFX11-NEXT: global_load_u16 v31, v[1:2], off offset:48
; GFX11-NEXT: global_load_u16 v32, v[1:2], off offset:54
; GFX11-NEXT: global_load_u16 v33, v[1:2], off offset:58
; GFX11-NEXT: global_load_u16 v1, v[1:2], off offset:62
; GFX11-NEXT: s_waitcnt vmcnt(31)
; GFX11-NEXT: v_lshlrev_b32_e32 v39, 16, v3
; GFX11-NEXT: s_waitcnt vmcnt(30)
; GFX11-NEXT: v_lshlrev_b32_e32 v38, 16, v4
; GFX11-NEXT: s_waitcnt vmcnt(29)
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT: s_waitcnt vmcnt(28)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v6
; GFX11-NEXT: s_waitcnt vmcnt(27)
; GFX11-NEXT: v_lshlrev_b32_e32 v37, 16, v7
; GFX11-NEXT: s_waitcnt vmcnt(26)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v8
; GFX11-NEXT: s_waitcnt vmcnt(25)
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v9
; GFX11-NEXT: s_waitcnt vmcnt(24)
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT: s_waitcnt vmcnt(23)
; GFX11-NEXT: v_lshlrev_b32_e32 v102, 16, v11
; GFX11-NEXT: s_waitcnt vmcnt(22)
; GFX11-NEXT: v_lshlrev_b32_e32 v101, 16, v12
; GFX11-NEXT: s_waitcnt vmcnt(21)
; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v13
; GFX11-NEXT: s_waitcnt vmcnt(20)
; GFX11-NEXT: v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT: s_waitcnt vmcnt(19)
; GFX11-NEXT: v_lshlrev_b32_e32 v100, 16, v15
; GFX11-NEXT: s_waitcnt vmcnt(18)
; GFX11-NEXT: v_lshlrev_b32_e32 v34, 16, v16
; GFX11-NEXT: s_waitcnt vmcnt(17)
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v17
; GFX11-NEXT: s_waitcnt vmcnt(16)
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT: s_waitcnt vmcnt(15)
; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v19
; GFX11-NEXT: s_waitcnt vmcnt(14)
; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v20
; GFX11-NEXT: s_waitcnt vmcnt(13)
; GFX11-NEXT: v_lshlrev_b32_e32 v21, 16, v21
; GFX11-NEXT: s_waitcnt vmcnt(12)
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT: s_waitcnt vmcnt(11)
; GFX11-NEXT: v_lshlrev_b32_e32 v103, 16, v23
; GFX11-NEXT: s_waitcnt vmcnt(10)
; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v24
; GFX11-NEXT: s_waitcnt vmcnt(9)
; GFX11-NEXT: v_lshlrev_b32_e32 v25, 16, v25
; GFX11-NEXT: s_waitcnt vmcnt(8)
; GFX11-NEXT: v_lshlrev_b32_e32 v26, 16, v26
; GFX11-NEXT: s_waitcnt vmcnt(7)
; GFX11-NEXT: v_lshlrev_b32_e32 v68, 16, v27
; GFX11-NEXT: s_waitcnt vmcnt(6)
; GFX11-NEXT: v_lshlrev_b32_e32 v65, 16, v28
; GFX11-NEXT: s_waitcnt vmcnt(5)
; GFX11-NEXT: v_lshlrev_b32_e32 v29, 16, v29
; GFX11-NEXT: s_waitcnt vmcnt(4)
; GFX11-NEXT: v_lshlrev_b32_e32 v30, 16, v30
; GFX11-NEXT: s_waitcnt vmcnt(3)
; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v31
; GFX11-NEXT: s_waitcnt vmcnt(2)
; GFX11-NEXT: v_lshlrev_b32_e32 v64, 16, v32
; GFX11-NEXT: s_waitcnt vmcnt(1)
; GFX11-NEXT: v_lshlrev_b32_e32 v33, 16, v33
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_cvt_f64_f32_e32 v[96:97], v68
; GFX11-NEXT: v_cvt_f64_f32_e32 v[84:85], v65
; GFX11-NEXT: v_cvt_f64_f32_e32 v[82:83], v64
; GFX11-NEXT: v_cvt_f64_f32_e32 v[86:87], v33
; GFX11-NEXT: v_cvt_f64_f32_e32 v[98:99], v1
; GFX11-NEXT: v_cvt_f64_f32_e32 v[80:81], v29
; GFX11-NEXT: v_cvt_f64_f32_e32 v[70:71], v30
; GFX11-NEXT: v_cvt_f64_f32_e32 v[68:69], v53
; GFX11-NEXT: v_cvt_f64_f32_e32 v[66:67], v26
; GFX11-NEXT: v_cvt_f64_f32_e32 v[64:65], v52
; GFX11-NEXT: v_cvt_f64_f32_e32 v[54:55], v25
; GFX11-NEXT: v_cvt_f64_f32_e32 v[52:53], v49
; GFX11-NEXT: v_cvt_f64_f32_e32 v[50:51], v48
; GFX11-NEXT: v_cvt_f64_f32_e32 v[48:49], v21
; GFX11-NEXT: v_cvt_f64_f32_e32 v[23:24], v34
; GFX11-NEXT: v_cvt_f64_f32_e32 v[35:36], v22
; GFX11-NEXT: v_cvt_f64_f32_e32 v[33:34], v103
; GFX11-NEXT: v_cvt_f64_f32_e32 v[31:32], v18
; GFX11-NEXT: v_cvt_f64_f32_e32 v[29:30], v102
; GFX11-NEXT: v_cvt_f64_f32_e32 v[27:28], v17
; GFX11-NEXT: v_cvt_f64_f32_e32 v[25:26], v101
; GFX11-NEXT: v_cvt_f64_f32_e32 v[21:22], v13
; GFX11-NEXT: v_cvt_f64_f32_e32 v[19:20], v14
; GFX11-NEXT: v_cvt_f64_f32_e32 v[17:18], v100
; GFX11-NEXT: v_cvt_f64_f32_e32 v[15:16], v10
; GFX11-NEXT: v_cvt_f64_f32_e32 v[13:14], v39
; GFX11-NEXT: v_cvt_f64_f32_e32 v[11:12], v9
; GFX11-NEXT: v_cvt_f64_f32_e32 v[9:10], v38
; GFX11-NEXT: v_cvt_f64_f32_e32 v[7:8], v6
; GFX11-NEXT: v_cvt_f64_f32_e32 v[5:6], v5
; GFX11-NEXT: v_cvt_f64_f32_e32 v[3:4], v2
; GFX11-NEXT: v_cvt_f64_f32_e32 v[1:2], v37
; GFX11-NEXT: s_clause 0xf
; GFX11-NEXT: scratch_store_b128 v0, v[96:99], off offset:240
; GFX11-NEXT: scratch_store_b128 v0, v[84:87], off offset:224
; GFX11-NEXT: scratch_store_b128 v0, v[80:83], off offset:208
; GFX11-NEXT: scratch_store_b128 v0, v[68:71], off offset:192
; GFX11-NEXT: scratch_store_b128 v0, v[64:67], off offset:176
; GFX11-NEXT: scratch_store_b128 v0, v[52:55], off offset:160
; GFX11-NEXT: scratch_store_b128 v0, v[48:51], off offset:144
; GFX11-NEXT: scratch_store_b128 v0, v[33:36], off offset:128
; GFX11-NEXT: scratch_store_b128 v0, v[29:32], off offset:112
; GFX11-NEXT: scratch_store_b128 v0, v[25:28], off offset:96
; GFX11-NEXT: scratch_store_b128 v0, v[21:24], off offset:80
; GFX11-NEXT: scratch_store_b128 v0, v[17:20], off offset:64
; GFX11-NEXT: scratch_store_b128 v0, v[13:16], off offset:48
; GFX11-NEXT: scratch_store_b128 v0, v[9:12], off offset:32
; GFX11-NEXT: scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-NEXT: scratch_store_b128 v0, v[1:4], off
; GFX11-NEXT: s_setpc_b64 s[30:31]
%load = load <32 x bfloat>, ptr addrspace(1) %ptr
%fpext = fpext <32 x bfloat> %load to <32 x double>
ret <32 x double> %fpext
}
define bfloat @v_fadd_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fadd_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd bfloat %a, %b
ret bfloat %op
}
define <2 x bfloat> @v_fadd_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_fadd_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v1, v1, v3
; GCN-NEXT: v_add_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v1, v1, v3
; GFX7-NEXT: v_add_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_add_f32_e32 v2, v3, v2
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_add_f32_e32 v2, v3, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_add_f32_e32 v2, v3, v2
; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
; GFX11-NEXT: v_add_f32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd <2 x bfloat> %a, %b
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_fadd_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_fadd_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v2, v2, v5
; GCN-NEXT: v_add_f32_e32 v1, v1, v4
; GCN-NEXT: v_add_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
; GFX7-NEXT: v_add_f32_e32 v1, v1, v4
; GFX7-NEXT: v_add_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_add_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_add_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_add_f32_e32 v3, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_add_f32_e32 v4, v5, v4
; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fadd_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_add_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11TRUE16-NEXT: v_dual_add_f32 v0, v0, v2 :: v_dual_add_f32 v1, v1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fadd_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_add_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11FAKE16-NEXT: v_dual_add_f32 v0, v0, v2 :: v_dual_add_f32 v1, v1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fadd <3 x bfloat> %a, %b
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_fadd_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_fadd_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v3, v3, v7
; GCN-NEXT: v_add_f32_e32 v2, v2, v6
; GCN-NEXT: v_add_f32_e32 v1, v1, v5
; GCN-NEXT: v_add_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v3, v3, v7
; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
; GFX7-NEXT: v_add_f32_e32 v1, v1, v5
; GFX7-NEXT: v_add_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_add_f32_e32 v4, v5, v4
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_add_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v4, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_add_f32_e32 v4, v5, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_add_f32_e32 v3, v5, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_add_f32_e32 v4, v5, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_add_f32_e32 v3, v7, v6
; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_add_f32 v0, v0, v2 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add_f32_e32 v1, v1, v3
; GFX11-NEXT: v_dual_add_f32 v3, v7, v6 :: v_dual_add_f32 v4, v5, v4
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd <4 x bfloat> %a, %b
ret <4 x bfloat> %op
}
define <8 x bfloat> @v_fadd_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_fadd_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v7, v7, v15
; GCN-NEXT: v_add_f32_e32 v6, v6, v14
; GCN-NEXT: v_add_f32_e32 v5, v5, v13
; GCN-NEXT: v_add_f32_e32 v4, v4, v12
; GCN-NEXT: v_add_f32_e32 v3, v3, v11
; GCN-NEXT: v_add_f32_e32 v2, v2, v10
; GCN-NEXT: v_add_f32_e32 v1, v1, v9
; GCN-NEXT: v_add_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v7, v7, v15
; GFX7-NEXT: v_add_f32_e32 v6, v6, v14
; GFX7-NEXT: v_add_f32_e32 v5, v5, v13
; GFX7-NEXT: v_add_f32_e32 v4, v4, v12
; GFX7-NEXT: v_add_f32_e32 v3, v3, v11
; GFX7-NEXT: v_add_f32_e32 v2, v2, v10
; GFX7-NEXT: v_add_f32_e32 v1, v1, v9
; GFX7-NEXT: v_add_f32_e32 v0, v0, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX8-NEXT: v_add_f32_e32 v8, v9, v8
; GFX8-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v8
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
; GFX8-NEXT: v_add_f32_e32 v3, v3, v7
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX8-NEXT: v_add_f32_e32 v7, v9, v7
; GFX8-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v7
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_add_f32_e32 v2, v2, v6
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX8-NEXT: v_add_f32_e32 v6, v9, v6
; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_add_f32_e32 v1, v1, v5
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX8-NEXT: v_add_f32_e32 v5, v9, v5
; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_add_f32_e32 v0, v0, v4
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v5, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v7, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v8, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX9-NEXT: v_add_f32_e32 v8, v9, v8
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add_f32_e32 v3, v3, v7
; GFX9-NEXT: v_add3_u32 v9, v9, v8, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v7, v7, v3, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX9-NEXT: v_add_f32_e32 v7, v9, v7
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX9-NEXT: v_add_f32_e32 v2, v2, v6
; GFX9-NEXT: v_add3_u32 v9, v9, v7, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v6, v6, v2, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX9-NEXT: v_add_f32_e32 v6, v9, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX9-NEXT: v_add_f32_e32 v1, v1, v5
; GFX9-NEXT: v_add3_u32 v9, v9, v6, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v5, v5, v1, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX9-NEXT: v_add_f32_e32 v5, v9, v5
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v4
; GFX9-NEXT: v_add3_u32 v9, v9, v5, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v5, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v7, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v8, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_add_f32_e32 v8, v9, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_add_f32_e32 v3, v3, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX10-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX10-NEXT: v_add_f32_e32 v7, v10, v9
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
; GFX10-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX10-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX10-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX10-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_add_f32_e32 v6, v10, v6
; GFX10-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX10-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v5
; GFX10-NEXT: v_add_f32_e32 v5, v15, v13
; GFX10-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX10-NEXT: v_add_f32_e32 v0, v0, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX10-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v10, vcc_lo
; GFX10-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX10-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_add_f32 v8, v9, v8 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_f32_e32 v3, v3, v7
; GFX11-NEXT: v_add_f32_e32 v7, v10, v9
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_cndmask_b32 v7, v11, v12 :: v_dual_add_f32 v2, v2, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_f32_e32 v6, v10, v6
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v0, v0, v4
; GFX11-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX11-NEXT: v_dual_add_f32 v1, v1, v5 :: v_dual_cndmask_b32 v4, v4, v10
; GFX11-NEXT: v_add_f32_e32 v5, v15, v13
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX11-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX11-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd <8 x bfloat> %a, %b
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_fadd_v16bf16(<16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_fadd_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_add_f32_e32 v14, v14, v30
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_add_f32_e32 v13, v13, v29
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_add_f32_e32 v12, v12, v28
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_add_f32_e32 v11, v11, v27
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_add_f32_e32 v10, v10, v26
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_add_f32_e32 v9, v9, v25
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_add_f32_e32 v8, v8, v24
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_add_f32_e32 v7, v7, v23
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_add_f32_e32 v6, v6, v22
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_add_f32_e32 v5, v5, v21
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_add_f32_e32 v4, v4, v20
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v3, v3, v19
; GCN-NEXT: v_add_f32_e32 v2, v2, v18
; GCN-NEXT: v_add_f32_e32 v1, v1, v17
; GCN-NEXT: v_add_f32_e32 v0, v0, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_add_f32_e32 v15, v15, v16
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_add_f32_e32 v6, v6, v22
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v14, v14, v30
; GFX7-NEXT: v_add_f32_e32 v13, v13, v29
; GFX7-NEXT: v_add_f32_e32 v12, v12, v28
; GFX7-NEXT: v_add_f32_e32 v11, v11, v27
; GFX7-NEXT: v_add_f32_e32 v10, v10, v26
; GFX7-NEXT: v_add_f32_e32 v9, v9, v25
; GFX7-NEXT: v_add_f32_e32 v8, v8, v24
; GFX7-NEXT: v_add_f32_e32 v7, v7, v23
; GFX7-NEXT: v_add_f32_e32 v5, v5, v21
; GFX7-NEXT: v_add_f32_e32 v4, v4, v20
; GFX7-NEXT: v_add_f32_e32 v3, v3, v19
; GFX7-NEXT: v_add_f32_e32 v2, v2, v18
; GFX7-NEXT: v_add_f32_e32 v1, v1, v17
; GFX7-NEXT: v_add_f32_e32 v0, v0, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_add_f32_e32 v15, v15, v22
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX8-NEXT: v_add_f32_e32 v16, v17, v16
; GFX8-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v16
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v7, v7, v15
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX8-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v7
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX8-NEXT: v_add_f32_e32 v15, v17, v15
; GFX8-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v15
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v6, v6, v14
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v14, vcc, v14, v6
; GFX8-NEXT: v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX8-NEXT: v_add_f32_e32 v14, v17, v14
; GFX8-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v5, v5, v13
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v13, vcc, v13, v5
; GFX8-NEXT: v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX8-NEXT: v_add_f32_e32 v13, v17, v13
; GFX8-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v13
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v4, v4, v12
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v12, vcc, v12, v4
; GFX8-NEXT: v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX8-NEXT: v_add_f32_e32 v12, v17, v12
; GFX8-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v12
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v3, v3, v11
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v11, vcc, v11, v3
; GFX8-NEXT: v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX8-NEXT: v_add_f32_e32 v11, v17, v11
; GFX8-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v11
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v2, v2, v10
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v2
; GFX8-NEXT: v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX8-NEXT: v_add_f32_e32 v10, v17, v10
; GFX8-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v10
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v1, v1, v9
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX8-NEXT: v_add_f32_e32 v9, v17, v9
; GFX8-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v9
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_add_f32_e32 v0, v0, v8
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v0
; GFX8-NEXT: v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v9, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v10, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v11, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v12, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v13, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v14, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v15, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v16, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX9-NEXT: v_add_f32_e32 v16, v17, v16
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add_f32_e32 v7, v7, v15
; GFX9-NEXT: v_add3_u32 v17, v17, v16, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v15, v15, v7, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX9-NEXT: v_add_f32_e32 v15, v17, v15
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX9-NEXT: v_add_f32_e32 v6, v6, v14
; GFX9-NEXT: v_add3_u32 v17, v17, v15, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v14, v14, v6, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX9-NEXT: v_add_f32_e32 v14, v17, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX9-NEXT: v_add_f32_e32 v5, v5, v13
; GFX9-NEXT: v_add3_u32 v17, v17, v14, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v13, v13, v5, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX9-NEXT: v_add_f32_e32 v13, v17, v13
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX9-NEXT: v_add_f32_e32 v4, v4, v12
; GFX9-NEXT: v_add3_u32 v17, v17, v13, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v12, v12, v4, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX9-NEXT: v_add_f32_e32 v12, v17, v12
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX9-NEXT: v_add_f32_e32 v3, v3, v11
; GFX9-NEXT: v_add3_u32 v17, v17, v12, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v11, v11, v3, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX9-NEXT: v_add_f32_e32 v11, v17, v11
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX9-NEXT: v_add_f32_e32 v2, v2, v10
; GFX9-NEXT: v_add3_u32 v17, v17, v11, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v10, v10, v2, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX9-NEXT: v_add_f32_e32 v10, v17, v10
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX9-NEXT: v_add_f32_e32 v1, v1, v9
; GFX9-NEXT: v_add3_u32 v17, v17, v10, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v9, v9, v1, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX9-NEXT: v_add_f32_e32 v9, v17, v9
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v8
; GFX9-NEXT: v_add3_u32 v17, v17, v9, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v8, v8, v0, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v9, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v10, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v11, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v12, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v13, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v14, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v15, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v16, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_add_f32_e32 v16, v17, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX10-NEXT: v_add_f32_e32 v7, v7, v15
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX10-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT: v_add_f32_e32 v17, v18, v17
; GFX10-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX10-NEXT: v_add_f32_e32 v6, v6, v14
; GFX10-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX10-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v15, v15, v20, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v5
; GFX10-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX10-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX10-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX10-NEXT: v_add_f32_e32 v17, v20, v19
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX10-NEXT: v_add_f32_e32 v5, v5, v13
; GFX10-NEXT: v_cndmask_b32_e32 v14, v14, v16, vcc_lo
; GFX10-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX10-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v6, v16, v13, vcc_lo
; GFX10-NEXT: v_add_f32_e32 v13, v19, v18
; GFX10-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX10-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX10-NEXT: v_add_f32_e32 v4, v4, v12
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX10-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_add_f32_e32 v12, v18, v12
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX10-NEXT: v_add_f32_e32 v3, v3, v11
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX10-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX10-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX10-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX10-NEXT: v_add_f32_e32 v18, v19, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX10-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX10-NEXT: v_add_f32_e32 v2, v2, v10
; GFX10-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX10-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX10-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX10-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX10-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX10-NEXT: v_add_f32_e32 v19, v22, v20
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX10-NEXT: v_add_f32_e32 v1, v1, v9
; GFX10-NEXT: v_add_f32_e32 v9, v22, v20
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX10-NEXT: v_add_f32_e32 v0, v0, v8
; GFX10-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX10-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX10-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX10-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX10-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_add_f32 v16, v17, v16 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add_f32_e32 v17, v18, v17
; GFX11-NEXT: v_add_f32_e32 v6, v6, v14
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX11-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v7, v7, v15
; GFX11-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX11-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_cndmask_b32 v15, v15, v20 :: v_dual_lshlrev_b32 v20, 16, v5
; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_dual_add_f32 v17, v20, v19 :: v_dual_cndmask_b32 v14, v14, v16
; GFX11-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v4, v4, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_add_f32_e32 v5, v5, v13
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_cndmask_b32 v6, v16, v13 :: v_dual_add_f32 v13, v19, v18
; GFX11-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_add_f32_e32 v12, v18, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX11-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX11-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add_f32_e32 v18, v19, v18
; GFX11-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_add_f32 v2, v2, v10 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_add_f32_e32 v3, v3, v11
; GFX11-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX11-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX11-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX11-NEXT: v_add_f32_e32 v19, v22, v20
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX11-NEXT: v_dual_add_f32 v0, v0, v8 :: v_dual_add_f32 v1, v1, v9
; GFX11-NEXT: v_add_f32_e32 v9, v22, v20
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX11-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd <16 x bfloat> %a, %b
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_fadd_v32bf16(<32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_fadd_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:124
; GCN-NEXT: v_add_f32_e32 v31, v31, v32
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:120
; GCN-NEXT: v_add_f32_e32 v30, v30, v32
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
; GCN-NEXT: v_add_f32_e32 v29, v29, v32
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112
; GCN-NEXT: v_add_f32_e32 v28, v28, v32
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:108
; GCN-NEXT: v_add_f32_e32 v27, v27, v32
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
; GCN-NEXT: v_add_f32_e32 v26, v26, v32
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100
; GCN-NEXT: v_add_f32_e32 v25, v25, v32
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
; GCN-NEXT: v_add_f32_e32 v24, v24, v32
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:92
; GCN-NEXT: v_add_f32_e32 v23, v23, v32
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
; GCN-NEXT: v_add_f32_e32 v22, v22, v32
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:84
; GCN-NEXT: v_add_f32_e32 v21, v21, v32
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:80
; GCN-NEXT: v_add_f32_e32 v20, v20, v32
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:76
; GCN-NEXT: v_add_f32_e32 v19, v19, v32
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72
; GCN-NEXT: v_add_f32_e32 v18, v18, v32
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:68
; GCN-NEXT: v_add_f32_e32 v17, v17, v32
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:64
; GCN-NEXT: v_add_f32_e32 v16, v16, v32
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:60
; GCN-NEXT: v_add_f32_e32 v15, v15, v32
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:56
; GCN-NEXT: v_add_f32_e32 v14, v14, v32
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:52
; GCN-NEXT: v_add_f32_e32 v13, v13, v32
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:48
; GCN-NEXT: v_add_f32_e32 v12, v12, v32
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:44
; GCN-NEXT: v_add_f32_e32 v11, v11, v32
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
; GCN-NEXT: v_add_f32_e32 v10, v10, v32
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:36
; GCN-NEXT: v_add_f32_e32 v9, v9, v32
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
; GCN-NEXT: v_add_f32_e32 v8, v8, v32
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28
; GCN-NEXT: v_add_f32_e32 v7, v7, v32
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:24
; GCN-NEXT: v_add_f32_e32 v6, v6, v32
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
; GCN-NEXT: v_add_f32_e32 v5, v5, v32
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:16
; GCN-NEXT: v_add_f32_e32 v4, v4, v32
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:12
; GCN-NEXT: v_add_f32_e32 v3, v3, v32
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GCN-NEXT: v_add_f32_e32 v2, v2, v32
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
; GCN-NEXT: v_add_f32_e32 v1, v1, v32
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_add_f32_e32 v0, v0, v32
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: v_add_f32_e32 v31, v31, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v30, v30, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v29, v29, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v28, v28, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v27, v27, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v26, v26, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v25, v25, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v24, v24, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v23, v23, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v22, v22, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v21, v21, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v20, v20, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v19, v19, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v18, v18, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v17, v17, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v16, v16, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v15, v15, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v14, v14, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v13, v13, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v12, v12, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v11, v11, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v10, v10, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v9, v9, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v8, v8, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v7, v7, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v6, v6, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v5, v5, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v4, v4, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v3, v3, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v2, v2, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v1, v1, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_add_f32_e32 v0, v0, v32
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX8-NEXT: v_add_f32_e32 v31, v32, v31
; GFX8-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v32, vcc, v32, v31
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_add_u32_e32 v32, vcc, s4, v32
; GFX8-NEXT: v_add_f32_e32 v14, v14, v30
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX8-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v30, vcc, v30, v14
; GFX8-NEXT: v_add_u32_e32 v30, vcc, s4, v30
; GFX8-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX8-NEXT: v_add_f32_e32 v32, v32, v30
; GFX8-NEXT: buffer_load_dword v30, off, s[0:3], s32
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_add_f32_e32 v13, v13, v29
; GFX8-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX8-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX8-NEXT: v_alignbit_b32 v14, v14, v31, 16
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; GFX8-NEXT: v_add_f32_e32 v33, v33, v34
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_add_f32_e32 v30, v15, v30
; GFX8-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v33
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX8-NEXT: v_bfe_u32 v33, v30, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v30
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v30
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX8-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v32
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX8-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v29, vcc, v29, v13
; GFX8-NEXT: v_add_u32_e32 v29, vcc, s4, v29
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_cndmask_b32_e32 v13, v29, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v12
; GFX8-NEXT: v_add_f32_e32 v29, v33, v29
; GFX8-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v29
; GFX8-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v12, v12, v28
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v28, vcc, v28, v12
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX8-NEXT: v_add_f32_e32 v28, v33, v28
; GFX8-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v28
; GFX8-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v11, v11, v27
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v27, vcc, v27, v11
; GFX8-NEXT: v_add_u32_e32 v27, vcc, s4, v27
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX8-NEXT: v_add_f32_e32 v27, v33, v27
; GFX8-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v27
; GFX8-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v10, v10, v26
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v26, vcc, v26, v10
; GFX8-NEXT: v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX8-NEXT: v_add_f32_e32 v26, v33, v26
; GFX8-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v26
; GFX8-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v9, v9, v25
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v25, vcc, v25, v9
; GFX8-NEXT: v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX8-NEXT: v_add_f32_e32 v25, v33, v25
; GFX8-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v25
; GFX8-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v8, v8, v24
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v24, vcc, v24, v8
; GFX8-NEXT: v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX8-NEXT: v_add_f32_e32 v24, v33, v24
; GFX8-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v24
; GFX8-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v7, v7, v23
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v23, vcc, v23, v7
; GFX8-NEXT: v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX8-NEXT: v_add_f32_e32 v23, v33, v23
; GFX8-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v23
; GFX8-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v6, v6, v22
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v22, vcc, v22, v6
; GFX8-NEXT: v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX8-NEXT: v_add_f32_e32 v22, v33, v22
; GFX8-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v22
; GFX8-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v5, v5, v21
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v21, vcc, v21, v5
; GFX8-NEXT: v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX8-NEXT: v_add_f32_e32 v21, v33, v21
; GFX8-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v21
; GFX8-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v4, v4, v20
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v20, vcc, v20, v4
; GFX8-NEXT: v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX8-NEXT: v_add_f32_e32 v20, v33, v20
; GFX8-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v20
; GFX8-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v3, v3, v19
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v19, vcc, v19, v3
; GFX8-NEXT: v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX8-NEXT: v_add_f32_e32 v19, v33, v19
; GFX8-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v19
; GFX8-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v2, v2, v18
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v18, vcc, v18, v2
; GFX8-NEXT: v_add_u32_e32 v18, vcc, s4, v18
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX8-NEXT: v_add_f32_e32 v18, v33, v18
; GFX8-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v18
; GFX8-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v1, v1, v17
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX8-NEXT: v_add_f32_e32 v17, v33, v17
; GFX8-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v17
; GFX8-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_add_f32_e32 v0, v0, v16
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v16, vcc, v16, v0
; GFX8-NEXT: v_add_u32_e32 v16, vcc, s4, v16
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v30
; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX8-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX8-NEXT: v_alignbit_b32 v0, v0, v17, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v18, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v19, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v20, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v21, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v22, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v23, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v24, 16
; GFX8-NEXT: v_alignbit_b32 v8, v8, v25, 16
; GFX8-NEXT: v_alignbit_b32 v9, v9, v26, 16
; GFX8-NEXT: v_alignbit_b32 v10, v10, v27, 16
; GFX8-NEXT: v_alignbit_b32 v11, v11, v28, 16
; GFX8-NEXT: v_alignbit_b32 v12, v12, v29, 16
; GFX8-NEXT: v_alignbit_b32 v13, v13, v32, 16
; GFX8-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT: v_add_f32_e32 v31, v32, v31
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX9-NEXT: v_add_f32_e32 v14, v14, v30
; GFX9-NEXT: v_add3_u32 v32, v32, v31, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v30, v30, v14, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT: v_add_f32_e32 v30, v32, v30
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_bfe_u32 v32, v30, 16, 1
; GFX9-NEXT: v_add_f32_e32 v13, v13, v29
; GFX9-NEXT: v_add3_u32 v32, v32, v30, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v30
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v29, v29, v13, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_cndmask_b32_e32 v13, v29, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT: v_add_f32_e32 v32, v32, v29
; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_add_f32_e32 v12, v12, v28
; GFX9-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX9-NEXT: v_add3_u32 v28, v28, v12, s4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v29
; GFX9-NEXT: v_add_f32_e32 v33, v33, v34
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_add_f32_e32 v29, v15, v29
; GFX9-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX9-NEXT: v_add3_u32 v15, v15, v33, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX9-NEXT: v_add3_u32 v33, v33, v29, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT: v_add3_u32 v33, v33, v32, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX9-NEXT: v_add_f32_e32 v28, v33, v28
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT: v_add_f32_e32 v11, v11, v27
; GFX9-NEXT: v_add3_u32 v33, v33, v28, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v27, v27, v11, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX9-NEXT: v_add_f32_e32 v27, v33, v27
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT: v_add_f32_e32 v10, v10, v26
; GFX9-NEXT: v_add3_u32 v33, v33, v27, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v26, v26, v10, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX9-NEXT: v_add_f32_e32 v26, v33, v26
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT: v_add_f32_e32 v9, v9, v25
; GFX9-NEXT: v_add3_u32 v33, v33, v26, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v25, v25, v9, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX9-NEXT: v_add_f32_e32 v25, v33, v25
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT: v_add_f32_e32 v8, v8, v24
; GFX9-NEXT: v_add3_u32 v33, v33, v25, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v24, v24, v8, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX9-NEXT: v_add_f32_e32 v24, v33, v24
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT: v_add_f32_e32 v7, v7, v23
; GFX9-NEXT: v_add3_u32 v33, v33, v24, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v23, v23, v7, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX9-NEXT: v_add_f32_e32 v23, v33, v23
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT: v_add_f32_e32 v6, v6, v22
; GFX9-NEXT: v_add3_u32 v33, v33, v23, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v22, v22, v6, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX9-NEXT: v_add_f32_e32 v22, v33, v22
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT: v_add_f32_e32 v5, v5, v21
; GFX9-NEXT: v_add3_u32 v33, v33, v22, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v21, v21, v5, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX9-NEXT: v_add_f32_e32 v21, v33, v21
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT: v_add_f32_e32 v4, v4, v20
; GFX9-NEXT: v_add3_u32 v33, v33, v21, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v20, v20, v4, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX9-NEXT: v_add_f32_e32 v20, v33, v20
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT: v_add_f32_e32 v3, v3, v19
; GFX9-NEXT: v_add3_u32 v33, v33, v20, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v19, v19, v3, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX9-NEXT: v_add_f32_e32 v19, v33, v19
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT: v_add_f32_e32 v2, v2, v18
; GFX9-NEXT: v_add3_u32 v33, v33, v19, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v18, v18, v2, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX9-NEXT: v_add_f32_e32 v18, v33, v18
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT: v_add_f32_e32 v1, v1, v17
; GFX9-NEXT: v_add3_u32 v33, v33, v18, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v17, v17, v1, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX9-NEXT: v_add_f32_e32 v17, v33, v17
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT: v_add_f32_e32 v0, v0, v16
; GFX9-NEXT: v_add3_u32 v33, v33, v17, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v16, v16, v0, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v17, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v18, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v19, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v20, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v21, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v22, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v23, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v24, s4
; GFX9-NEXT: v_perm_b32 v8, v8, v25, s4
; GFX9-NEXT: v_perm_b32 v9, v9, v26, s4
; GFX9-NEXT: v_perm_b32 v10, v10, v27, s4
; GFX9-NEXT: v_perm_b32 v11, v11, v28, s4
; GFX9-NEXT: v_perm_b32 v12, v12, v32, s4
; GFX9-NEXT: v_perm_b32 v13, v13, v30, s4
; GFX9-NEXT: v_perm_b32 v14, v14, v31, s4
; GFX9-NEXT: v_perm_b32 v15, v29, v15, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX10-NEXT: v_lshlrev_b32_e32 v39, 16, v27
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX10-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX10-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_lshlrev_b32_e32 v37, 16, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v12
; GFX10-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX10-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX10-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX10-NEXT: v_lshlrev_b32_e32 v54, 16, v8
; GFX10-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX10-NEXT: v_lshlrev_b32_e32 v64, 16, v7
; GFX10-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX10-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX10-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX10-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX10-NEXT: v_add_f32_e32 v39, v48, v39
; GFX10-NEXT: v_add_f32_e32 v11, v11, v27
; GFX10-NEXT: v_add_f32_e32 v49, v50, v49
; GFX10-NEXT: v_add_f32_e32 v10, v10, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX10-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_add_f32_e32 v37, v38, v37
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v18
; GFX10-NEXT: v_add_f32_e32 v12, v12, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v2
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v17
; GFX10-NEXT: v_lshlrev_b32_e32 v27, 16, v1
; GFX10-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v26, 16, v0
; GFX10-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_add_f32_e32 v9, v9, v25
; GFX10-NEXT: v_add_f32_e32 v25, v54, v53
; GFX10-NEXT: v_add_f32_e32 v8, v8, v24
; GFX10-NEXT: v_add_f32_e32 v24, v64, v55
; GFX10-NEXT: v_add_f32_e32 v7, v7, v23
; GFX10-NEXT: v_add_f32_e32 v23, v66, v65
; GFX10-NEXT: v_add_f32_e32 v6, v6, v22
; GFX10-NEXT: v_add_f32_e32 v22, v68, v67
; GFX10-NEXT: v_bfe_u32 v53, v39, 16, 1
; GFX10-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX10-NEXT: v_bfe_u32 v65, v49, 16, 1
; GFX10-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v33, 16, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v14
; GFX10-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_add_f32_e32 v35, v36, v35
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v19
; GFX10-NEXT: v_add_f32_e32 v13, v13, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_add_f32_e32 v2, v2, v18
; GFX10-NEXT: v_add_f32_e32 v18, v27, v48
; GFX10-NEXT: v_add_f32_e32 v1, v1, v17
; GFX10-NEXT: v_add_f32_e32 v17, v26, v50
; GFX10-NEXT: v_add_f32_e32 v0, v0, v16
; GFX10-NEXT: v_or_b32_e32 v54, 0x400000, v39
; GFX10-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX10-NEXT: v_or_b32_e32 v66, 0x400000, v49
; GFX10-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v39, v39
; GFX10-NEXT: v_add3_u32 v39, v53, v39, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v11, v11
; GFX10-NEXT: v_add3_u32 v11, v55, v11, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v49, v49
; GFX10-NEXT: v_add3_u32 v49, v65, v49, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v10, v10
; GFX10-NEXT: v_add3_u32 v10, v67, v10, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_add_f32_e32 v33, v34, v33
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v20
; GFX10-NEXT: v_add_f32_e32 v14, v14, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v30, 16, v4
; GFX10-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_add_f32_e32 v3, v3, v19
; GFX10-NEXT: v_add_f32_e32 v19, v28, v38
; GFX10-NEXT: v_bfe_u32 v38, v37, 16, 1
; GFX10-NEXT: v_bfe_u32 v50, v12, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v39, v39, v54, s9
; GFX10-NEXT: v_bfe_u32 v54, v18, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v11, v11, v64, s10
; GFX10-NEXT: v_bfe_u32 v64, v1, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v49, v49, v66, s11
; GFX10-NEXT: v_bfe_u32 v66, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v10, v10, v68, s12
; GFX10-NEXT: v_bfe_u32 v68, v0, 16, 1
; GFX10-NEXT: v_add_f32_e32 v51, v52, v51
; GFX10-NEXT: v_add_f32_e32 v5, v5, v21
; GFX10-NEXT: v_add_f32_e32 v21, v30, v34
; GFX10-NEXT: v_add_f32_e32 v4, v4, v20
; GFX10-NEXT: v_add_f32_e32 v20, v29, v36
; GFX10-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX10-NEXT: v_bfe_u32 v27, v14, 16, 1
; GFX10-NEXT: v_bfe_u32 v29, v35, 16, 1
; GFX10-NEXT: v_bfe_u32 v34, v13, 16, 1
; GFX10-NEXT: v_or_b32_e32 v48, 0x400000, v37
; GFX10-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v37, v37
; GFX10-NEXT: v_add3_u32 v37, v38, v37, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v12, v12
; GFX10-NEXT: v_add3_u32 v12, v50, v12, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v18, v18
; GFX10-NEXT: v_add3_u32 v54, v54, v18, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v18
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v1, v1
; GFX10-NEXT: v_add3_u32 v64, v64, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v17, v17
; GFX10-NEXT: v_add3_u32 v66, v66, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s22, v0, v0
; GFX10-NEXT: v_add3_u32 v68, v68, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v0, 0x400000, v0
; GFX10-NEXT: v_or_b32_e32 v26, 0x400000, v33
; GFX10-NEXT: v_or_b32_e32 v28, 0x400000, v14
; GFX10-NEXT: v_or_b32_e32 v30, 0x400000, v35
; GFX10-NEXT: v_or_b32_e32 v36, 0x400000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX10-NEXT: v_bfe_u32 v33, v51, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v14, v14
; GFX10-NEXT: v_add3_u32 v14, v27, v14, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v35, v35
; GFX10-NEXT: v_add3_u32 v29, v29, v35, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v13, v13
; GFX10-NEXT: v_add3_u32 v13, v34, v13, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v24, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v37, v37, v48, s7
; GFX10-NEXT: v_bfe_u32 v48, v19, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v12, v12, v52, s8
; GFX10-NEXT: v_bfe_u32 v52, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v18, v54, v18, s10
; GFX10-NEXT: v_cndmask_b32_e64 v17, v66, v17, s12
; GFX10-NEXT: v_cndmask_b32_e64 v0, v68, v0, s22
; GFX10-NEXT: v_cndmask_b32_e64 v1, v64, v1, s11
; GFX10-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_or_b32_e32 v27, 0x400000, v51
; GFX10-NEXT: v_bfe_u32 v35, v9, 16, 1
; GFX10-NEXT: v_bfe_u32 v38, v25, 16, 1
; GFX10-NEXT: v_or_b32_e32 v67, 0x400000, v24
; GFX10-NEXT: v_cmp_u_f32_e64 s13, v51, v51
; GFX10-NEXT: v_add3_u32 v33, v33, v51, 0x7fff
; GFX10-NEXT: v_bfe_u32 v51, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s17, v24, v24
; GFX10-NEXT: v_add3_u32 v24, v65, v24, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v6, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v26, vcc_lo
; GFX10-NEXT: v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v14, v14, v28, s4
; GFX10-NEXT: v_bfe_u32 v28, v4, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v29, v29, v30, s5
; GFX10-NEXT: v_bfe_u32 v30, v20, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v13, v13, v36, s6
; GFX10-NEXT: v_bfe_u32 v36, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v19, v19
; GFX10-NEXT: v_add3_u32 v48, v48, v19, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v19
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v2, v2
; GFX10-NEXT: v_add3_u32 v52, v52, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX10-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX10-NEXT: v_or_b32_e32 v34, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v50, 0x400000, v25
; GFX10-NEXT: v_bfe_u32 v53, v8, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s14, v9, v9
; GFX10-NEXT: v_add3_u32 v9, v35, v9, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v35, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e64 s15, v25, v25
; GFX10-NEXT: v_add3_u32 v25, v38, v25, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v23, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s18, v7, v7
; GFX10-NEXT: v_add3_u32 v7, v51, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v51, 0x400000, v6
; GFX10-NEXT: v_cmp_u_f32_e64 s20, v6, v6
; GFX10-NEXT: v_add3_u32 v6, v65, v6, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v21, v21
; GFX10-NEXT: v_add3_u32 v26, v26, v21, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v21
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v4, v4
; GFX10-NEXT: v_add3_u32 v28, v28, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v4
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v20, v20
; GFX10-NEXT: v_add3_u32 v30, v30, v20, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v20
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v3, v3
; GFX10-NEXT: v_add3_u32 v36, v36, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e64 v19, v48, v19, s8
; GFX10-NEXT: v_cndmask_b32_e64 v2, v52, v2, s9
; GFX10-NEXT: v_or_b32_e32 v55, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e64 s16, v8, v8
; GFX10-NEXT: v_add3_u32 v8, v53, v8, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v53, 0x400000, v23
; GFX10-NEXT: v_cmp_u_f32_e64 s19, v23, v23
; GFX10-NEXT: v_add3_u32 v23, v38, v23, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v22, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v65, v65, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e64 v21, v26, v21, s4
; GFX10-NEXT: v_cndmask_b32_e64 v4, v28, v4, s5
; GFX10-NEXT: v_cndmask_b32_e64 v20, v30, v20, s6
; GFX10-NEXT: v_cndmask_b32_e64 v3, v36, v3, s7
; GFX10-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX10-NEXT: v_cmp_u_f32_e64 s21, v22, v22
; GFX10-NEXT: v_add3_u32 v38, v38, v22, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v22
; GFX10-NEXT: v_cndmask_b32_e32 v5, v65, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX10-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e64 v27, v33, v27, s13
; GFX10-NEXT: v_cndmask_b32_e64 v9, v9, v34, s14
; GFX10-NEXT: v_cndmask_b32_e64 v25, v25, v50, s15
; GFX10-NEXT: v_cndmask_b32_e64 v8, v8, v55, s16
; GFX10-NEXT: v_cndmask_b32_e64 v24, v24, v67, s17
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v35, s18
; GFX10-NEXT: v_cndmask_b32_e64 v23, v23, v53, s19
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v51, s20
; GFX10-NEXT: v_cndmask_b32_e64 v22, v38, v22, s21
; GFX10-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX10-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX10-NEXT: v_perm_b32 v9, v9, v27, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX10-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX10-NEXT: v_perm_b32 v10, v10, v49, 0x7060302
; GFX10-NEXT: v_perm_b32 v11, v11, v39, 0x7060302
; GFX10-NEXT: v_perm_b32 v12, v12, v37, 0x7060302
; GFX10-NEXT: v_perm_b32 v13, v13, v29, 0x7060302
; GFX10-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v32
; GFX10-NEXT: v_add_f32_e32 v17, v31, v17
; GFX10-NEXT: v_add_f32_e32 v15, v15, v18
; GFX10-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX10-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v15, v15
; GFX10-NEXT: v_add3_u32 v17, v18, v17, 0x7fff
; GFX10-NEXT: v_add3_u32 v15, v19, v15, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v15, v15, v21, s4
; GFX10-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_load_b32 v32, off, s32
; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX11-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v1
; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX11-NEXT: v_dual_add_f32 v5, v5, v21 :: v_dual_and_b32 v26, 0xffff0000, v26
; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_add_f32 v1, v1, v17 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v19
; GFX11-NEXT: v_bfe_u32 v103, v5, 16, 1
; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v18
; GFX11-NEXT: v_bfe_u32 v135, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v5
; GFX11-NEXT: v_or_b32_e32 v144, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v103, v103, v5, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_add3_u32 v135, v135, v1, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_add_f32 v3, v3, v19 :: v_dual_lshlrev_b32 v54, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v16
; GFX11-NEXT: v_dual_add_f32 v19, v82, v81 :: v_dual_lshlrev_b32 v64, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v129, v19, 16, 1
; GFX11-NEXT: v_or_b32_e32 v130, 0x400000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX11-NEXT: v_bfe_u32 v119, v3, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX11-NEXT: v_add3_u32 v129, v129, v19, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v0
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_add_f32 v17, v86, v85 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: v_dual_add_f32 v8, v8, v24 :: v_dual_lshlrev_b32 v39, 16, v27
; GFX11-NEXT: v_or_b32_e32 v128, 0x400000, v3
; GFX11-NEXT: v_add3_u32 v119, v119, v3, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v145, v17, 16, 1
; GFX11-NEXT: v_or_b32_e32 v146, 0x400000, v17
; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-NEXT: v_lshlrev_b32_e32 v70, 16, v4
; GFX11-NEXT: v_add3_u32 v145, v145, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX11-NEXT: v_add_f32_e32 v2, v2, v18
; GFX11-NEXT: v_add_f32_e32 v0, v0, v16
; GFX11-NEXT: v_dual_add_f32 v24, v64, v55 :: v_dual_lshlrev_b32 v37, 16, v28
; GFX11-NEXT: v_add_f32_e32 v7, v7, v23
; GFX11-NEXT: v_dual_add_f32 v23, v66, v65 :: v_dual_add_f32 v18, v84, v83
; GFX11-NEXT: v_dual_add_f32 v9, v9, v25 :: v_dual_and_b32 v28, 0xffff0000, v28
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v85, v24, 16, 1
; GFX11-NEXT: v_bfe_u32 v97, v23, 16, 1
; GFX11-NEXT: v_or_b32_e32 v86, 0x400000, v24
; GFX11-NEXT: v_or_b32_e32 v98, 0x400000, v23
; GFX11-NEXT: v_bfe_u32 v87, v7, 16, 1
; GFX11-NEXT: v_add3_u32 v85, v85, v24, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v20
; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-NEXT: v_add3_u32 v97, v97, v23, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v7
; GFX11-NEXT: v_add3_u32 v87, v87, v7, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_add_f32_e32 v4, v4, v20
; GFX11-NEXT: v_add_f32_e32 v20, v80, v71
; GFX11-NEXT: v_bfe_u32 v71, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX11-NEXT: v_dual_add_f32 v21, v70, v69 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v71, v71, v9, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-NEXT: v_dual_add_f32 v10, v10, v26 :: v_dual_and_b32 v29, 0xffff0000, v29
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-NEXT: v_add_f32_e32 v26, v52, v51
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_add_f32_e32 v6, v6, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_dual_add_f32 v11, v11, v27 :: v_dual_lshlrev_b32 v34, 16, v14
; GFX11-NEXT: v_dual_add_f32 v22, v68, v67 :: v_dual_lshlrev_b32 v33, 16, v30
; GFX11-NEXT: v_dual_add_f32 v27, v50, v49 :: v_dual_lshlrev_b32 v38, 16, v12
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_dual_add_f32 v25, v54, v53 :: v_dual_and_b32 v12, 0xffff0000, v12
; GFX11-NEXT: v_dual_add_f32 v13, v13, v29 :: v_dual_and_b32 v30, 0xffff0000, v30
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_f32_e32 v29, v38, v37
; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX11-NEXT: v_dual_add_f32 v12, v12, v28 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_f32_e32 v14, v14, v30
; GFX11-NEXT: v_add_f32_e32 v28, v48, v39
; GFX11-NEXT: v_dual_add_f32 v30, v36, v35 :: v_dual_add_f32 v33, v34, v33
; GFX11-NEXT: v_bfe_u32 v39, v13, 16, 1
; GFX11-NEXT: v_bfe_u32 v35, v14, 16, 1
; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v14
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v37, v30, 16, 1
; GFX11-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-NEXT: v_add3_u32 v35, v35, v14, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
; GFX11-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX11-NEXT: v_add3_u32 v37, v37, v30, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v13
; GFX11-NEXT: v_bfe_u32 v49, v29, 16, 1
; GFX11-NEXT: v_add3_u32 v39, v39, v13, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v34, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v29
; GFX11-NEXT: v_bfe_u32 v51, v12, 16, 1
; GFX11-NEXT: v_add3_u32 v49, v49, v29, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX11-NEXT: v_cndmask_b32_e32 v14, v35, v36, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-NEXT: v_bfe_u32 v53, v28, 16, 1
; GFX11-NEXT: v_add3_u32 v51, v51, v12, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v28
; GFX11-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v30, v37, v38, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_add3_u32 v53, v53, v28, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX11-NEXT: v_bfe_u32 v65, v27, 16, 1
; GFX11-NEXT: v_add3_u32 v55, v55, v11, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v13, v39, v48, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v27
; GFX11-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX11-NEXT: v_add3_u32 v65, v65, v27, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX11-NEXT: v_cndmask_b32_e32 v29, v49, v50, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_bfe_u32 v69, v26, 16, 1
; GFX11-NEXT: v_add3_u32 v67, v67, v10, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v26
; GFX11-NEXT: v_bfe_u32 v81, v25, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v12, v51, v52, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-NEXT: v_add3_u32 v69, v69, v26, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v25
; GFX11-NEXT: v_bfe_u32 v83, v8, 16, 1
; GFX11-NEXT: v_add3_u32 v81, v81, v25, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v83, v83, v8, 0x7fff
; GFX11-NEXT: v_bfe_u32 v99, v6, 16, 1
; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v6
; GFX11-NEXT: v_cndmask_b32_e32 v11, v55, v64, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-NEXT: v_bfe_u32 v101, v22, 16, 1
; GFX11-NEXT: v_add3_u32 v99, v99, v6, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v22
; GFX11-NEXT: v_bfe_u32 v113, v21, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v27, v65, v66, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-NEXT: v_add3_u32 v101, v101, v22, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v114, 0x400000, v21
; GFX11-NEXT: v_bfe_u32 v115, v4, 16, 1
; GFX11-NEXT: v_add3_u32 v113, v113, v21, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v10, v67, v68, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-NEXT: v_or_b32_e32 v116, 0x400000, v4
; GFX11-NEXT: v_bfe_u32 v117, v20, 16, 1
; GFX11-NEXT: v_add3_u32 v115, v115, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v118, 0x400000, v20
; GFX11-NEXT: v_cndmask_b32_e32 v26, v69, v70, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v117, v117, v20, 0x7fff
; GFX11-NEXT: v_bfe_u32 v133, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v134, 0x400000, v18
; GFX11-NEXT: v_bfe_u32 v147, v0, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v9, v71, v80, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-NEXT: v_add3_u32 v133, v133, v18, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v147, v147, v0, 0x7fff
; GFX11-NEXT: v_bfe_u32 v131, v2, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v25, v81, v82, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_or_b32_e32 v132, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v9, v9, v26, 0x7060302
; GFX11-NEXT: v_add3_u32 v131, v131, v2, 0x7fff
; GFX11-NEXT: v_perm_b32 v10, v10, v27, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v83, v84, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-NEXT: v_perm_b32 v11, v11, v28, 0x7060302
; GFX11-NEXT: v_perm_b32 v12, v12, v29, 0x7060302
; GFX11-NEXT: v_perm_b32 v13, v13, v30, 0x7060302
; GFX11-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v24, v85, v86, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v7, v87, v96, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v23, v97, v98, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_cndmask_b32_e32 v6, v99, v100, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v22, v101, v102, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_cndmask_b32_e32 v5, v103, v112, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v21, v113, v114, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_cndmask_b32_e32 v4, v115, v116, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v20, v117, v118, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_cndmask_b32_e32 v19, v129, v130, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_cndmask_b32_e32 v18, v133, v134, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v135, v144, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v17, v145, v146, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v147, v33, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v131, v132, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v119, v128, vcc_lo
; GFX11-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_add_f32 v17, v31, v17 :: v_dual_and_b32 v18, 0xffff0000, v32
; GFX11-NEXT: v_add_f32_e32 v15, v15, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX11-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v17, v18, v20, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fadd <32 x bfloat> %a, %b
ret <32 x bfloat> %op
}
define bfloat @v_fadd_bf16_fpimm_0(bfloat %arg0) {
; GCN-LABEL: v_fadd_bf16_fpimm_0:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_bf16_fpimm_0:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_bf16_fpimm_0:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_add_f32_e32 v0, 1.0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_bf16_fpimm_0:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_bf16_fpimm_0:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_add_f32_e32 v0, 1.0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_bf16_fpimm_0:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v0, 1.0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%add = fadd bfloat %arg0, 1.0
ret bfloat %add
}
define bfloat @v_fadd_bf16_fpimm_1(bfloat %arg0) {
; GCN-LABEL: v_fadd_bf16_fpimm_1:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fadd_bf16_fpimm_1:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fadd_bf16_fpimm_1:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fadd_bf16_fpimm_1:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fadd_bf16_fpimm_1:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fadd_bf16_fpimm_1:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v0, 0x42280000, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%add = fadd bfloat %arg0, 42.0
ret bfloat %add
}
define bfloat @v_fsub_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fsub_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_sub_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fsub_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fsub_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fsub_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fsub_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fsub_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fsub bfloat %a, %b
ret bfloat %op
}
define <2 x bfloat> @v_fsub_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_fsub_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_sub_f32_e32 v1, v1, v3
; GCN-NEXT: v_sub_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fsub_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fsub_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v2, v3, v2
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fsub_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_sub_f32_e32 v2, v3, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fsub_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_sub_f32_e32 v2, v3, v2
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fsub_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX11-NEXT: v_sub_f32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fsub <2 x bfloat> %a, %b
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_fsub_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_fsub_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_sub_f32_e32 v2, v2, v5
; GCN-NEXT: v_sub_f32_e32 v1, v1, v4
; GCN-NEXT: v_sub_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fsub_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_sub_f32_e32 v2, v2, v5
; GFX7-NEXT: v_sub_f32_e32 v1, v1, v4
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fsub_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fsub_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_sub_f32_e32 v3, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fsub_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_sub_f32_e32 v4, v5, v4
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX10-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fsub_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_sub_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11TRUE16-NEXT: v_dual_sub_f32 v0, v0, v2 :: v_dual_sub_f32 v1, v1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fsub_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_sub_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11FAKE16-NEXT: v_dual_sub_f32 v0, v0, v2 :: v_dual_sub_f32 v1, v1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fsub <3 x bfloat> %a, %b
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_fsub_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_fsub_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_sub_f32_e32 v3, v3, v7
; GCN-NEXT: v_sub_f32_e32 v2, v2, v6
; GCN-NEXT: v_sub_f32_e32 v1, v1, v5
; GCN-NEXT: v_sub_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fsub_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_sub_f32_e32 v3, v3, v7
; GFX7-NEXT: v_sub_f32_e32 v2, v2, v6
; GFX7-NEXT: v_sub_f32_e32 v1, v1, v5
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fsub_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_sub_f32_e32 v4, v5, v4
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v4, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fsub_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_sub_f32_e32 v4, v5, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_sub_f32_e32 v3, v5, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fsub_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_sub_f32_e32 v4, v5, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX10-NEXT: v_sub_f32_e32 v3, v7, v6
; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fsub_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_sub_f32 v0, v0, v2 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX11-NEXT: v_dual_sub_f32 v3, v7, v6 :: v_dual_sub_f32 v4, v5, v4
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fsub <4 x bfloat> %a, %b
ret <4 x bfloat> %op
}
define bfloat @v_fmul_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fmul_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul bfloat %a, %b
ret bfloat %op
}
define <2 x bfloat> @v_fmul_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_fmul_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v1, v1, v3
; GCN-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v2, v3, v2
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v2, v3, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v2, v3, v2
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: v_mul_f32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul <2 x bfloat> %a, %b
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_fmul_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_fmul_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v2, v2, v5
; GCN-NEXT: v_mul_f32_e32 v1, v1, v4
; GCN-NEXT: v_mul_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v5
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v4
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v3, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_mul_f32_e32 v4, v5, v4
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fmul_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_mul_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fmul_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_mul_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11FAKE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_mul_f32 v1, v1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fmul <3 x bfloat> %a, %b
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_fmul_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_fmul_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v3, v3, v7
; GCN-NEXT: v_mul_f32_e32 v2, v2, v6
; GCN-NEXT: v_mul_f32_e32 v1, v1, v5
; GCN-NEXT: v_mul_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v6
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v5
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v4, v5, v4
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v4, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v4, v5, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v4, v5, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX10-NEXT: v_mul_f32_e32 v3, v7, v6
; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX11-NEXT: v_dual_mul_f32 v3, v7, v6 :: v_dual_mul_f32 v4, v5, v4
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul <4 x bfloat> %a, %b
ret <4 x bfloat> %op
}
define <8 x bfloat> @v_fmul_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_fmul_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v7, v7, v15
; GCN-NEXT: v_mul_f32_e32 v6, v6, v14
; GCN-NEXT: v_mul_f32_e32 v5, v5, v13
; GCN-NEXT: v_mul_f32_e32 v4, v4, v12
; GCN-NEXT: v_mul_f32_e32 v3, v3, v11
; GCN-NEXT: v_mul_f32_e32 v2, v2, v10
; GCN-NEXT: v_mul_f32_e32 v1, v1, v9
; GCN-NEXT: v_mul_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v7, v7, v15
; GFX7-NEXT: v_mul_f32_e32 v6, v6, v14
; GFX7-NEXT: v_mul_f32_e32 v5, v5, v13
; GFX7-NEXT: v_mul_f32_e32 v4, v4, v12
; GFX7-NEXT: v_mul_f32_e32 v3, v3, v11
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v10
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v9
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX8-NEXT: v_mul_f32_e32 v8, v9, v8
; GFX8-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v8
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
; GFX8-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX8-NEXT: v_mul_f32_e32 v7, v9, v7
; GFX8-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v7
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_mul_f32_e32 v2, v2, v6
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v6, v9, v6
; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v5
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v5, v9, v5
; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v5, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v7, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v8, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX9-NEXT: v_mul_f32_e32 v8, v9, v8
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX9-NEXT: v_add3_u32 v9, v9, v8, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v7, v7, v3, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX9-NEXT: v_mul_f32_e32 v7, v9, v7
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v2, v2, v6
; GFX9-NEXT: v_add3_u32 v9, v9, v7, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v6, v6, v2, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v6, v9, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v5
; GFX9-NEXT: v_add3_u32 v9, v9, v6, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v5, v5, v1, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v5, v9, v5
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX9-NEXT: v_add3_u32 v9, v9, v5, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v5, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v7, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v8, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_mul_f32_e32 v8, v9, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX10-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX10-NEXT: v_mul_f32_e32 v7, v10, v9
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT: v_mul_f32_e32 v2, v2, v6
; GFX10-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX10-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX10-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX10-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_mul_f32_e32 v6, v10, v6
; GFX10-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX10-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v5
; GFX10-NEXT: v_mul_f32_e32 v5, v15, v13
; GFX10-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX10-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v10, vcc_lo
; GFX10-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX10-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_mul_f32 v8, v9, v8 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX11-NEXT: v_mul_f32_e32 v7, v10, v9
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_cndmask_b32 v7, v11, v12 :: v_dual_mul_f32 v2, v2, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e32 v6, v10, v6
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX11-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX11-NEXT: v_dual_mul_f32 v1, v1, v5 :: v_dual_cndmask_b32 v4, v4, v10
; GFX11-NEXT: v_mul_f32_e32 v5, v15, v13
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX11-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX11-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul <8 x bfloat> %a, %b
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_fmul_v16bf16(<16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_fmul_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_mul_f32_e32 v14, v14, v30
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_mul_f32_e32 v13, v13, v29
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_mul_f32_e32 v12, v12, v28
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_mul_f32_e32 v11, v11, v27
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_mul_f32_e32 v10, v10, v26
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_mul_f32_e32 v9, v9, v25
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_mul_f32_e32 v8, v8, v24
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_mul_f32_e32 v7, v7, v23
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_mul_f32_e32 v6, v6, v22
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_mul_f32_e32 v5, v5, v21
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_mul_f32_e32 v4, v4, v20
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v3, v3, v19
; GCN-NEXT: v_mul_f32_e32 v2, v2, v18
; GCN-NEXT: v_mul_f32_e32 v1, v1, v17
; GCN-NEXT: v_mul_f32_e32 v0, v0, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_mul_f32_e32 v15, v15, v16
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_mul_f32_e32 v6, v6, v22
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v14, v14, v30
; GFX7-NEXT: v_mul_f32_e32 v13, v13, v29
; GFX7-NEXT: v_mul_f32_e32 v12, v12, v28
; GFX7-NEXT: v_mul_f32_e32 v11, v11, v27
; GFX7-NEXT: v_mul_f32_e32 v10, v10, v26
; GFX7-NEXT: v_mul_f32_e32 v9, v9, v25
; GFX7-NEXT: v_mul_f32_e32 v8, v8, v24
; GFX7-NEXT: v_mul_f32_e32 v7, v7, v23
; GFX7-NEXT: v_mul_f32_e32 v5, v5, v21
; GFX7-NEXT: v_mul_f32_e32 v4, v4, v20
; GFX7-NEXT: v_mul_f32_e32 v3, v3, v19
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v18
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v17
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_mul_f32_e32 v15, v15, v22
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX8-NEXT: v_mul_f32_e32 v16, v17, v16
; GFX8-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v16
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v7, v7, v15
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX8-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v7
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX8-NEXT: v_mul_f32_e32 v15, v17, v15
; GFX8-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v15
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v6, v6, v14
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v14, vcc, v14, v6
; GFX8-NEXT: v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX8-NEXT: v_mul_f32_e32 v14, v17, v14
; GFX8-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v5, v5, v13
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v13, vcc, v13, v5
; GFX8-NEXT: v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX8-NEXT: v_mul_f32_e32 v13, v17, v13
; GFX8-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v13
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v4, v4, v12
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v12, vcc, v12, v4
; GFX8-NEXT: v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX8-NEXT: v_mul_f32_e32 v12, v17, v12
; GFX8-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v12
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v3, v3, v11
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v11, vcc, v11, v3
; GFX8-NEXT: v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX8-NEXT: v_mul_f32_e32 v11, v17, v11
; GFX8-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v11
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v2, v2, v10
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v2
; GFX8-NEXT: v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v10, v17, v10
; GFX8-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v10
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v9
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v9, v17, v9
; GFX8-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v9
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v8
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v0
; GFX8-NEXT: v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v9, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v10, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v11, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v12, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v13, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v14, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v15, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v16, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX9-NEXT: v_mul_f32_e32 v16, v17, v16
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mul_f32_e32 v7, v7, v15
; GFX9-NEXT: v_add3_u32 v17, v17, v16, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v15, v15, v7, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX9-NEXT: v_mul_f32_e32 v15, v17, v15
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v6, v6, v14
; GFX9-NEXT: v_add3_u32 v17, v17, v15, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v14, v14, v6, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX9-NEXT: v_mul_f32_e32 v14, v17, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v5, v5, v13
; GFX9-NEXT: v_add3_u32 v17, v17, v14, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v13, v13, v5, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX9-NEXT: v_mul_f32_e32 v13, v17, v13
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v4, v4, v12
; GFX9-NEXT: v_add3_u32 v17, v17, v13, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v12, v12, v4, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX9-NEXT: v_mul_f32_e32 v12, v17, v12
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v3, v3, v11
; GFX9-NEXT: v_add3_u32 v17, v17, v12, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v11, v11, v3, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX9-NEXT: v_mul_f32_e32 v11, v17, v11
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v2, v2, v10
; GFX9-NEXT: v_add3_u32 v17, v17, v11, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v10, v10, v2, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v10, v17, v10
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v9
; GFX9-NEXT: v_add3_u32 v17, v17, v10, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v9, v9, v1, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v9, v17, v9
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v8
; GFX9-NEXT: v_add3_u32 v17, v17, v9, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v8, v8, v0, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v9, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v10, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v11, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v12, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v13, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v14, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v15, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v16, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_mul_f32_e32 v16, v17, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX10-NEXT: v_mul_f32_e32 v7, v7, v15
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX10-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT: v_mul_f32_e32 v17, v18, v17
; GFX10-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX10-NEXT: v_mul_f32_e32 v6, v6, v14
; GFX10-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX10-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v15, v15, v20, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v5
; GFX10-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX10-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX10-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX10-NEXT: v_mul_f32_e32 v17, v20, v19
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX10-NEXT: v_mul_f32_e32 v5, v5, v13
; GFX10-NEXT: v_cndmask_b32_e32 v14, v14, v16, vcc_lo
; GFX10-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX10-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v6, v16, v13, vcc_lo
; GFX10-NEXT: v_mul_f32_e32 v13, v19, v18
; GFX10-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX10-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX10-NEXT: v_mul_f32_e32 v4, v4, v12
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX10-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_mul_f32_e32 v12, v18, v12
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX10-NEXT: v_mul_f32_e32 v3, v3, v11
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX10-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX10-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX10-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX10-NEXT: v_mul_f32_e32 v18, v19, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX10-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX10-NEXT: v_mul_f32_e32 v2, v2, v10
; GFX10-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX10-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX10-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX10-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX10-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX10-NEXT: v_mul_f32_e32 v19, v22, v20
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v9
; GFX10-NEXT: v_mul_f32_e32 v9, v22, v20
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v8
; GFX10-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX10-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX10-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX10-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX10-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_mul_f32 v16, v17, v16 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_mul_f32_e32 v17, v18, v17
; GFX11-NEXT: v_mul_f32_e32 v6, v6, v14
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX11-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v7, v7, v15
; GFX11-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX11-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_cndmask_b32 v15, v15, v20 :: v_dual_lshlrev_b32 v20, 16, v5
; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_dual_mul_f32 v17, v20, v19 :: v_dual_cndmask_b32 v14, v14, v16
; GFX11-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v4, v4, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_mul_f32_e32 v5, v5, v13
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_cndmask_b32 v6, v16, v13 :: v_dual_mul_f32 v13, v19, v18
; GFX11-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_mul_f32_e32 v12, v18, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX11-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX11-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_mul_f32_e32 v18, v19, v18
; GFX11-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_mul_f32 v2, v2, v10 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_mul_f32_e32 v3, v3, v11
; GFX11-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX11-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX11-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX11-NEXT: v_mul_f32_e32 v19, v22, v20
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX11-NEXT: v_dual_mul_f32 v0, v0, v8 :: v_dual_mul_f32 v1, v1, v9
; GFX11-NEXT: v_mul_f32_e32 v9, v22, v20
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX11-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul <16 x bfloat> %a, %b
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_fmul_v32bf16(<32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_fmul_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:124
; GCN-NEXT: v_mul_f32_e32 v31, v31, v32
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:120
; GCN-NEXT: v_mul_f32_e32 v30, v30, v32
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
; GCN-NEXT: v_mul_f32_e32 v29, v29, v32
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112
; GCN-NEXT: v_mul_f32_e32 v28, v28, v32
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:108
; GCN-NEXT: v_mul_f32_e32 v27, v27, v32
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
; GCN-NEXT: v_mul_f32_e32 v26, v26, v32
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100
; GCN-NEXT: v_mul_f32_e32 v25, v25, v32
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
; GCN-NEXT: v_mul_f32_e32 v24, v24, v32
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:92
; GCN-NEXT: v_mul_f32_e32 v23, v23, v32
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
; GCN-NEXT: v_mul_f32_e32 v22, v22, v32
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:84
; GCN-NEXT: v_mul_f32_e32 v21, v21, v32
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:80
; GCN-NEXT: v_mul_f32_e32 v20, v20, v32
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:76
; GCN-NEXT: v_mul_f32_e32 v19, v19, v32
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72
; GCN-NEXT: v_mul_f32_e32 v18, v18, v32
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:68
; GCN-NEXT: v_mul_f32_e32 v17, v17, v32
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:64
; GCN-NEXT: v_mul_f32_e32 v16, v16, v32
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:60
; GCN-NEXT: v_mul_f32_e32 v15, v15, v32
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:56
; GCN-NEXT: v_mul_f32_e32 v14, v14, v32
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:52
; GCN-NEXT: v_mul_f32_e32 v13, v13, v32
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:48
; GCN-NEXT: v_mul_f32_e32 v12, v12, v32
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:44
; GCN-NEXT: v_mul_f32_e32 v11, v11, v32
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
; GCN-NEXT: v_mul_f32_e32 v10, v10, v32
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:36
; GCN-NEXT: v_mul_f32_e32 v9, v9, v32
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
; GCN-NEXT: v_mul_f32_e32 v8, v8, v32
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28
; GCN-NEXT: v_mul_f32_e32 v7, v7, v32
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:24
; GCN-NEXT: v_mul_f32_e32 v6, v6, v32
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
; GCN-NEXT: v_mul_f32_e32 v5, v5, v32
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:16
; GCN-NEXT: v_mul_f32_e32 v4, v4, v32
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:12
; GCN-NEXT: v_mul_f32_e32 v3, v3, v32
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GCN-NEXT: v_mul_f32_e32 v2, v2, v32
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
; GCN-NEXT: v_mul_f32_e32 v1, v1, v32
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_mul_f32_e32 v0, v0, v32
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmul_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: v_mul_f32_e32 v31, v31, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v30, v30, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v29, v29, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v28, v28, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v27, v27, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v26, v26, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v25, v25, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v24, v24, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v23, v23, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v22, v22, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v21, v21, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v20, v20, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v19, v19, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v18, v18, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v17, v17, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v16, v16, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v15, v15, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v14, v14, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v13, v13, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v12, v12, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v11, v11, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v10, v10, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v9, v9, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v8, v8, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v7, v7, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v6, v6, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v5, v5, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v4, v4, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v3, v3, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v32
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmul_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX8-NEXT: v_mul_f32_e32 v31, v32, v31
; GFX8-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v32, vcc, v32, v31
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_add_u32_e32 v32, vcc, s4, v32
; GFX8-NEXT: v_mul_f32_e32 v14, v14, v30
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX8-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v30, vcc, v30, v14
; GFX8-NEXT: v_add_u32_e32 v30, vcc, s4, v30
; GFX8-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX8-NEXT: v_mul_f32_e32 v32, v32, v30
; GFX8-NEXT: buffer_load_dword v30, off, s[0:3], s32
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_mul_f32_e32 v13, v13, v29
; GFX8-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX8-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX8-NEXT: v_alignbit_b32 v14, v14, v31, 16
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; GFX8-NEXT: v_mul_f32_e32 v33, v33, v34
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_mul_f32_e32 v30, v15, v30
; GFX8-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v33
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX8-NEXT: v_bfe_u32 v33, v30, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v30
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v30
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX8-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v32
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX8-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v29, vcc, v29, v13
; GFX8-NEXT: v_add_u32_e32 v29, vcc, s4, v29
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_cndmask_b32_e32 v13, v29, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v12
; GFX8-NEXT: v_mul_f32_e32 v29, v33, v29
; GFX8-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v29
; GFX8-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v12, v12, v28
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v28, vcc, v28, v12
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX8-NEXT: v_mul_f32_e32 v28, v33, v28
; GFX8-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v28
; GFX8-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v11, v11, v27
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v27, vcc, v27, v11
; GFX8-NEXT: v_add_u32_e32 v27, vcc, s4, v27
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX8-NEXT: v_mul_f32_e32 v27, v33, v27
; GFX8-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v27
; GFX8-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v10, v10, v26
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v26, vcc, v26, v10
; GFX8-NEXT: v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX8-NEXT: v_mul_f32_e32 v26, v33, v26
; GFX8-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v26
; GFX8-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v9, v9, v25
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v25, vcc, v25, v9
; GFX8-NEXT: v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX8-NEXT: v_mul_f32_e32 v25, v33, v25
; GFX8-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v25
; GFX8-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v8, v8, v24
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v24, vcc, v24, v8
; GFX8-NEXT: v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX8-NEXT: v_mul_f32_e32 v24, v33, v24
; GFX8-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v24
; GFX8-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v7, v7, v23
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v23, vcc, v23, v7
; GFX8-NEXT: v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX8-NEXT: v_mul_f32_e32 v23, v33, v23
; GFX8-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v23
; GFX8-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v6, v6, v22
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v22, vcc, v22, v6
; GFX8-NEXT: v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX8-NEXT: v_mul_f32_e32 v22, v33, v22
; GFX8-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v22
; GFX8-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v5, v5, v21
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v21, vcc, v21, v5
; GFX8-NEXT: v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX8-NEXT: v_mul_f32_e32 v21, v33, v21
; GFX8-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v21
; GFX8-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v4, v4, v20
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v20, vcc, v20, v4
; GFX8-NEXT: v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX8-NEXT: v_mul_f32_e32 v20, v33, v20
; GFX8-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v20
; GFX8-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v3, v3, v19
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v19, vcc, v19, v3
; GFX8-NEXT: v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX8-NEXT: v_mul_f32_e32 v19, v33, v19
; GFX8-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v19
; GFX8-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v2, v2, v18
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v18, vcc, v18, v2
; GFX8-NEXT: v_add_u32_e32 v18, vcc, s4, v18
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v18, v33, v18
; GFX8-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v18
; GFX8-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v17
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v17, v33, v17
; GFX8-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v17
; GFX8-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v16
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v16, vcc, v16, v0
; GFX8-NEXT: v_add_u32_e32 v16, vcc, s4, v16
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v30
; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX8-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX8-NEXT: v_alignbit_b32 v0, v0, v17, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v18, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v19, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v20, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v21, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v22, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v23, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v24, 16
; GFX8-NEXT: v_alignbit_b32 v8, v8, v25, 16
; GFX8-NEXT: v_alignbit_b32 v9, v9, v26, 16
; GFX8-NEXT: v_alignbit_b32 v10, v10, v27, 16
; GFX8-NEXT: v_alignbit_b32 v11, v11, v28, 16
; GFX8-NEXT: v_alignbit_b32 v12, v12, v29, 16
; GFX8-NEXT: v_alignbit_b32 v13, v13, v32, 16
; GFX8-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmul_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT: v_mul_f32_e32 v31, v32, v31
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v14, v14, v30
; GFX9-NEXT: v_add3_u32 v32, v32, v31, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v30, v30, v14, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT: v_mul_f32_e32 v30, v32, v30
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_bfe_u32 v32, v30, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v13, v13, v29
; GFX9-NEXT: v_add3_u32 v32, v32, v30, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v30
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v29, v29, v13, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_cndmask_b32_e32 v13, v29, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT: v_mul_f32_e32 v32, v32, v29
; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_mul_f32_e32 v12, v12, v28
; GFX9-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX9-NEXT: v_add3_u32 v28, v28, v12, s4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v29
; GFX9-NEXT: v_mul_f32_e32 v33, v33, v34
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_mul_f32_e32 v29, v15, v29
; GFX9-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX9-NEXT: v_add3_u32 v15, v15, v33, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX9-NEXT: v_add3_u32 v33, v33, v29, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT: v_add3_u32 v33, v33, v32, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX9-NEXT: v_mul_f32_e32 v28, v33, v28
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v11, v11, v27
; GFX9-NEXT: v_add3_u32 v33, v33, v28, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v27, v27, v11, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX9-NEXT: v_mul_f32_e32 v27, v33, v27
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v10, v10, v26
; GFX9-NEXT: v_add3_u32 v33, v33, v27, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v26, v26, v10, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX9-NEXT: v_mul_f32_e32 v26, v33, v26
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v9, v9, v25
; GFX9-NEXT: v_add3_u32 v33, v33, v26, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v25, v25, v9, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX9-NEXT: v_mul_f32_e32 v25, v33, v25
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v8, v8, v24
; GFX9-NEXT: v_add3_u32 v33, v33, v25, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v24, v24, v8, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX9-NEXT: v_mul_f32_e32 v24, v33, v24
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v7, v7, v23
; GFX9-NEXT: v_add3_u32 v33, v33, v24, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v23, v23, v7, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX9-NEXT: v_mul_f32_e32 v23, v33, v23
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v6, v6, v22
; GFX9-NEXT: v_add3_u32 v33, v33, v23, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v22, v22, v6, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX9-NEXT: v_mul_f32_e32 v22, v33, v22
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v5, v5, v21
; GFX9-NEXT: v_add3_u32 v33, v33, v22, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v21, v21, v5, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX9-NEXT: v_mul_f32_e32 v21, v33, v21
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v4, v4, v20
; GFX9-NEXT: v_add3_u32 v33, v33, v21, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v20, v20, v4, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX9-NEXT: v_mul_f32_e32 v20, v33, v20
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v3, v3, v19
; GFX9-NEXT: v_add3_u32 v33, v33, v20, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v19, v19, v3, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX9-NEXT: v_mul_f32_e32 v19, v33, v19
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v2, v2, v18
; GFX9-NEXT: v_add3_u32 v33, v33, v19, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v18, v18, v2, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v18, v33, v18
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v17
; GFX9-NEXT: v_add3_u32 v33, v33, v18, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v17, v17, v1, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v17, v33, v17
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v16
; GFX9-NEXT: v_add3_u32 v33, v33, v17, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v16, v16, v0, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v17, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v18, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v19, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v20, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v21, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v22, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v23, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v24, s4
; GFX9-NEXT: v_perm_b32 v8, v8, v25, s4
; GFX9-NEXT: v_perm_b32 v9, v9, v26, s4
; GFX9-NEXT: v_perm_b32 v10, v10, v27, s4
; GFX9-NEXT: v_perm_b32 v11, v11, v28, s4
; GFX9-NEXT: v_perm_b32 v12, v12, v32, s4
; GFX9-NEXT: v_perm_b32 v13, v13, v30, s4
; GFX9-NEXT: v_perm_b32 v14, v14, v31, s4
; GFX9-NEXT: v_perm_b32 v15, v29, v15, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmul_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX10-NEXT: v_lshlrev_b32_e32 v39, 16, v27
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX10-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX10-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_lshlrev_b32_e32 v37, 16, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v12
; GFX10-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX10-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX10-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX10-NEXT: v_lshlrev_b32_e32 v54, 16, v8
; GFX10-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX10-NEXT: v_lshlrev_b32_e32 v64, 16, v7
; GFX10-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX10-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX10-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX10-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX10-NEXT: v_mul_f32_e32 v39, v48, v39
; GFX10-NEXT: v_mul_f32_e32 v11, v11, v27
; GFX10-NEXT: v_mul_f32_e32 v49, v50, v49
; GFX10-NEXT: v_mul_f32_e32 v10, v10, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX10-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_mul_f32_e32 v37, v38, v37
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v18
; GFX10-NEXT: v_mul_f32_e32 v12, v12, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v2
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v17
; GFX10-NEXT: v_lshlrev_b32_e32 v27, 16, v1
; GFX10-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v26, 16, v0
; GFX10-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v9, v9, v25
; GFX10-NEXT: v_mul_f32_e32 v25, v54, v53
; GFX10-NEXT: v_mul_f32_e32 v8, v8, v24
; GFX10-NEXT: v_mul_f32_e32 v24, v64, v55
; GFX10-NEXT: v_mul_f32_e32 v7, v7, v23
; GFX10-NEXT: v_mul_f32_e32 v23, v66, v65
; GFX10-NEXT: v_mul_f32_e32 v6, v6, v22
; GFX10-NEXT: v_mul_f32_e32 v22, v68, v67
; GFX10-NEXT: v_bfe_u32 v53, v39, 16, 1
; GFX10-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX10-NEXT: v_bfe_u32 v65, v49, 16, 1
; GFX10-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v33, 16, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v14
; GFX10-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_mul_f32_e32 v35, v36, v35
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v19
; GFX10-NEXT: v_mul_f32_e32 v13, v13, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_mul_f32_e32 v2, v2, v18
; GFX10-NEXT: v_mul_f32_e32 v18, v27, v48
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v17
; GFX10-NEXT: v_mul_f32_e32 v17, v26, v50
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v16
; GFX10-NEXT: v_or_b32_e32 v54, 0x400000, v39
; GFX10-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX10-NEXT: v_or_b32_e32 v66, 0x400000, v49
; GFX10-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v39, v39
; GFX10-NEXT: v_add3_u32 v39, v53, v39, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v11, v11
; GFX10-NEXT: v_add3_u32 v11, v55, v11, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v49, v49
; GFX10-NEXT: v_add3_u32 v49, v65, v49, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v10, v10
; GFX10-NEXT: v_add3_u32 v10, v67, v10, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_mul_f32_e32 v33, v34, v33
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v20
; GFX10-NEXT: v_mul_f32_e32 v14, v14, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v30, 16, v4
; GFX10-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_mul_f32_e32 v3, v3, v19
; GFX10-NEXT: v_mul_f32_e32 v19, v28, v38
; GFX10-NEXT: v_bfe_u32 v38, v37, 16, 1
; GFX10-NEXT: v_bfe_u32 v50, v12, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v39, v39, v54, s9
; GFX10-NEXT: v_bfe_u32 v54, v18, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v11, v11, v64, s10
; GFX10-NEXT: v_bfe_u32 v64, v1, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v49, v49, v66, s11
; GFX10-NEXT: v_bfe_u32 v66, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v10, v10, v68, s12
; GFX10-NEXT: v_bfe_u32 v68, v0, 16, 1
; GFX10-NEXT: v_mul_f32_e32 v51, v52, v51
; GFX10-NEXT: v_mul_f32_e32 v5, v5, v21
; GFX10-NEXT: v_mul_f32_e32 v21, v30, v34
; GFX10-NEXT: v_mul_f32_e32 v4, v4, v20
; GFX10-NEXT: v_mul_f32_e32 v20, v29, v36
; GFX10-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX10-NEXT: v_bfe_u32 v27, v14, 16, 1
; GFX10-NEXT: v_bfe_u32 v29, v35, 16, 1
; GFX10-NEXT: v_bfe_u32 v34, v13, 16, 1
; GFX10-NEXT: v_or_b32_e32 v48, 0x400000, v37
; GFX10-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v37, v37
; GFX10-NEXT: v_add3_u32 v37, v38, v37, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v12, v12
; GFX10-NEXT: v_add3_u32 v12, v50, v12, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v18, v18
; GFX10-NEXT: v_add3_u32 v54, v54, v18, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v18
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v1, v1
; GFX10-NEXT: v_add3_u32 v64, v64, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v17, v17
; GFX10-NEXT: v_add3_u32 v66, v66, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s22, v0, v0
; GFX10-NEXT: v_add3_u32 v68, v68, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v0, 0x400000, v0
; GFX10-NEXT: v_or_b32_e32 v26, 0x400000, v33
; GFX10-NEXT: v_or_b32_e32 v28, 0x400000, v14
; GFX10-NEXT: v_or_b32_e32 v30, 0x400000, v35
; GFX10-NEXT: v_or_b32_e32 v36, 0x400000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX10-NEXT: v_bfe_u32 v33, v51, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v14, v14
; GFX10-NEXT: v_add3_u32 v14, v27, v14, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v35, v35
; GFX10-NEXT: v_add3_u32 v29, v29, v35, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v13, v13
; GFX10-NEXT: v_add3_u32 v13, v34, v13, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v24, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v37, v37, v48, s7
; GFX10-NEXT: v_bfe_u32 v48, v19, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v12, v12, v52, s8
; GFX10-NEXT: v_bfe_u32 v52, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v18, v54, v18, s10
; GFX10-NEXT: v_cndmask_b32_e64 v17, v66, v17, s12
; GFX10-NEXT: v_cndmask_b32_e64 v0, v68, v0, s22
; GFX10-NEXT: v_cndmask_b32_e64 v1, v64, v1, s11
; GFX10-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_or_b32_e32 v27, 0x400000, v51
; GFX10-NEXT: v_bfe_u32 v35, v9, 16, 1
; GFX10-NEXT: v_bfe_u32 v38, v25, 16, 1
; GFX10-NEXT: v_or_b32_e32 v67, 0x400000, v24
; GFX10-NEXT: v_cmp_u_f32_e64 s13, v51, v51
; GFX10-NEXT: v_add3_u32 v33, v33, v51, 0x7fff
; GFX10-NEXT: v_bfe_u32 v51, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s17, v24, v24
; GFX10-NEXT: v_add3_u32 v24, v65, v24, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v6, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v26, vcc_lo
; GFX10-NEXT: v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v14, v14, v28, s4
; GFX10-NEXT: v_bfe_u32 v28, v4, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v29, v29, v30, s5
; GFX10-NEXT: v_bfe_u32 v30, v20, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v13, v13, v36, s6
; GFX10-NEXT: v_bfe_u32 v36, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v19, v19
; GFX10-NEXT: v_add3_u32 v48, v48, v19, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v19
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v2, v2
; GFX10-NEXT: v_add3_u32 v52, v52, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX10-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX10-NEXT: v_or_b32_e32 v34, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v50, 0x400000, v25
; GFX10-NEXT: v_bfe_u32 v53, v8, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s14, v9, v9
; GFX10-NEXT: v_add3_u32 v9, v35, v9, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v35, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e64 s15, v25, v25
; GFX10-NEXT: v_add3_u32 v25, v38, v25, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v23, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s18, v7, v7
; GFX10-NEXT: v_add3_u32 v7, v51, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v51, 0x400000, v6
; GFX10-NEXT: v_cmp_u_f32_e64 s20, v6, v6
; GFX10-NEXT: v_add3_u32 v6, v65, v6, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v21, v21
; GFX10-NEXT: v_add3_u32 v26, v26, v21, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v21
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v4, v4
; GFX10-NEXT: v_add3_u32 v28, v28, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v4
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v20, v20
; GFX10-NEXT: v_add3_u32 v30, v30, v20, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v20
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v3, v3
; GFX10-NEXT: v_add3_u32 v36, v36, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e64 v19, v48, v19, s8
; GFX10-NEXT: v_cndmask_b32_e64 v2, v52, v2, s9
; GFX10-NEXT: v_or_b32_e32 v55, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e64 s16, v8, v8
; GFX10-NEXT: v_add3_u32 v8, v53, v8, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v53, 0x400000, v23
; GFX10-NEXT: v_cmp_u_f32_e64 s19, v23, v23
; GFX10-NEXT: v_add3_u32 v23, v38, v23, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v22, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v65, v65, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e64 v21, v26, v21, s4
; GFX10-NEXT: v_cndmask_b32_e64 v4, v28, v4, s5
; GFX10-NEXT: v_cndmask_b32_e64 v20, v30, v20, s6
; GFX10-NEXT: v_cndmask_b32_e64 v3, v36, v3, s7
; GFX10-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX10-NEXT: v_cmp_u_f32_e64 s21, v22, v22
; GFX10-NEXT: v_add3_u32 v38, v38, v22, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v22
; GFX10-NEXT: v_cndmask_b32_e32 v5, v65, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX10-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e64 v27, v33, v27, s13
; GFX10-NEXT: v_cndmask_b32_e64 v9, v9, v34, s14
; GFX10-NEXT: v_cndmask_b32_e64 v25, v25, v50, s15
; GFX10-NEXT: v_cndmask_b32_e64 v8, v8, v55, s16
; GFX10-NEXT: v_cndmask_b32_e64 v24, v24, v67, s17
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v35, s18
; GFX10-NEXT: v_cndmask_b32_e64 v23, v23, v53, s19
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v51, s20
; GFX10-NEXT: v_cndmask_b32_e64 v22, v38, v22, s21
; GFX10-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX10-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX10-NEXT: v_perm_b32 v9, v9, v27, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX10-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX10-NEXT: v_perm_b32 v10, v10, v49, 0x7060302
; GFX10-NEXT: v_perm_b32 v11, v11, v39, 0x7060302
; GFX10-NEXT: v_perm_b32 v12, v12, v37, 0x7060302
; GFX10-NEXT: v_perm_b32 v13, v13, v29, 0x7060302
; GFX10-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v32
; GFX10-NEXT: v_mul_f32_e32 v17, v31, v17
; GFX10-NEXT: v_mul_f32_e32 v15, v15, v18
; GFX10-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX10-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v15, v15
; GFX10-NEXT: v_add3_u32 v17, v18, v17, 0x7fff
; GFX10-NEXT: v_add3_u32 v15, v19, v15, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v15, v15, v21, s4
; GFX10-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmul_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_load_b32 v32, off, s32
; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX11-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v1
; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX11-NEXT: v_dual_mul_f32 v5, v5, v21 :: v_dual_and_b32 v26, 0xffff0000, v26
; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_mul_f32 v1, v1, v17 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v19
; GFX11-NEXT: v_bfe_u32 v103, v5, 16, 1
; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v18
; GFX11-NEXT: v_bfe_u32 v135, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v5
; GFX11-NEXT: v_or_b32_e32 v144, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v103, v103, v5, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_add3_u32 v135, v135, v1, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_mul_f32 v3, v3, v19 :: v_dual_lshlrev_b32 v54, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v16
; GFX11-NEXT: v_dual_mul_f32 v19, v82, v81 :: v_dual_lshlrev_b32 v64, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v129, v19, 16, 1
; GFX11-NEXT: v_or_b32_e32 v130, 0x400000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX11-NEXT: v_bfe_u32 v119, v3, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX11-NEXT: v_add3_u32 v129, v129, v19, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v0
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_mul_f32 v17, v86, v85 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: v_dual_mul_f32 v8, v8, v24 :: v_dual_lshlrev_b32 v39, 16, v27
; GFX11-NEXT: v_or_b32_e32 v128, 0x400000, v3
; GFX11-NEXT: v_add3_u32 v119, v119, v3, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v145, v17, 16, 1
; GFX11-NEXT: v_or_b32_e32 v146, 0x400000, v17
; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-NEXT: v_lshlrev_b32_e32 v70, 16, v4
; GFX11-NEXT: v_add3_u32 v145, v145, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX11-NEXT: v_mul_f32_e32 v2, v2, v18
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v16
; GFX11-NEXT: v_dual_mul_f32 v24, v64, v55 :: v_dual_lshlrev_b32 v37, 16, v28
; GFX11-NEXT: v_mul_f32_e32 v7, v7, v23
; GFX11-NEXT: v_dual_mul_f32 v23, v66, v65 :: v_dual_mul_f32 v18, v84, v83
; GFX11-NEXT: v_dual_mul_f32 v9, v9, v25 :: v_dual_and_b32 v28, 0xffff0000, v28
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v85, v24, 16, 1
; GFX11-NEXT: v_bfe_u32 v97, v23, 16, 1
; GFX11-NEXT: v_or_b32_e32 v86, 0x400000, v24
; GFX11-NEXT: v_or_b32_e32 v98, 0x400000, v23
; GFX11-NEXT: v_bfe_u32 v87, v7, 16, 1
; GFX11-NEXT: v_add3_u32 v85, v85, v24, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v20
; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-NEXT: v_add3_u32 v97, v97, v23, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v7
; GFX11-NEXT: v_add3_u32 v87, v87, v7, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_mul_f32_e32 v4, v4, v20
; GFX11-NEXT: v_mul_f32_e32 v20, v80, v71
; GFX11-NEXT: v_bfe_u32 v71, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX11-NEXT: v_dual_mul_f32 v21, v70, v69 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v71, v71, v9, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-NEXT: v_dual_mul_f32 v10, v10, v26 :: v_dual_and_b32 v29, 0xffff0000, v29
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-NEXT: v_mul_f32_e32 v26, v52, v51
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e32 v6, v6, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_dual_mul_f32 v11, v11, v27 :: v_dual_lshlrev_b32 v34, 16, v14
; GFX11-NEXT: v_dual_mul_f32 v22, v68, v67 :: v_dual_lshlrev_b32 v33, 16, v30
; GFX11-NEXT: v_dual_mul_f32 v27, v50, v49 :: v_dual_lshlrev_b32 v38, 16, v12
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_dual_mul_f32 v25, v54, v53 :: v_dual_and_b32 v12, 0xffff0000, v12
; GFX11-NEXT: v_dual_mul_f32 v13, v13, v29 :: v_dual_and_b32 v30, 0xffff0000, v30
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e32 v29, v38, v37
; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX11-NEXT: v_dual_mul_f32 v12, v12, v28 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e32 v14, v14, v30
; GFX11-NEXT: v_mul_f32_e32 v28, v48, v39
; GFX11-NEXT: v_dual_mul_f32 v30, v36, v35 :: v_dual_mul_f32 v33, v34, v33
; GFX11-NEXT: v_bfe_u32 v39, v13, 16, 1
; GFX11-NEXT: v_bfe_u32 v35, v14, 16, 1
; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v14
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v37, v30, 16, 1
; GFX11-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-NEXT: v_add3_u32 v35, v35, v14, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
; GFX11-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX11-NEXT: v_add3_u32 v37, v37, v30, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v13
; GFX11-NEXT: v_bfe_u32 v49, v29, 16, 1
; GFX11-NEXT: v_add3_u32 v39, v39, v13, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v34, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v29
; GFX11-NEXT: v_bfe_u32 v51, v12, 16, 1
; GFX11-NEXT: v_add3_u32 v49, v49, v29, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX11-NEXT: v_cndmask_b32_e32 v14, v35, v36, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-NEXT: v_bfe_u32 v53, v28, 16, 1
; GFX11-NEXT: v_add3_u32 v51, v51, v12, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v28
; GFX11-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v30, v37, v38, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_add3_u32 v53, v53, v28, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX11-NEXT: v_bfe_u32 v65, v27, 16, 1
; GFX11-NEXT: v_add3_u32 v55, v55, v11, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v13, v39, v48, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v27
; GFX11-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX11-NEXT: v_add3_u32 v65, v65, v27, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX11-NEXT: v_cndmask_b32_e32 v29, v49, v50, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_bfe_u32 v69, v26, 16, 1
; GFX11-NEXT: v_add3_u32 v67, v67, v10, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v26
; GFX11-NEXT: v_bfe_u32 v81, v25, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v12, v51, v52, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-NEXT: v_add3_u32 v69, v69, v26, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v25
; GFX11-NEXT: v_bfe_u32 v83, v8, 16, 1
; GFX11-NEXT: v_add3_u32 v81, v81, v25, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v83, v83, v8, 0x7fff
; GFX11-NEXT: v_bfe_u32 v99, v6, 16, 1
; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v6
; GFX11-NEXT: v_cndmask_b32_e32 v11, v55, v64, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-NEXT: v_bfe_u32 v101, v22, 16, 1
; GFX11-NEXT: v_add3_u32 v99, v99, v6, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v22
; GFX11-NEXT: v_bfe_u32 v113, v21, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v27, v65, v66, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-NEXT: v_add3_u32 v101, v101, v22, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v114, 0x400000, v21
; GFX11-NEXT: v_bfe_u32 v115, v4, 16, 1
; GFX11-NEXT: v_add3_u32 v113, v113, v21, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v10, v67, v68, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-NEXT: v_or_b32_e32 v116, 0x400000, v4
; GFX11-NEXT: v_bfe_u32 v117, v20, 16, 1
; GFX11-NEXT: v_add3_u32 v115, v115, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v118, 0x400000, v20
; GFX11-NEXT: v_cndmask_b32_e32 v26, v69, v70, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v117, v117, v20, 0x7fff
; GFX11-NEXT: v_bfe_u32 v133, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v134, 0x400000, v18
; GFX11-NEXT: v_bfe_u32 v147, v0, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v9, v71, v80, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-NEXT: v_add3_u32 v133, v133, v18, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v147, v147, v0, 0x7fff
; GFX11-NEXT: v_bfe_u32 v131, v2, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v25, v81, v82, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_or_b32_e32 v132, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v9, v9, v26, 0x7060302
; GFX11-NEXT: v_add3_u32 v131, v131, v2, 0x7fff
; GFX11-NEXT: v_perm_b32 v10, v10, v27, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v83, v84, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-NEXT: v_perm_b32 v11, v11, v28, 0x7060302
; GFX11-NEXT: v_perm_b32 v12, v12, v29, 0x7060302
; GFX11-NEXT: v_perm_b32 v13, v13, v30, 0x7060302
; GFX11-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v24, v85, v86, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v7, v87, v96, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v23, v97, v98, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_cndmask_b32_e32 v6, v99, v100, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v22, v101, v102, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_cndmask_b32_e32 v5, v103, v112, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v21, v113, v114, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_cndmask_b32_e32 v4, v115, v116, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v20, v117, v118, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_cndmask_b32_e32 v19, v129, v130, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_cndmask_b32_e32 v18, v133, v134, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v135, v144, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v17, v145, v146, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v147, v33, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v131, v132, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v119, v128, vcc_lo
; GFX11-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_mul_f32 v17, v31, v17 :: v_dual_and_b32 v18, 0xffff0000, v32
; GFX11-NEXT: v_mul_f32_e32 v15, v15, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX11-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v17, v18, v20, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fmul <32 x bfloat> %a, %b
ret <32 x bfloat> %op
}
define bfloat @v_fdiv_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fdiv_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v0
; GCN-NEXT: v_rcp_f32_e32 v3, v2
; GCN-NEXT: v_fma_f32 v4, -v2, v3, 1.0
; GCN-NEXT: v_fma_f32 v3, v4, v3, v3
; GCN-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0
; GCN-NEXT: v_mul_f32_e32 v5, v4, v3
; GCN-NEXT: v_fma_f32 v6, -v2, v5, v4
; GCN-NEXT: v_fma_f32 v5, v6, v3, v5
; GCN-NEXT: v_fma_f32 v2, -v2, v5, v4
; GCN-NEXT: v_div_fmas_f32 v2, v2, v3, v5
; GCN-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fdiv_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v0
; GFX7-NEXT: v_rcp_f32_e32 v3, v2
; GFX7-NEXT: v_fma_f32 v4, -v2, v3, 1.0
; GFX7-NEXT: v_fma_f32 v3, v4, v3, v3
; GFX7-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0
; GFX7-NEXT: v_mul_f32_e32 v5, v4, v3
; GFX7-NEXT: v_fma_f32 v6, -v2, v5, v4
; GFX7-NEXT: v_fma_f32 v5, v6, v3, v5
; GFX7-NEXT: v_fma_f32 v2, -v2, v5, v4
; GFX7-NEXT: v_div_fmas_f32 v2, v2, v3, v5
; GFX7-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fdiv_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v0
; GFX8-NEXT: v_div_scale_f32 v3, vcc, v0, v1, v0
; GFX8-NEXT: v_rcp_f32_e32 v4, v2
; GFX8-NEXT: v_fma_f32 v5, -v2, v4, 1.0
; GFX8-NEXT: v_fma_f32 v4, v5, v4, v4
; GFX8-NEXT: v_mul_f32_e32 v5, v3, v4
; GFX8-NEXT: v_fma_f32 v6, -v2, v5, v3
; GFX8-NEXT: v_fma_f32 v5, v6, v4, v5
; GFX8-NEXT: v_fma_f32 v2, -v2, v5, v3
; GFX8-NEXT: v_div_fmas_f32 v2, v2, v4, v5
; GFX8-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fdiv_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v0
; GFX9-NEXT: v_div_scale_f32 v3, vcc, v0, v1, v0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_rcp_f32_e32 v4, v2
; GFX9-NEXT: v_fma_f32 v5, -v2, v4, 1.0
; GFX9-NEXT: v_fma_f32 v4, v5, v4, v4
; GFX9-NEXT: v_mul_f32_e32 v5, v3, v4
; GFX9-NEXT: v_fma_f32 v6, -v2, v5, v3
; GFX9-NEXT: v_fma_f32 v5, v6, v4, v5
; GFX9-NEXT: v_fma_f32 v2, -v2, v5, v3
; GFX9-NEXT: v_div_fmas_f32 v2, v2, v4, v5
; GFX9-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fdiv_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_div_scale_f32 v2, s4, v1, v1, v0
; GFX10-NEXT: v_div_scale_f32 v5, vcc_lo, v0, v1, v0
; GFX10-NEXT: v_rcp_f32_e32 v3, v2
; GFX10-NEXT: v_fma_f32 v4, -v2, v3, 1.0
; GFX10-NEXT: v_fmac_f32_e32 v3, v4, v3
; GFX10-NEXT: v_mul_f32_e32 v4, v5, v3
; GFX10-NEXT: v_fma_f32 v6, -v2, v4, v5
; GFX10-NEXT: v_fmac_f32_e32 v4, v6, v3
; GFX10-NEXT: v_fma_f32 v2, -v2, v4, v5
; GFX10-NEXT: v_div_fmas_f32 v2, v2, v3, v4
; GFX10-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fdiv_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_div_scale_f32 v2, null, v1, v1, v0
; GFX11-NEXT: v_rcp_f32_e32 v3, v2
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_fma_f32 v4, -v2, v3, 1.0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fmac_f32_e32 v3, v4, v3
; GFX11-NEXT: v_div_scale_f32 v5, vcc_lo, v0, v1, v0
; GFX11-NEXT: v_mul_f32_e32 v4, v5, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fma_f32 v6, -v2, v4, v5
; GFX11-NEXT: v_fmac_f32_e32 v4, v6, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fma_f32 v2, -v2, v4, v5
; GFX11-NEXT: v_div_fmas_f32 v2, v2, v3, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_div_fixup_f32 v0, v2, v1, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fdiv bfloat %a, %b
ret bfloat %op
}
declare bfloat @llvm.fabs.bf16(bfloat)
define bfloat @v_fabs_bf16(bfloat %a) {
; GCN-LABEL: v_fabs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fabs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fabs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 0x7fff, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fabs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 0x7fff, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fabs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 0x7fff, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fabs_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b16 v0.l, 0x7fff, v0.l
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fabs_bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0x7fff, v0
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.fabs.bf16(bfloat %a)
ret bfloat %op
}
define amdgpu_ps i32 @s_fabs_bf16(bfloat inreg %a) {
; GCN-LABEL: s_fabs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_fabs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_fabs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_and_b32 s0, s0, 0x7fff
; GFX8-NEXT: s_and_b32 s0, 0xffff, s0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_fabs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_and_b32 s0, s0, 0x7fff
; GFX9-NEXT: s_and_b32 s0, 0xffff, s0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_fabs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_and_b32 s0, s0, 0x7fff
; GFX10-NEXT: s_and_b32 s0, 0xffff, s0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_fabs_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_and_b32 s0, s0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s0, 0xffff, s0
; GFX11-NEXT: ; return to shader part epilog
%op = call bfloat @llvm.fabs.bf16(bfloat %a)
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define bfloat @v_fneg_bf16(bfloat %a) {
; GCN-LABEL: v_fneg_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fneg_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fneg_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_xor_b32_e32 v0, 0x8000, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fneg_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_xor_b32_e32 v0, 0x8000, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fneg_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_xor_b32_e32 v0, 0x8000, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fneg_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_xor_b16 v0.l, 0x8000, v0.l
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fneg_bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_xor_b32_e32 v0, 0x8000, v0
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fneg bfloat %a
ret bfloat %op
}
declare i32 @llvm.amdgcn.readfirstlane(i32)
; FIXME: readfirstlane hack for other bugs
define amdgpu_ps i32 @s_fneg_bf16(bfloat inreg %a) {
; GCN-LABEL: s_fneg_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, -1.0, s0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_fneg_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, -1.0, s0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_fneg_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_xor_b32 s0, s0, 0x8000
; GFX8-NEXT: s_and_b32 s0, 0xffff, s0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_fneg_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_xor_b32 s0, s0, 0x8000
; GFX9-NEXT: s_and_b32 s0, 0xffff, s0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_fneg_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_xor_b32 s0, s0, 0x8000
; GFX10-NEXT: s_and_b32 s0, 0xffff, s0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_fneg_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_xor_b32 s0, s0, 0x8000
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s0, 0xffff, s0
; GFX11-NEXT: ; return to shader part epilog
%op = fneg bfloat %a
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define bfloat @v_fneg_fabs_bf16(bfloat %a) {
; GCN-LABEL: v_fneg_fabs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fneg_fabs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fneg_fabs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_or_b32_e32 v0, 0x8000, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fneg_fabs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_or_b32_e32 v0, 0x8000, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fneg_fabs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_or_b32_e32 v0, 0x8000, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fneg_fabs_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_or_b16 v0.l, 0x8000, v0.l
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fneg_fabs_bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_or_b32_e32 v0, 0x8000, v0
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%fabs = call bfloat @llvm.fabs.bf16(bfloat %a)
%op = fneg bfloat %fabs
ret bfloat %op
}
; FIXME: readfirstlane hack for other bugs
define amdgpu_ps i32 @s_fneg_fabs_bf16(bfloat inreg %a) {
; GCN-LABEL: s_fneg_fabs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: s_and_b32 s0, s0, 0xffff0000
; GCN-NEXT: s_bitset0_b32 s0, 31
; GCN-NEXT: s_and_b32 s0, s0, 0xffff0000
; GCN-NEXT: s_xor_b32 s0, s0, 0x80000000
; GCN-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_fneg_fabs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: s_and_b32 s0, s0, 0xffff0000
; GFX7-NEXT: s_bitset0_b32 s0, 31
; GFX7-NEXT: s_and_b32 s0, s0, 0xffff0000
; GFX7-NEXT: s_xor_b32 s0, s0, 0x80000000
; GFX7-NEXT: s_lshr_b32 s0, s0, 16
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_fneg_fabs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_bitset1_b32 s0, 15
; GFX8-NEXT: s_and_b32 s0, 0xffff, s0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_fneg_fabs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_bitset1_b32 s0, 15
; GFX9-NEXT: s_and_b32 s0, 0xffff, s0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_fneg_fabs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_bitset1_b32 s0, 15
; GFX10-NEXT: s_and_b32 s0, 0xffff, s0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_fneg_fabs_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_bitset1_b32 s0, 15
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s0, 0xffff, s0
; GFX11-NEXT: ; return to shader part epilog
%fabs = call bfloat @llvm.fabs.bf16(bfloat %a)
%op = fneg bfloat %fabs
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
declare bfloat @llvm.minnum.bf16(bfloat, bfloat)
declare <2 x bfloat> @llvm.minnum.v2bf16(<2 x bfloat>, <2 x bfloat>)
declare <3 x bfloat> @llvm.minnum.v3bf16(<3 x bfloat>, <3 x bfloat>)
declare <4 x bfloat> @llvm.minnum.v4bf16(<4 x bfloat>, <4 x bfloat>)
declare <8 x bfloat> @llvm.minnum.v8bf16(<8 x bfloat>, <8 x bfloat>)
declare <16 x bfloat> @llvm.minnum.v16bf16(<16 x bfloat>, <16 x bfloat>)
declare <32 x bfloat> @llvm.minnum.v32bf16(<32 x bfloat>, <32 x bfloat>)
define bfloat @v_minnum_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_minnum_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_min_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.minnum.bf16(bfloat %a, bfloat %b)
ret bfloat %op
}
define <2 x bfloat> @v_minnum_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_minnum_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v1, v1, v3
; GCN-NEXT: v_min_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v1, v1, v3
; GFX7-NEXT: v_min_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_min_f32_e32 v2, v3, v2
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_min_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_min_f32_e32 v2, v3, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_min_f32_e32 v2, v3, v2
; GFX10-NEXT: v_min_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_min_f32_e32 v0, v0, v1
; GFX11-NEXT: v_min_f32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <2 x bfloat> @llvm.minnum.v2bf16(<2 x bfloat> %a, <2 x bfloat> %b)
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_minnum_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_minnum_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v2, v2, v5
; GCN-NEXT: v_min_f32_e32 v1, v1, v4
; GCN-NEXT: v_min_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v2, v2, v5
; GFX7-NEXT: v_min_f32_e32 v1, v1, v4
; GFX7-NEXT: v_min_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_min_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_min_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_min_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_min_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_min_f32_e32 v3, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_min_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_min_f32_e32 v4, v5, v4
; GFX10-NEXT: v_min_f32_e32 v0, v0, v2
; GFX10-NEXT: v_min_f32_e32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_minnum_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_min_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11TRUE16-NEXT: v_dual_min_f32 v0, v0, v2 :: v_dual_min_f32 v1, v1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_minnum_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_min_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11FAKE16-NEXT: v_dual_min_f32 v0, v0, v2 :: v_dual_min_f32 v1, v1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = call <3 x bfloat> @llvm.minnum.v3bf16(<3 x bfloat> %a, <3 x bfloat> %b)
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_minnum_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_minnum_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v3, v3, v7
; GCN-NEXT: v_min_f32_e32 v2, v2, v6
; GCN-NEXT: v_min_f32_e32 v1, v1, v5
; GCN-NEXT: v_min_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v3, v3, v7
; GFX7-NEXT: v_min_f32_e32 v2, v2, v6
; GFX7-NEXT: v_min_f32_e32 v1, v1, v5
; GFX7-NEXT: v_min_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_min_f32_e32 v4, v5, v4
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_min_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_min_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_min_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v4, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_min_f32_e32 v4, v5, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_min_f32_e32 v3, v5, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_min_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_min_f32_e32 v4, v5, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_min_f32_e32 v1, v1, v3
; GFX10-NEXT: v_min_f32_e32 v3, v7, v6
; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_min_f32_e32 v0, v0, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_min_f32 v0, v0, v2 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_f32_e32 v1, v1, v3
; GFX11-NEXT: v_dual_min_f32 v3, v7, v6 :: v_dual_min_f32 v4, v5, v4
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <4 x bfloat> @llvm.minnum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
ret <4 x bfloat> %op
}
define <8 x bfloat> @v_minnum_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_minnum_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v7, v7, v15
; GCN-NEXT: v_min_f32_e32 v6, v6, v14
; GCN-NEXT: v_min_f32_e32 v5, v5, v13
; GCN-NEXT: v_min_f32_e32 v4, v4, v12
; GCN-NEXT: v_min_f32_e32 v3, v3, v11
; GCN-NEXT: v_min_f32_e32 v2, v2, v10
; GCN-NEXT: v_min_f32_e32 v1, v1, v9
; GCN-NEXT: v_min_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v7, v7, v15
; GFX7-NEXT: v_min_f32_e32 v6, v6, v14
; GFX7-NEXT: v_min_f32_e32 v5, v5, v13
; GFX7-NEXT: v_min_f32_e32 v4, v4, v12
; GFX7-NEXT: v_min_f32_e32 v3, v3, v11
; GFX7-NEXT: v_min_f32_e32 v2, v2, v10
; GFX7-NEXT: v_min_f32_e32 v1, v1, v9
; GFX7-NEXT: v_min_f32_e32 v0, v0, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX8-NEXT: v_min_f32_e32 v8, v9, v8
; GFX8-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v8
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
; GFX8-NEXT: v_min_f32_e32 v3, v3, v7
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX8-NEXT: v_min_f32_e32 v7, v9, v7
; GFX8-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v7
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_min_f32_e32 v2, v2, v6
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX8-NEXT: v_min_f32_e32 v6, v9, v6
; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_min_f32_e32 v1, v1, v5
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX8-NEXT: v_min_f32_e32 v5, v9, v5
; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_min_f32_e32 v0, v0, v4
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v5, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v7, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v8, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX9-NEXT: v_min_f32_e32 v8, v9, v8
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_f32_e32 v3, v3, v7
; GFX9-NEXT: v_add3_u32 v9, v9, v8, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v7, v7, v3, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX9-NEXT: v_min_f32_e32 v7, v9, v7
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX9-NEXT: v_min_f32_e32 v2, v2, v6
; GFX9-NEXT: v_add3_u32 v9, v9, v7, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v6, v6, v2, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX9-NEXT: v_min_f32_e32 v6, v9, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX9-NEXT: v_min_f32_e32 v1, v1, v5
; GFX9-NEXT: v_add3_u32 v9, v9, v6, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v5, v5, v1, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX9-NEXT: v_min_f32_e32 v5, v9, v5
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX9-NEXT: v_min_f32_e32 v0, v0, v4
; GFX9-NEXT: v_add3_u32 v9, v9, v5, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v5, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v7, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v8, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_min_f32_e32 v8, v9, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_min_f32_e32 v3, v3, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX10-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX10-NEXT: v_min_f32_e32 v7, v10, v9
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT: v_min_f32_e32 v2, v2, v6
; GFX10-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX10-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX10-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX10-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_min_f32_e32 v6, v10, v6
; GFX10-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX10-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_min_f32_e32 v1, v1, v5
; GFX10-NEXT: v_min_f32_e32 v5, v15, v13
; GFX10-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX10-NEXT: v_min_f32_e32 v0, v0, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX10-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v10, vcc_lo
; GFX10-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX10-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_min_f32 v8, v9, v8 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_f32_e32 v3, v3, v7
; GFX11-NEXT: v_min_f32_e32 v7, v10, v9
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_cndmask_b32 v7, v11, v12 :: v_dual_min_f32 v2, v2, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_f32_e32 v6, v10, v6
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_f32_e32 v0, v0, v4
; GFX11-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX11-NEXT: v_dual_min_f32 v1, v1, v5 :: v_dual_cndmask_b32 v4, v4, v10
; GFX11-NEXT: v_min_f32_e32 v5, v15, v13
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX11-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX11-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <8 x bfloat> @llvm.minnum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_minnum_v16bf16(<16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_minnum_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_min_f32_e32 v14, v14, v30
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_min_f32_e32 v13, v13, v29
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_min_f32_e32 v12, v12, v28
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_min_f32_e32 v11, v11, v27
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_min_f32_e32 v10, v10, v26
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_min_f32_e32 v9, v9, v25
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_min_f32_e32 v8, v8, v24
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_min_f32_e32 v7, v7, v23
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_min_f32_e32 v6, v6, v22
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_min_f32_e32 v5, v5, v21
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_min_f32_e32 v4, v4, v20
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_min_f32_e32 v3, v3, v19
; GCN-NEXT: v_min_f32_e32 v2, v2, v18
; GCN-NEXT: v_min_f32_e32 v1, v1, v17
; GCN-NEXT: v_min_f32_e32 v0, v0, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_min_f32_e32 v15, v15, v16
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_min_f32_e32 v6, v6, v22
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_min_f32_e32 v14, v14, v30
; GFX7-NEXT: v_min_f32_e32 v13, v13, v29
; GFX7-NEXT: v_min_f32_e32 v12, v12, v28
; GFX7-NEXT: v_min_f32_e32 v11, v11, v27
; GFX7-NEXT: v_min_f32_e32 v10, v10, v26
; GFX7-NEXT: v_min_f32_e32 v9, v9, v25
; GFX7-NEXT: v_min_f32_e32 v8, v8, v24
; GFX7-NEXT: v_min_f32_e32 v7, v7, v23
; GFX7-NEXT: v_min_f32_e32 v5, v5, v21
; GFX7-NEXT: v_min_f32_e32 v4, v4, v20
; GFX7-NEXT: v_min_f32_e32 v3, v3, v19
; GFX7-NEXT: v_min_f32_e32 v2, v2, v18
; GFX7-NEXT: v_min_f32_e32 v1, v1, v17
; GFX7-NEXT: v_min_f32_e32 v0, v0, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_min_f32_e32 v15, v15, v22
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX8-NEXT: v_min_f32_e32 v16, v17, v16
; GFX8-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v16
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v7, v7, v15
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX8-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v7
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX8-NEXT: v_min_f32_e32 v15, v17, v15
; GFX8-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v15
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v6, v6, v14
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v14, vcc, v14, v6
; GFX8-NEXT: v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX8-NEXT: v_min_f32_e32 v14, v17, v14
; GFX8-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v5, v5, v13
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v13, vcc, v13, v5
; GFX8-NEXT: v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX8-NEXT: v_min_f32_e32 v13, v17, v13
; GFX8-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v13
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v4, v4, v12
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v12, vcc, v12, v4
; GFX8-NEXT: v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX8-NEXT: v_min_f32_e32 v12, v17, v12
; GFX8-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v12
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v3, v3, v11
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v11, vcc, v11, v3
; GFX8-NEXT: v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX8-NEXT: v_min_f32_e32 v11, v17, v11
; GFX8-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v11
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v2, v2, v10
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v2
; GFX8-NEXT: v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX8-NEXT: v_min_f32_e32 v10, v17, v10
; GFX8-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v10
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v1, v1, v9
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX8-NEXT: v_min_f32_e32 v9, v17, v9
; GFX8-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v9
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_min_f32_e32 v0, v0, v8
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v0
; GFX8-NEXT: v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v9, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v10, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v11, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v12, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v13, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v14, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v15, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v16, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX9-NEXT: v_min_f32_e32 v16, v17, v16
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_f32_e32 v7, v7, v15
; GFX9-NEXT: v_add3_u32 v17, v17, v16, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v15, v15, v7, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX9-NEXT: v_min_f32_e32 v15, v17, v15
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX9-NEXT: v_min_f32_e32 v6, v6, v14
; GFX9-NEXT: v_add3_u32 v17, v17, v15, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v14, v14, v6, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX9-NEXT: v_min_f32_e32 v14, v17, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX9-NEXT: v_min_f32_e32 v5, v5, v13
; GFX9-NEXT: v_add3_u32 v17, v17, v14, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v13, v13, v5, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX9-NEXT: v_min_f32_e32 v13, v17, v13
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX9-NEXT: v_min_f32_e32 v4, v4, v12
; GFX9-NEXT: v_add3_u32 v17, v17, v13, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v12, v12, v4, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX9-NEXT: v_min_f32_e32 v12, v17, v12
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX9-NEXT: v_min_f32_e32 v3, v3, v11
; GFX9-NEXT: v_add3_u32 v17, v17, v12, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v11, v11, v3, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX9-NEXT: v_min_f32_e32 v11, v17, v11
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX9-NEXT: v_min_f32_e32 v2, v2, v10
; GFX9-NEXT: v_add3_u32 v17, v17, v11, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v10, v10, v2, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX9-NEXT: v_min_f32_e32 v10, v17, v10
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX9-NEXT: v_min_f32_e32 v1, v1, v9
; GFX9-NEXT: v_add3_u32 v17, v17, v10, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v9, v9, v1, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX9-NEXT: v_min_f32_e32 v9, v17, v9
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX9-NEXT: v_min_f32_e32 v0, v0, v8
; GFX9-NEXT: v_add3_u32 v17, v17, v9, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v8, v8, v0, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v9, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v10, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v11, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v12, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v13, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v14, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v15, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v16, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_min_f32_e32 v16, v17, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX10-NEXT: v_min_f32_e32 v7, v7, v15
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX10-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT: v_min_f32_e32 v17, v18, v17
; GFX10-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX10-NEXT: v_min_f32_e32 v6, v6, v14
; GFX10-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX10-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v15, v15, v20, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v5
; GFX10-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX10-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX10-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX10-NEXT: v_min_f32_e32 v17, v20, v19
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX10-NEXT: v_min_f32_e32 v5, v5, v13
; GFX10-NEXT: v_cndmask_b32_e32 v14, v14, v16, vcc_lo
; GFX10-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX10-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v6, v16, v13, vcc_lo
; GFX10-NEXT: v_min_f32_e32 v13, v19, v18
; GFX10-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX10-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX10-NEXT: v_min_f32_e32 v4, v4, v12
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX10-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_min_f32_e32 v12, v18, v12
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX10-NEXT: v_min_f32_e32 v3, v3, v11
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX10-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX10-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX10-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX10-NEXT: v_min_f32_e32 v18, v19, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX10-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX10-NEXT: v_min_f32_e32 v2, v2, v10
; GFX10-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX10-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX10-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX10-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX10-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX10-NEXT: v_min_f32_e32 v19, v22, v20
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX10-NEXT: v_min_f32_e32 v1, v1, v9
; GFX10-NEXT: v_min_f32_e32 v9, v22, v20
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX10-NEXT: v_min_f32_e32 v0, v0, v8
; GFX10-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX10-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX10-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX10-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX10-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_min_f32 v16, v17, v16 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_min_f32_e32 v17, v18, v17
; GFX11-NEXT: v_min_f32_e32 v6, v6, v14
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX11-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_f32_e32 v7, v7, v15
; GFX11-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX11-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_cndmask_b32 v15, v15, v20 :: v_dual_lshlrev_b32 v20, 16, v5
; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_dual_min_f32 v17, v20, v19 :: v_dual_cndmask_b32 v14, v14, v16
; GFX11-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_f32_e32 v4, v4, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_min_f32_e32 v5, v5, v13
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_cndmask_b32 v6, v16, v13 :: v_dual_min_f32 v13, v19, v18
; GFX11-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_min_f32_e32 v12, v18, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX11-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX11-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_f32_e32 v18, v19, v18
; GFX11-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_min_f32 v2, v2, v10 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_min_f32_e32 v3, v3, v11
; GFX11-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX11-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX11-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX11-NEXT: v_min_f32_e32 v19, v22, v20
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX11-NEXT: v_dual_min_f32 v0, v0, v8 :: v_dual_min_f32 v1, v1, v9
; GFX11-NEXT: v_min_f32_e32 v9, v22, v20
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX11-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <16 x bfloat> @llvm.minnum.v16bf16(<16 x bfloat> %a, <16 x bfloat> %b)
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_minnum_v32bf16(<32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_minnum_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:124
; GCN-NEXT: v_min_f32_e32 v31, v31, v32
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:120
; GCN-NEXT: v_min_f32_e32 v30, v30, v32
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
; GCN-NEXT: v_min_f32_e32 v29, v29, v32
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112
; GCN-NEXT: v_min_f32_e32 v28, v28, v32
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:108
; GCN-NEXT: v_min_f32_e32 v27, v27, v32
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
; GCN-NEXT: v_min_f32_e32 v26, v26, v32
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100
; GCN-NEXT: v_min_f32_e32 v25, v25, v32
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
; GCN-NEXT: v_min_f32_e32 v24, v24, v32
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:92
; GCN-NEXT: v_min_f32_e32 v23, v23, v32
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
; GCN-NEXT: v_min_f32_e32 v22, v22, v32
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:84
; GCN-NEXT: v_min_f32_e32 v21, v21, v32
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:80
; GCN-NEXT: v_min_f32_e32 v20, v20, v32
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:76
; GCN-NEXT: v_min_f32_e32 v19, v19, v32
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72
; GCN-NEXT: v_min_f32_e32 v18, v18, v32
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:68
; GCN-NEXT: v_min_f32_e32 v17, v17, v32
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:64
; GCN-NEXT: v_min_f32_e32 v16, v16, v32
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:60
; GCN-NEXT: v_min_f32_e32 v15, v15, v32
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:56
; GCN-NEXT: v_min_f32_e32 v14, v14, v32
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:52
; GCN-NEXT: v_min_f32_e32 v13, v13, v32
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:48
; GCN-NEXT: v_min_f32_e32 v12, v12, v32
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:44
; GCN-NEXT: v_min_f32_e32 v11, v11, v32
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
; GCN-NEXT: v_min_f32_e32 v10, v10, v32
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:36
; GCN-NEXT: v_min_f32_e32 v9, v9, v32
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
; GCN-NEXT: v_min_f32_e32 v8, v8, v32
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28
; GCN-NEXT: v_min_f32_e32 v7, v7, v32
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:24
; GCN-NEXT: v_min_f32_e32 v6, v6, v32
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
; GCN-NEXT: v_min_f32_e32 v5, v5, v32
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:16
; GCN-NEXT: v_min_f32_e32 v4, v4, v32
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:12
; GCN-NEXT: v_min_f32_e32 v3, v3, v32
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GCN-NEXT: v_min_f32_e32 v2, v2, v32
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
; GCN-NEXT: v_min_f32_e32 v1, v1, v32
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_min_f32_e32 v0, v0, v32
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_minnum_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: v_min_f32_e32 v31, v31, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v30, v30, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v29, v29, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v28, v28, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v27, v27, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v26, v26, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v25, v25, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v24, v24, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v23, v23, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v22, v22, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v21, v21, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v20, v20, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v19, v19, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v18, v18, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v17, v17, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v16, v16, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v15, v15, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v14, v14, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v13, v13, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v12, v12, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v11, v11, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v10, v10, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v9, v9, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v8, v8, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v7, v7, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v6, v6, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v5, v5, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v4, v4, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v3, v3, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v2, v2, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v1, v1, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_min_f32_e32 v0, v0, v32
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_minnum_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX8-NEXT: v_min_f32_e32 v31, v32, v31
; GFX8-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v32, vcc, v32, v31
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_add_u32_e32 v32, vcc, s4, v32
; GFX8-NEXT: v_min_f32_e32 v14, v14, v30
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX8-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v30, vcc, v30, v14
; GFX8-NEXT: v_add_u32_e32 v30, vcc, s4, v30
; GFX8-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX8-NEXT: v_min_f32_e32 v32, v32, v30
; GFX8-NEXT: buffer_load_dword v30, off, s[0:3], s32
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_min_f32_e32 v13, v13, v29
; GFX8-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX8-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX8-NEXT: v_alignbit_b32 v14, v14, v31, 16
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; GFX8-NEXT: v_min_f32_e32 v33, v33, v34
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_min_f32_e32 v30, v15, v30
; GFX8-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v33
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX8-NEXT: v_bfe_u32 v33, v30, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v30
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v30
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX8-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v32
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX8-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v29, vcc, v29, v13
; GFX8-NEXT: v_add_u32_e32 v29, vcc, s4, v29
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_cndmask_b32_e32 v13, v29, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v12
; GFX8-NEXT: v_min_f32_e32 v29, v33, v29
; GFX8-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v29
; GFX8-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v12, v12, v28
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v28, vcc, v28, v12
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX8-NEXT: v_min_f32_e32 v28, v33, v28
; GFX8-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v28
; GFX8-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v11, v11, v27
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v27, vcc, v27, v11
; GFX8-NEXT: v_add_u32_e32 v27, vcc, s4, v27
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX8-NEXT: v_min_f32_e32 v27, v33, v27
; GFX8-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v27
; GFX8-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v10, v10, v26
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v26, vcc, v26, v10
; GFX8-NEXT: v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX8-NEXT: v_min_f32_e32 v26, v33, v26
; GFX8-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v26
; GFX8-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v9, v9, v25
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v25, vcc, v25, v9
; GFX8-NEXT: v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX8-NEXT: v_min_f32_e32 v25, v33, v25
; GFX8-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v25
; GFX8-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v8, v8, v24
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v24, vcc, v24, v8
; GFX8-NEXT: v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX8-NEXT: v_min_f32_e32 v24, v33, v24
; GFX8-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v24
; GFX8-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v7, v7, v23
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v23, vcc, v23, v7
; GFX8-NEXT: v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX8-NEXT: v_min_f32_e32 v23, v33, v23
; GFX8-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v23
; GFX8-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v6, v6, v22
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v22, vcc, v22, v6
; GFX8-NEXT: v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX8-NEXT: v_min_f32_e32 v22, v33, v22
; GFX8-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v22
; GFX8-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v5, v5, v21
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v21, vcc, v21, v5
; GFX8-NEXT: v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX8-NEXT: v_min_f32_e32 v21, v33, v21
; GFX8-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v21
; GFX8-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v4, v4, v20
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v20, vcc, v20, v4
; GFX8-NEXT: v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX8-NEXT: v_min_f32_e32 v20, v33, v20
; GFX8-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v20
; GFX8-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v3, v3, v19
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v19, vcc, v19, v3
; GFX8-NEXT: v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX8-NEXT: v_min_f32_e32 v19, v33, v19
; GFX8-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v19
; GFX8-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v2, v2, v18
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v18, vcc, v18, v2
; GFX8-NEXT: v_add_u32_e32 v18, vcc, s4, v18
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX8-NEXT: v_min_f32_e32 v18, v33, v18
; GFX8-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v18
; GFX8-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v1, v1, v17
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX8-NEXT: v_min_f32_e32 v17, v33, v17
; GFX8-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v17
; GFX8-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_min_f32_e32 v0, v0, v16
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v16, vcc, v16, v0
; GFX8-NEXT: v_add_u32_e32 v16, vcc, s4, v16
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v30
; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX8-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX8-NEXT: v_alignbit_b32 v0, v0, v17, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v18, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v19, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v20, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v21, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v22, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v23, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v24, 16
; GFX8-NEXT: v_alignbit_b32 v8, v8, v25, 16
; GFX8-NEXT: v_alignbit_b32 v9, v9, v26, 16
; GFX8-NEXT: v_alignbit_b32 v10, v10, v27, 16
; GFX8-NEXT: v_alignbit_b32 v11, v11, v28, 16
; GFX8-NEXT: v_alignbit_b32 v12, v12, v29, 16
; GFX8-NEXT: v_alignbit_b32 v13, v13, v32, 16
; GFX8-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_minnum_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT: v_min_f32_e32 v31, v32, v31
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX9-NEXT: v_min_f32_e32 v14, v14, v30
; GFX9-NEXT: v_add3_u32 v32, v32, v31, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v30, v30, v14, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT: v_min_f32_e32 v30, v32, v30
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_bfe_u32 v32, v30, 16, 1
; GFX9-NEXT: v_min_f32_e32 v13, v13, v29
; GFX9-NEXT: v_add3_u32 v32, v32, v30, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v30
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v29, v29, v13, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_cndmask_b32_e32 v13, v29, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT: v_min_f32_e32 v32, v32, v29
; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_min_f32_e32 v12, v12, v28
; GFX9-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX9-NEXT: v_add3_u32 v28, v28, v12, s4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v29
; GFX9-NEXT: v_min_f32_e32 v33, v33, v34
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_min_f32_e32 v29, v15, v29
; GFX9-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX9-NEXT: v_add3_u32 v15, v15, v33, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX9-NEXT: v_add3_u32 v33, v33, v29, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT: v_add3_u32 v33, v33, v32, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX9-NEXT: v_min_f32_e32 v28, v33, v28
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT: v_min_f32_e32 v11, v11, v27
; GFX9-NEXT: v_add3_u32 v33, v33, v28, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v27, v27, v11, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX9-NEXT: v_min_f32_e32 v27, v33, v27
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT: v_min_f32_e32 v10, v10, v26
; GFX9-NEXT: v_add3_u32 v33, v33, v27, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v26, v26, v10, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX9-NEXT: v_min_f32_e32 v26, v33, v26
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT: v_min_f32_e32 v9, v9, v25
; GFX9-NEXT: v_add3_u32 v33, v33, v26, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v25, v25, v9, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX9-NEXT: v_min_f32_e32 v25, v33, v25
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT: v_min_f32_e32 v8, v8, v24
; GFX9-NEXT: v_add3_u32 v33, v33, v25, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v24, v24, v8, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX9-NEXT: v_min_f32_e32 v24, v33, v24
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT: v_min_f32_e32 v7, v7, v23
; GFX9-NEXT: v_add3_u32 v33, v33, v24, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v23, v23, v7, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX9-NEXT: v_min_f32_e32 v23, v33, v23
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT: v_min_f32_e32 v6, v6, v22
; GFX9-NEXT: v_add3_u32 v33, v33, v23, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v22, v22, v6, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX9-NEXT: v_min_f32_e32 v22, v33, v22
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT: v_min_f32_e32 v5, v5, v21
; GFX9-NEXT: v_add3_u32 v33, v33, v22, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v21, v21, v5, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX9-NEXT: v_min_f32_e32 v21, v33, v21
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT: v_min_f32_e32 v4, v4, v20
; GFX9-NEXT: v_add3_u32 v33, v33, v21, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v20, v20, v4, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX9-NEXT: v_min_f32_e32 v20, v33, v20
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT: v_min_f32_e32 v3, v3, v19
; GFX9-NEXT: v_add3_u32 v33, v33, v20, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v19, v19, v3, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX9-NEXT: v_min_f32_e32 v19, v33, v19
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT: v_min_f32_e32 v2, v2, v18
; GFX9-NEXT: v_add3_u32 v33, v33, v19, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v18, v18, v2, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX9-NEXT: v_min_f32_e32 v18, v33, v18
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT: v_min_f32_e32 v1, v1, v17
; GFX9-NEXT: v_add3_u32 v33, v33, v18, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v17, v17, v1, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX9-NEXT: v_min_f32_e32 v17, v33, v17
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT: v_min_f32_e32 v0, v0, v16
; GFX9-NEXT: v_add3_u32 v33, v33, v17, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v16, v16, v0, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v17, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v18, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v19, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v20, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v21, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v22, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v23, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v24, s4
; GFX9-NEXT: v_perm_b32 v8, v8, v25, s4
; GFX9-NEXT: v_perm_b32 v9, v9, v26, s4
; GFX9-NEXT: v_perm_b32 v10, v10, v27, s4
; GFX9-NEXT: v_perm_b32 v11, v11, v28, s4
; GFX9-NEXT: v_perm_b32 v12, v12, v32, s4
; GFX9-NEXT: v_perm_b32 v13, v13, v30, s4
; GFX9-NEXT: v_perm_b32 v14, v14, v31, s4
; GFX9-NEXT: v_perm_b32 v15, v29, v15, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_minnum_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX10-NEXT: v_lshlrev_b32_e32 v39, 16, v27
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX10-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX10-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_lshlrev_b32_e32 v37, 16, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v12
; GFX10-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX10-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX10-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX10-NEXT: v_lshlrev_b32_e32 v54, 16, v8
; GFX10-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX10-NEXT: v_lshlrev_b32_e32 v64, 16, v7
; GFX10-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX10-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX10-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX10-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX10-NEXT: v_min_f32_e32 v39, v48, v39
; GFX10-NEXT: v_min_f32_e32 v11, v11, v27
; GFX10-NEXT: v_min_f32_e32 v49, v50, v49
; GFX10-NEXT: v_min_f32_e32 v10, v10, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX10-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_min_f32_e32 v37, v38, v37
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v18
; GFX10-NEXT: v_min_f32_e32 v12, v12, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v2
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v17
; GFX10-NEXT: v_lshlrev_b32_e32 v27, 16, v1
; GFX10-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v26, 16, v0
; GFX10-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_min_f32_e32 v9, v9, v25
; GFX10-NEXT: v_min_f32_e32 v25, v54, v53
; GFX10-NEXT: v_min_f32_e32 v8, v8, v24
; GFX10-NEXT: v_min_f32_e32 v24, v64, v55
; GFX10-NEXT: v_min_f32_e32 v7, v7, v23
; GFX10-NEXT: v_min_f32_e32 v23, v66, v65
; GFX10-NEXT: v_min_f32_e32 v6, v6, v22
; GFX10-NEXT: v_min_f32_e32 v22, v68, v67
; GFX10-NEXT: v_bfe_u32 v53, v39, 16, 1
; GFX10-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX10-NEXT: v_bfe_u32 v65, v49, 16, 1
; GFX10-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v33, 16, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v14
; GFX10-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_min_f32_e32 v35, v36, v35
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v19
; GFX10-NEXT: v_min_f32_e32 v13, v13, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_min_f32_e32 v2, v2, v18
; GFX10-NEXT: v_min_f32_e32 v18, v27, v48
; GFX10-NEXT: v_min_f32_e32 v1, v1, v17
; GFX10-NEXT: v_min_f32_e32 v17, v26, v50
; GFX10-NEXT: v_min_f32_e32 v0, v0, v16
; GFX10-NEXT: v_or_b32_e32 v54, 0x400000, v39
; GFX10-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX10-NEXT: v_or_b32_e32 v66, 0x400000, v49
; GFX10-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v39, v39
; GFX10-NEXT: v_add3_u32 v39, v53, v39, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v11, v11
; GFX10-NEXT: v_add3_u32 v11, v55, v11, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v49, v49
; GFX10-NEXT: v_add3_u32 v49, v65, v49, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v10, v10
; GFX10-NEXT: v_add3_u32 v10, v67, v10, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_min_f32_e32 v33, v34, v33
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v20
; GFX10-NEXT: v_min_f32_e32 v14, v14, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v30, 16, v4
; GFX10-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_min_f32_e32 v3, v3, v19
; GFX10-NEXT: v_min_f32_e32 v19, v28, v38
; GFX10-NEXT: v_bfe_u32 v38, v37, 16, 1
; GFX10-NEXT: v_bfe_u32 v50, v12, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v39, v39, v54, s9
; GFX10-NEXT: v_bfe_u32 v54, v18, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v11, v11, v64, s10
; GFX10-NEXT: v_bfe_u32 v64, v1, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v49, v49, v66, s11
; GFX10-NEXT: v_bfe_u32 v66, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v10, v10, v68, s12
; GFX10-NEXT: v_bfe_u32 v68, v0, 16, 1
; GFX10-NEXT: v_min_f32_e32 v51, v52, v51
; GFX10-NEXT: v_min_f32_e32 v5, v5, v21
; GFX10-NEXT: v_min_f32_e32 v21, v30, v34
; GFX10-NEXT: v_min_f32_e32 v4, v4, v20
; GFX10-NEXT: v_min_f32_e32 v20, v29, v36
; GFX10-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX10-NEXT: v_bfe_u32 v27, v14, 16, 1
; GFX10-NEXT: v_bfe_u32 v29, v35, 16, 1
; GFX10-NEXT: v_bfe_u32 v34, v13, 16, 1
; GFX10-NEXT: v_or_b32_e32 v48, 0x400000, v37
; GFX10-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v37, v37
; GFX10-NEXT: v_add3_u32 v37, v38, v37, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v12, v12
; GFX10-NEXT: v_add3_u32 v12, v50, v12, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v18, v18
; GFX10-NEXT: v_add3_u32 v54, v54, v18, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v18
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v1, v1
; GFX10-NEXT: v_add3_u32 v64, v64, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v17, v17
; GFX10-NEXT: v_add3_u32 v66, v66, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s22, v0, v0
; GFX10-NEXT: v_add3_u32 v68, v68, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v0, 0x400000, v0
; GFX10-NEXT: v_or_b32_e32 v26, 0x400000, v33
; GFX10-NEXT: v_or_b32_e32 v28, 0x400000, v14
; GFX10-NEXT: v_or_b32_e32 v30, 0x400000, v35
; GFX10-NEXT: v_or_b32_e32 v36, 0x400000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX10-NEXT: v_bfe_u32 v33, v51, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v14, v14
; GFX10-NEXT: v_add3_u32 v14, v27, v14, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v35, v35
; GFX10-NEXT: v_add3_u32 v29, v29, v35, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v13, v13
; GFX10-NEXT: v_add3_u32 v13, v34, v13, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v24, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v37, v37, v48, s7
; GFX10-NEXT: v_bfe_u32 v48, v19, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v12, v12, v52, s8
; GFX10-NEXT: v_bfe_u32 v52, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v18, v54, v18, s10
; GFX10-NEXT: v_cndmask_b32_e64 v17, v66, v17, s12
; GFX10-NEXT: v_cndmask_b32_e64 v0, v68, v0, s22
; GFX10-NEXT: v_cndmask_b32_e64 v1, v64, v1, s11
; GFX10-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_or_b32_e32 v27, 0x400000, v51
; GFX10-NEXT: v_bfe_u32 v35, v9, 16, 1
; GFX10-NEXT: v_bfe_u32 v38, v25, 16, 1
; GFX10-NEXT: v_or_b32_e32 v67, 0x400000, v24
; GFX10-NEXT: v_cmp_u_f32_e64 s13, v51, v51
; GFX10-NEXT: v_add3_u32 v33, v33, v51, 0x7fff
; GFX10-NEXT: v_bfe_u32 v51, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s17, v24, v24
; GFX10-NEXT: v_add3_u32 v24, v65, v24, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v6, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v26, vcc_lo
; GFX10-NEXT: v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v14, v14, v28, s4
; GFX10-NEXT: v_bfe_u32 v28, v4, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v29, v29, v30, s5
; GFX10-NEXT: v_bfe_u32 v30, v20, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v13, v13, v36, s6
; GFX10-NEXT: v_bfe_u32 v36, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v19, v19
; GFX10-NEXT: v_add3_u32 v48, v48, v19, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v19
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v2, v2
; GFX10-NEXT: v_add3_u32 v52, v52, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX10-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX10-NEXT: v_or_b32_e32 v34, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v50, 0x400000, v25
; GFX10-NEXT: v_bfe_u32 v53, v8, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s14, v9, v9
; GFX10-NEXT: v_add3_u32 v9, v35, v9, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v35, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e64 s15, v25, v25
; GFX10-NEXT: v_add3_u32 v25, v38, v25, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v23, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s18, v7, v7
; GFX10-NEXT: v_add3_u32 v7, v51, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v51, 0x400000, v6
; GFX10-NEXT: v_cmp_u_f32_e64 s20, v6, v6
; GFX10-NEXT: v_add3_u32 v6, v65, v6, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v21, v21
; GFX10-NEXT: v_add3_u32 v26, v26, v21, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v21
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v4, v4
; GFX10-NEXT: v_add3_u32 v28, v28, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v4
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v20, v20
; GFX10-NEXT: v_add3_u32 v30, v30, v20, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v20
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v3, v3
; GFX10-NEXT: v_add3_u32 v36, v36, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e64 v19, v48, v19, s8
; GFX10-NEXT: v_cndmask_b32_e64 v2, v52, v2, s9
; GFX10-NEXT: v_or_b32_e32 v55, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e64 s16, v8, v8
; GFX10-NEXT: v_add3_u32 v8, v53, v8, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v53, 0x400000, v23
; GFX10-NEXT: v_cmp_u_f32_e64 s19, v23, v23
; GFX10-NEXT: v_add3_u32 v23, v38, v23, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v22, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v65, v65, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e64 v21, v26, v21, s4
; GFX10-NEXT: v_cndmask_b32_e64 v4, v28, v4, s5
; GFX10-NEXT: v_cndmask_b32_e64 v20, v30, v20, s6
; GFX10-NEXT: v_cndmask_b32_e64 v3, v36, v3, s7
; GFX10-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX10-NEXT: v_cmp_u_f32_e64 s21, v22, v22
; GFX10-NEXT: v_add3_u32 v38, v38, v22, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v22
; GFX10-NEXT: v_cndmask_b32_e32 v5, v65, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX10-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e64 v27, v33, v27, s13
; GFX10-NEXT: v_cndmask_b32_e64 v9, v9, v34, s14
; GFX10-NEXT: v_cndmask_b32_e64 v25, v25, v50, s15
; GFX10-NEXT: v_cndmask_b32_e64 v8, v8, v55, s16
; GFX10-NEXT: v_cndmask_b32_e64 v24, v24, v67, s17
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v35, s18
; GFX10-NEXT: v_cndmask_b32_e64 v23, v23, v53, s19
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v51, s20
; GFX10-NEXT: v_cndmask_b32_e64 v22, v38, v22, s21
; GFX10-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX10-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX10-NEXT: v_perm_b32 v9, v9, v27, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX10-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX10-NEXT: v_perm_b32 v10, v10, v49, 0x7060302
; GFX10-NEXT: v_perm_b32 v11, v11, v39, 0x7060302
; GFX10-NEXT: v_perm_b32 v12, v12, v37, 0x7060302
; GFX10-NEXT: v_perm_b32 v13, v13, v29, 0x7060302
; GFX10-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v32
; GFX10-NEXT: v_min_f32_e32 v17, v31, v17
; GFX10-NEXT: v_min_f32_e32 v15, v15, v18
; GFX10-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX10-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v15, v15
; GFX10-NEXT: v_add3_u32 v17, v18, v17, 0x7fff
; GFX10-NEXT: v_add3_u32 v15, v19, v15, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v15, v15, v21, s4
; GFX10-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_minnum_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_load_b32 v32, off, s32
; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX11-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v1
; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX11-NEXT: v_dual_min_f32 v5, v5, v21 :: v_dual_and_b32 v26, 0xffff0000, v26
; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_min_f32 v1, v1, v17 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v19
; GFX11-NEXT: v_bfe_u32 v103, v5, 16, 1
; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v18
; GFX11-NEXT: v_bfe_u32 v135, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v5
; GFX11-NEXT: v_or_b32_e32 v144, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v103, v103, v5, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_add3_u32 v135, v135, v1, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_min_f32 v3, v3, v19 :: v_dual_lshlrev_b32 v54, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v16
; GFX11-NEXT: v_dual_min_f32 v19, v82, v81 :: v_dual_lshlrev_b32 v64, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v129, v19, 16, 1
; GFX11-NEXT: v_or_b32_e32 v130, 0x400000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX11-NEXT: v_bfe_u32 v119, v3, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX11-NEXT: v_add3_u32 v129, v129, v19, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v0
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_min_f32 v17, v86, v85 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: v_dual_min_f32 v8, v8, v24 :: v_dual_lshlrev_b32 v39, 16, v27
; GFX11-NEXT: v_or_b32_e32 v128, 0x400000, v3
; GFX11-NEXT: v_add3_u32 v119, v119, v3, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v145, v17, 16, 1
; GFX11-NEXT: v_or_b32_e32 v146, 0x400000, v17
; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-NEXT: v_lshlrev_b32_e32 v70, 16, v4
; GFX11-NEXT: v_add3_u32 v145, v145, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX11-NEXT: v_min_f32_e32 v2, v2, v18
; GFX11-NEXT: v_min_f32_e32 v0, v0, v16
; GFX11-NEXT: v_dual_min_f32 v24, v64, v55 :: v_dual_lshlrev_b32 v37, 16, v28
; GFX11-NEXT: v_min_f32_e32 v7, v7, v23
; GFX11-NEXT: v_dual_min_f32 v23, v66, v65 :: v_dual_min_f32 v18, v84, v83
; GFX11-NEXT: v_dual_min_f32 v9, v9, v25 :: v_dual_and_b32 v28, 0xffff0000, v28
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v85, v24, 16, 1
; GFX11-NEXT: v_bfe_u32 v97, v23, 16, 1
; GFX11-NEXT: v_or_b32_e32 v86, 0x400000, v24
; GFX11-NEXT: v_or_b32_e32 v98, 0x400000, v23
; GFX11-NEXT: v_bfe_u32 v87, v7, 16, 1
; GFX11-NEXT: v_add3_u32 v85, v85, v24, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v20
; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-NEXT: v_add3_u32 v97, v97, v23, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v7
; GFX11-NEXT: v_add3_u32 v87, v87, v7, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_min_f32_e32 v4, v4, v20
; GFX11-NEXT: v_min_f32_e32 v20, v80, v71
; GFX11-NEXT: v_bfe_u32 v71, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX11-NEXT: v_dual_min_f32 v21, v70, v69 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v71, v71, v9, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-NEXT: v_dual_min_f32 v10, v10, v26 :: v_dual_and_b32 v29, 0xffff0000, v29
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-NEXT: v_min_f32_e32 v26, v52, v51
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_min_f32_e32 v6, v6, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_dual_min_f32 v11, v11, v27 :: v_dual_lshlrev_b32 v34, 16, v14
; GFX11-NEXT: v_dual_min_f32 v22, v68, v67 :: v_dual_lshlrev_b32 v33, 16, v30
; GFX11-NEXT: v_dual_min_f32 v27, v50, v49 :: v_dual_lshlrev_b32 v38, 16, v12
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_dual_min_f32 v25, v54, v53 :: v_dual_and_b32 v12, 0xffff0000, v12
; GFX11-NEXT: v_dual_min_f32 v13, v13, v29 :: v_dual_and_b32 v30, 0xffff0000, v30
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_f32_e32 v29, v38, v37
; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX11-NEXT: v_dual_min_f32 v12, v12, v28 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_f32_e32 v14, v14, v30
; GFX11-NEXT: v_min_f32_e32 v28, v48, v39
; GFX11-NEXT: v_dual_min_f32 v30, v36, v35 :: v_dual_min_f32 v33, v34, v33
; GFX11-NEXT: v_bfe_u32 v39, v13, 16, 1
; GFX11-NEXT: v_bfe_u32 v35, v14, 16, 1
; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v14
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v37, v30, 16, 1
; GFX11-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-NEXT: v_add3_u32 v35, v35, v14, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
; GFX11-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX11-NEXT: v_add3_u32 v37, v37, v30, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v13
; GFX11-NEXT: v_bfe_u32 v49, v29, 16, 1
; GFX11-NEXT: v_add3_u32 v39, v39, v13, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v34, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v29
; GFX11-NEXT: v_bfe_u32 v51, v12, 16, 1
; GFX11-NEXT: v_add3_u32 v49, v49, v29, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX11-NEXT: v_cndmask_b32_e32 v14, v35, v36, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-NEXT: v_bfe_u32 v53, v28, 16, 1
; GFX11-NEXT: v_add3_u32 v51, v51, v12, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v28
; GFX11-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v30, v37, v38, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_add3_u32 v53, v53, v28, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX11-NEXT: v_bfe_u32 v65, v27, 16, 1
; GFX11-NEXT: v_add3_u32 v55, v55, v11, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v13, v39, v48, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v27
; GFX11-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX11-NEXT: v_add3_u32 v65, v65, v27, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX11-NEXT: v_cndmask_b32_e32 v29, v49, v50, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_bfe_u32 v69, v26, 16, 1
; GFX11-NEXT: v_add3_u32 v67, v67, v10, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v26
; GFX11-NEXT: v_bfe_u32 v81, v25, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v12, v51, v52, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-NEXT: v_add3_u32 v69, v69, v26, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v25
; GFX11-NEXT: v_bfe_u32 v83, v8, 16, 1
; GFX11-NEXT: v_add3_u32 v81, v81, v25, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v83, v83, v8, 0x7fff
; GFX11-NEXT: v_bfe_u32 v99, v6, 16, 1
; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v6
; GFX11-NEXT: v_cndmask_b32_e32 v11, v55, v64, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-NEXT: v_bfe_u32 v101, v22, 16, 1
; GFX11-NEXT: v_add3_u32 v99, v99, v6, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v22
; GFX11-NEXT: v_bfe_u32 v113, v21, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v27, v65, v66, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-NEXT: v_add3_u32 v101, v101, v22, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v114, 0x400000, v21
; GFX11-NEXT: v_bfe_u32 v115, v4, 16, 1
; GFX11-NEXT: v_add3_u32 v113, v113, v21, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v10, v67, v68, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-NEXT: v_or_b32_e32 v116, 0x400000, v4
; GFX11-NEXT: v_bfe_u32 v117, v20, 16, 1
; GFX11-NEXT: v_add3_u32 v115, v115, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v118, 0x400000, v20
; GFX11-NEXT: v_cndmask_b32_e32 v26, v69, v70, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v117, v117, v20, 0x7fff
; GFX11-NEXT: v_bfe_u32 v133, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v134, 0x400000, v18
; GFX11-NEXT: v_bfe_u32 v147, v0, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v9, v71, v80, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-NEXT: v_add3_u32 v133, v133, v18, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v147, v147, v0, 0x7fff
; GFX11-NEXT: v_bfe_u32 v131, v2, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v25, v81, v82, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_or_b32_e32 v132, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v9, v9, v26, 0x7060302
; GFX11-NEXT: v_add3_u32 v131, v131, v2, 0x7fff
; GFX11-NEXT: v_perm_b32 v10, v10, v27, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v83, v84, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-NEXT: v_perm_b32 v11, v11, v28, 0x7060302
; GFX11-NEXT: v_perm_b32 v12, v12, v29, 0x7060302
; GFX11-NEXT: v_perm_b32 v13, v13, v30, 0x7060302
; GFX11-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v24, v85, v86, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v7, v87, v96, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v23, v97, v98, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_cndmask_b32_e32 v6, v99, v100, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v22, v101, v102, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_cndmask_b32_e32 v5, v103, v112, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v21, v113, v114, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_cndmask_b32_e32 v4, v115, v116, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v20, v117, v118, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_cndmask_b32_e32 v19, v129, v130, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_cndmask_b32_e32 v18, v133, v134, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v135, v144, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v17, v145, v146, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v147, v33, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v131, v132, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v119, v128, vcc_lo
; GFX11-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_min_f32 v17, v31, v17 :: v_dual_and_b32 v18, 0xffff0000, v32
; GFX11-NEXT: v_min_f32_e32 v15, v15, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX11-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v17, v18, v20, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <32 x bfloat> @llvm.minnum.v32bf16(<32 x bfloat> %a, <32 x bfloat> %b)
ret <32 x bfloat> %op
}
declare bfloat @llvm.maxnum.bf16(bfloat, bfloat)
declare <2 x bfloat> @llvm.maxnum.v2bf16(<2 x bfloat>, <2 x bfloat>)
declare <3 x bfloat> @llvm.maxnum.v3bf16(<3 x bfloat>, <3 x bfloat>)
declare <4 x bfloat> @llvm.maxnum.v4bf16(<4 x bfloat>, <4 x bfloat>)
declare <8 x bfloat> @llvm.maxnum.v8bf16(<8 x bfloat>, <8 x bfloat>)
declare <16 x bfloat> @llvm.maxnum.v16bf16(<16 x bfloat>, <16 x bfloat>)
declare <32 x bfloat> @llvm.maxnum.v32bf16(<32 x bfloat>, <32 x bfloat>)
define bfloat @v_maxnum_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_maxnum_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_max_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_max_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.maxnum.bf16(bfloat %a, bfloat %b)
ret bfloat %op
}
define <2 x bfloat> @v_maxnum_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v1, v1, v3
; GCN-NEXT: v_max_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v1, v1, v3
; GFX7-NEXT: v_max_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_max_f32_e32 v2, v3, v2
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_max_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_max_f32_e32 v2, v3, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_max_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_max_f32_e32 v2, v3, v2
; GFX10-NEXT: v_max_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_max_f32_e32 v0, v0, v1
; GFX11-NEXT: v_max_f32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <2 x bfloat> @llvm.maxnum.v2bf16(<2 x bfloat> %a, <2 x bfloat> %b)
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_maxnum_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v2, v2, v5
; GCN-NEXT: v_max_f32_e32 v1, v1, v4
; GCN-NEXT: v_max_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v2, v2, v5
; GFX7-NEXT: v_max_f32_e32 v1, v1, v4
; GFX7-NEXT: v_max_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_max_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_max_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_max_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_max_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_max_f32_e32 v3, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_max_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_max_f32_e32 v4, v5, v4
; GFX10-NEXT: v_max_f32_e32 v0, v0, v2
; GFX10-NEXT: v_max_f32_e32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_maxnum_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_max_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11TRUE16-NEXT: v_dual_max_f32 v0, v0, v2 :: v_dual_max_f32 v1, v1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_maxnum_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_max_f32 v4, v5, v4 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11FAKE16-NEXT: v_dual_max_f32 v0, v0, v2 :: v_dual_max_f32 v1, v1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = call <3 x bfloat> @llvm.maxnum.v3bf16(<3 x bfloat> %a, <3 x bfloat> %b)
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_maxnum_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v3, v3, v7
; GCN-NEXT: v_max_f32_e32 v2, v2, v6
; GCN-NEXT: v_max_f32_e32 v1, v1, v5
; GCN-NEXT: v_max_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v3, v3, v7
; GFX7-NEXT: v_max_f32_e32 v2, v2, v6
; GFX7-NEXT: v_max_f32_e32 v1, v1, v5
; GFX7-NEXT: v_max_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_max_f32_e32 v4, v5, v4
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_max_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_max_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_max_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v4, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_max_f32_e32 v4, v5, v4
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_max_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_max_f32_e32 v3, v5, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_max_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_max_f32_e32 v4, v5, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_max_f32_e32 v1, v1, v3
; GFX10-NEXT: v_max_f32_e32 v3, v7, v6
; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_max_f32_e32 v0, v0, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_max_f32 v0, v0, v2 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_max_f32_e32 v1, v1, v3
; GFX11-NEXT: v_dual_max_f32 v3, v7, v6 :: v_dual_max_f32 v4, v5, v4
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v9, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <4 x bfloat> @llvm.maxnum.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b)
ret <4 x bfloat> %op
}
define <8 x bfloat> @v_maxnum_v8bf16(<8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v7, v7, v15
; GCN-NEXT: v_max_f32_e32 v6, v6, v14
; GCN-NEXT: v_max_f32_e32 v5, v5, v13
; GCN-NEXT: v_max_f32_e32 v4, v4, v12
; GCN-NEXT: v_max_f32_e32 v3, v3, v11
; GCN-NEXT: v_max_f32_e32 v2, v2, v10
; GCN-NEXT: v_max_f32_e32 v1, v1, v9
; GCN-NEXT: v_max_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v7, v7, v15
; GFX7-NEXT: v_max_f32_e32 v6, v6, v14
; GFX7-NEXT: v_max_f32_e32 v5, v5, v13
; GFX7-NEXT: v_max_f32_e32 v4, v4, v12
; GFX7-NEXT: v_max_f32_e32 v3, v3, v11
; GFX7-NEXT: v_max_f32_e32 v2, v2, v10
; GFX7-NEXT: v_max_f32_e32 v1, v1, v9
; GFX7-NEXT: v_max_f32_e32 v0, v0, v8
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX8-NEXT: v_max_f32_e32 v8, v9, v8
; GFX8-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v8
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 0x7fff, v9
; GFX8-NEXT: v_max_f32_e32 v3, v3, v7
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v3
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX8-NEXT: v_max_f32_e32 v7, v9, v7
; GFX8-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v7
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_max_f32_e32 v2, v2, v6
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX8-NEXT: v_max_f32_e32 v6, v9, v6
; GFX8-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_max_f32_e32 v1, v1, v5
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX8-NEXT: v_max_f32_e32 v5, v9, v5
; GFX8-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v5
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_max_f32_e32 v0, v0, v4
; GFX8-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v5, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v7, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v8, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX9-NEXT: v_max_f32_e32 v8, v9, v8
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v9, v8, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_max_f32_e32 v3, v3, v7
; GFX9-NEXT: v_add3_u32 v9, v9, v8, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v8, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v7, v7, v3, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v2
; GFX9-NEXT: v_max_f32_e32 v7, v9, v7
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX9-NEXT: v_max_f32_e32 v2, v2, v6
; GFX9-NEXT: v_add3_u32 v9, v9, v7, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_bfe_u32 v6, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v7, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v6, v6, v2, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v6, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v1
; GFX9-NEXT: v_max_f32_e32 v6, v9, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v9, v6, 16, 1
; GFX9-NEXT: v_max_f32_e32 v1, v1, v5
; GFX9-NEXT: v_add3_u32 v9, v9, v6, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v5, v5, v1, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v9, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX9-NEXT: v_max_f32_e32 v5, v9, v5
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v9, v5, 16, 1
; GFX9-NEXT: v_max_f32_e32 v0, v0, v4
; GFX9-NEXT: v_add3_u32 v9, v9, v5, s4
; GFX9-NEXT: v_or_b32_e32 v10, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v10, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v9, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v5, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v7, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v8, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_max_f32_e32 v8, v9, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_max_f32_e32 v3, v3, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX10-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX10-NEXT: v_max_f32_e32 v7, v10, v9
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX10-NEXT: v_max_f32_e32 v2, v2, v6
; GFX10-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX10-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX10-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX10-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_max_f32_e32 v6, v10, v6
; GFX10-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX10-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_max_f32_e32 v1, v1, v5
; GFX10-NEXT: v_max_f32_e32 v5, v15, v13
; GFX10-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX10-NEXT: v_max_f32_e32 v0, v0, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX10-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v4, v4, v10, vcc_lo
; GFX10-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX10-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v7
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v15, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_max_f32 v8, v9, v8 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v11, v8, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_max_f32_e32 v3, v3, v7
; GFX11-NEXT: v_max_f32_e32 v7, v10, v9
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v10, v11, v8, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v12, v7, 16, 1
; GFX11-NEXT: v_or_b32_e32 v14, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v9, v11, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v11, v12, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v10, 16, v1
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_cndmask_b32 v7, v11, v12 :: v_dual_max_f32 v2, v2, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_bfe_u32 v13, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_max_f32_e32 v6, v10, v6
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v10, v13, v2, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v12, v6, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v2, v10, v11, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v13, 16, v4
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_perm_b32 v2, v2, v7, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_max_f32_e32 v0, v0, v4
; GFX11-NEXT: v_add3_u32 v4, v12, v6, 0x7fff
; GFX11-NEXT: v_dual_max_f32 v1, v1, v5 :: v_dual_cndmask_b32 v4, v4, v10
; GFX11-NEXT: v_max_f32_e32 v5, v15, v13
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v11, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v13, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v15, 0x400000, v1
; GFX11-NEXT: v_bfe_u32 v12, v5, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v6, v11, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_add3_u32 v10, v12, v5, 0x7fff
; GFX11-NEXT: v_add3_u32 v12, v13, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v12, v13, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v5, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v15, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v1, v1, v4, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v9, v14, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v3, v3, v8, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <8 x bfloat> @llvm.maxnum.v8bf16(<8 x bfloat> %a, <8 x bfloat> %b)
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_maxnum_v16bf16(<16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_max_f32_e32 v14, v14, v30
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_max_f32_e32 v13, v13, v29
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_max_f32_e32 v12, v12, v28
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_max_f32_e32 v11, v11, v27
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_max_f32_e32 v10, v10, v26
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_max_f32_e32 v9, v9, v25
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_max_f32_e32 v8, v8, v24
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_max_f32_e32 v7, v7, v23
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_max_f32_e32 v6, v6, v22
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_max_f32_e32 v5, v5, v21
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_max_f32_e32 v4, v4, v20
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_max_f32_e32 v3, v3, v19
; GCN-NEXT: v_max_f32_e32 v2, v2, v18
; GCN-NEXT: v_max_f32_e32 v1, v1, v17
; GCN-NEXT: v_max_f32_e32 v0, v0, v16
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_max_f32_e32 v15, v15, v16
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_max_f32_e32 v6, v6, v22
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_max_f32_e32 v14, v14, v30
; GFX7-NEXT: v_max_f32_e32 v13, v13, v29
; GFX7-NEXT: v_max_f32_e32 v12, v12, v28
; GFX7-NEXT: v_max_f32_e32 v11, v11, v27
; GFX7-NEXT: v_max_f32_e32 v10, v10, v26
; GFX7-NEXT: v_max_f32_e32 v9, v9, v25
; GFX7-NEXT: v_max_f32_e32 v8, v8, v24
; GFX7-NEXT: v_max_f32_e32 v7, v7, v23
; GFX7-NEXT: v_max_f32_e32 v5, v5, v21
; GFX7-NEXT: v_max_f32_e32 v4, v4, v20
; GFX7-NEXT: v_max_f32_e32 v3, v3, v19
; GFX7-NEXT: v_max_f32_e32 v2, v2, v18
; GFX7-NEXT: v_max_f32_e32 v1, v1, v17
; GFX7-NEXT: v_max_f32_e32 v0, v0, v16
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_max_f32_e32 v15, v15, v22
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX8-NEXT: v_max_f32_e32 v16, v17, v16
; GFX8-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v16
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v7, v7, v15
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX8-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v7
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX8-NEXT: v_max_f32_e32 v15, v17, v15
; GFX8-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v15
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v6, v6, v14
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX8-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v14, vcc, v14, v6
; GFX8-NEXT: v_add_u32_e32 v14, vcc, s4, v14
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX8-NEXT: v_max_f32_e32 v14, v17, v14
; GFX8-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v14
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v5, v5, v13
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v13, vcc, v13, v5
; GFX8-NEXT: v_add_u32_e32 v13, vcc, s4, v13
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX8-NEXT: v_max_f32_e32 v13, v17, v13
; GFX8-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v13
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v4, v4, v12
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v12, vcc, v12, v4
; GFX8-NEXT: v_add_u32_e32 v12, vcc, s4, v12
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX8-NEXT: v_max_f32_e32 v12, v17, v12
; GFX8-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v12
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v3, v3, v11
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v11, vcc, v11, v3
; GFX8-NEXT: v_add_u32_e32 v11, vcc, s4, v11
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX8-NEXT: v_max_f32_e32 v11, v17, v11
; GFX8-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v11
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v2, v2, v10
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v10, vcc, v10, v2
; GFX8-NEXT: v_add_u32_e32 v10, vcc, s4, v10
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX8-NEXT: v_max_f32_e32 v10, v17, v10
; GFX8-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v10
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v1, v1, v9
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v9, vcc, v9, v1
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v9
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX8-NEXT: v_max_f32_e32 v9, v17, v9
; GFX8-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v9
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_max_f32_e32 v0, v0, v8
; GFX8-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v0
; GFX8-NEXT: v_add_u32_e32 v8, vcc, s4, v8
; GFX8-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v9, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v10, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v11, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v12, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v13, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v14, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v15, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v16, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX9-NEXT: v_max_f32_e32 v16, v17, v16
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v17, v16, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_max_f32_e32 v7, v7, v15
; GFX9-NEXT: v_add3_u32 v17, v17, v16, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v16
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT: v_bfe_u32 v15, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v16, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v15, v15, v7, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v15, 16, v14
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v6
; GFX9-NEXT: v_max_f32_e32 v15, v17, v15
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v17, v15, 16, 1
; GFX9-NEXT: v_max_f32_e32 v6, v6, v14
; GFX9-NEXT: v_add3_u32 v17, v17, v15, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v15
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT: v_bfe_u32 v14, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v14, v14, v6, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v14, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v14, 16, v13
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v5
; GFX9-NEXT: v_max_f32_e32 v14, v17, v14
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v17, v14, 16, 1
; GFX9-NEXT: v_max_f32_e32 v5, v5, v13
; GFX9-NEXT: v_add3_u32 v17, v17, v14, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_bfe_u32 v13, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v13, v13, v5, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v13, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v13, 16, v12
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v4
; GFX9-NEXT: v_max_f32_e32 v13, v17, v13
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v17, v13, 16, 1
; GFX9-NEXT: v_max_f32_e32 v4, v4, v12
; GFX9-NEXT: v_add3_u32 v17, v17, v13, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_bfe_u32 v12, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v13, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v12, v12, v4, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v12, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v3
; GFX9-NEXT: v_max_f32_e32 v12, v17, v12
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX9-NEXT: v_max_f32_e32 v3, v3, v11
; GFX9-NEXT: v_add3_u32 v17, v17, v12, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_bfe_u32 v11, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v12, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v11, v11, v3, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v11, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v11, 16, v10
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GFX9-NEXT: v_max_f32_e32 v11, v17, v11
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v17, v11, 16, 1
; GFX9-NEXT: v_max_f32_e32 v2, v2, v10
; GFX9-NEXT: v_add3_u32 v17, v17, v11, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_bfe_u32 v10, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v11, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v10, v10, v2, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v10, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v10, 16, v9
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v1
; GFX9-NEXT: v_max_f32_e32 v10, v17, v10
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v17, v10, 16, 1
; GFX9-NEXT: v_max_f32_e32 v1, v1, v9
; GFX9-NEXT: v_add3_u32 v17, v17, v10, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v10, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v9, v9, v1, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v17, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v9, 16, v8
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v0
; GFX9-NEXT: v_max_f32_e32 v9, v17, v9
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v17, v9, 16, 1
; GFX9-NEXT: v_max_f32_e32 v0, v0, v8
; GFX9-NEXT: v_add3_u32 v17, v17, v9, s4
; GFX9-NEXT: v_or_b32_e32 v18, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v9, v17, v18, vcc
; GFX9-NEXT: v_add3_u32 v8, v8, v0, s4
; GFX9-NEXT: v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v8, v17, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v9, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v10, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v11, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v12, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v13, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v14, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v15, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v16, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_max_f32_e32 v16, v17, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX10-NEXT: v_max_f32_e32 v7, v7, v15
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX10-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX10-NEXT: v_max_f32_e32 v17, v18, v17
; GFX10-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX10-NEXT: v_max_f32_e32 v6, v6, v14
; GFX10-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX10-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v15, v15, v20, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v5
; GFX10-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX10-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX10-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX10-NEXT: v_max_f32_e32 v17, v20, v19
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX10-NEXT: v_max_f32_e32 v5, v5, v13
; GFX10-NEXT: v_cndmask_b32_e32 v14, v14, v16, vcc_lo
; GFX10-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX10-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v6, v16, v13, vcc_lo
; GFX10-NEXT: v_max_f32_e32 v13, v19, v18
; GFX10-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX10-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX10-NEXT: v_max_f32_e32 v4, v4, v12
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX10-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_max_f32_e32 v12, v18, v12
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX10-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX10-NEXT: v_max_f32_e32 v3, v3, v11
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX10-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX10-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX10-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX10-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX10-NEXT: v_max_f32_e32 v18, v19, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX10-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX10-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX10-NEXT: v_max_f32_e32 v2, v2, v10
; GFX10-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX10-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX10-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX10-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX10-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX10-NEXT: v_max_f32_e32 v19, v22, v20
; GFX10-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX10-NEXT: v_max_f32_e32 v1, v1, v9
; GFX10-NEXT: v_max_f32_e32 v9, v22, v20
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX10-NEXT: v_max_f32_e32 v0, v0, v8
; GFX10-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX10-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX10-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX10-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX10-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX10-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX10-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v16, 16, v15
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_max_f32 v16, v17, v16 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v14
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_max_f32_e32 v17, v18, v17
; GFX11-NEXT: v_max_f32_e32 v6, v6, v14
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfe_u32 v21, v17, 16, 1
; GFX11-NEXT: v_add3_u32 v14, v21, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_max_f32_e32 v7, v7, v15
; GFX11-NEXT: v_bfe_u32 v15, v16, 16, 1
; GFX11-NEXT: v_add3_u32 v15, v15, v16, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v16, 0x400000, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_cndmask_b32 v15, v15, v20 :: v_dual_lshlrev_b32 v20, 16, v5
; GFX11-NEXT: v_bfe_u32 v19, v7, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v18, v19, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v7, v18, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v18, v6, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v7, v7, v15, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_dual_max_f32 v17, v20, v19 :: v_dual_cndmask_b32 v14, v14, v16
; GFX11-NEXT: v_add3_u32 v16, v18, v6, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v4
; GFX11-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_bfe_u32 v20, v17, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_max_f32_e32 v4, v4, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v12, 16, v11
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_max_f32_e32 v5, v5, v13
; GFX11-NEXT: v_or_b32_e32 v13, 0x400000, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_cndmask_b32 v6, v16, v13 :: v_dual_max_f32 v13, v19, v18
; GFX11-NEXT: v_add3_u32 v16, v20, v17, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_perm_b32 v6, v6, v14, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v3
; GFX11-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v5
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_max_f32_e32 v12, v18, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
; GFX11-NEXT: v_bfe_u32 v21, v13, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v18, 16, v10
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v12
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v5, v19, v20, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v21, v13, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v19, 0x400000, v13
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v4
; GFX11-NEXT: v_perm_b32 v5, v5, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v19, vcc_lo
; GFX11-NEXT: v_bfe_u32 v17, v12, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_lshlrev_b32_e32 v19, 16, v2
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_add3_u32 v17, v17, v12, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_max_f32_e32 v18, v19, v18
; GFX11-NEXT: v_cndmask_b32_e32 v12, v17, v22, vcc_lo
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v1
; GFX11-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v23, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v17, 0x400000, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_max_f32 v2, v2, v10 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_max_f32_e32 v3, v3, v11
; GFX11-NEXT: v_add3_u32 v11, v20, v4, 0x7fff
; GFX11-NEXT: v_add3_u32 v10, v23, v18, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v20, v3, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v19, v20, v3, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v20, vcc_lo
; GFX11-NEXT: v_bfe_u32 v19, v2, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v9
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_or_b32_e32 v18, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v3, v3, v12, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v10, v10, v17, vcc_lo
; GFX11-NEXT: v_add3_u32 v17, v19, v2, 0x7fff
; GFX11-NEXT: v_max_f32_e32 v19, v22, v20
; GFX11-NEXT: v_lshlrev_b32_e32 v20, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v22, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: v_bfe_u32 v23, v19, 16, 1
; GFX11-NEXT: v_dual_max_f32 v0, v0, v8 :: v_dual_max_f32 v1, v1, v9
; GFX11-NEXT: v_max_f32_e32 v9, v22, v20
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add3_u32 v20, v23, v19, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v19
; GFX11-NEXT: v_or_b32_e32 v25, 0x400000, v0
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v23, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v24, 0x400000, v9
; GFX11-NEXT: v_cndmask_b32_e32 v19, v20, v22, vcc_lo
; GFX11-NEXT: v_or_b32_e32 v22, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_bfe_u32 v20, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v23, v23, v9, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v8, v22, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v20, v20, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_perm_b32 v1, v1, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v23, v24, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v20, v25, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v0, v0, v8, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v17, v18, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_perm_b32 v2, v2, v10, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v4, v11, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v4, v4, v13, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <16 x bfloat> @llvm.maxnum.v16bf16(<16 x bfloat> %a, <16 x bfloat> %b)
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_maxnum_v32bf16(<32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_maxnum_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:124
; GCN-NEXT: v_max_f32_e32 v31, v31, v32
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:120
; GCN-NEXT: v_max_f32_e32 v30, v30, v32
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:116
; GCN-NEXT: v_max_f32_e32 v29, v29, v32
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:112
; GCN-NEXT: v_max_f32_e32 v28, v28, v32
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:108
; GCN-NEXT: v_max_f32_e32 v27, v27, v32
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:104
; GCN-NEXT: v_max_f32_e32 v26, v26, v32
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:100
; GCN-NEXT: v_max_f32_e32 v25, v25, v32
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
; GCN-NEXT: v_max_f32_e32 v24, v24, v32
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:92
; GCN-NEXT: v_max_f32_e32 v23, v23, v32
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:88
; GCN-NEXT: v_max_f32_e32 v22, v22, v32
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:84
; GCN-NEXT: v_max_f32_e32 v21, v21, v32
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:80
; GCN-NEXT: v_max_f32_e32 v20, v20, v32
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:76
; GCN-NEXT: v_max_f32_e32 v19, v19, v32
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:72
; GCN-NEXT: v_max_f32_e32 v18, v18, v32
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:68
; GCN-NEXT: v_max_f32_e32 v17, v17, v32
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:64
; GCN-NEXT: v_max_f32_e32 v16, v16, v32
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:60
; GCN-NEXT: v_max_f32_e32 v15, v15, v32
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:56
; GCN-NEXT: v_max_f32_e32 v14, v14, v32
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:52
; GCN-NEXT: v_max_f32_e32 v13, v13, v32
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:48
; GCN-NEXT: v_max_f32_e32 v12, v12, v32
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:44
; GCN-NEXT: v_max_f32_e32 v11, v11, v32
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
; GCN-NEXT: v_max_f32_e32 v10, v10, v32
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:36
; GCN-NEXT: v_max_f32_e32 v9, v9, v32
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:32
; GCN-NEXT: v_max_f32_e32 v8, v8, v32
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:28
; GCN-NEXT: v_max_f32_e32 v7, v7, v32
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:24
; GCN-NEXT: v_max_f32_e32 v6, v6, v32
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
; GCN-NEXT: v_max_f32_e32 v5, v5, v32
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:16
; GCN-NEXT: v_max_f32_e32 v4, v4, v32
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:12
; GCN-NEXT: v_max_f32_e32 v3, v3, v32
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:8
; GCN-NEXT: v_max_f32_e32 v2, v2, v32
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:4
; GCN-NEXT: v_max_f32_e32 v1, v1, v32
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GCN-NEXT: v_max_f32_e32 v0, v0, v32
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_maxnum_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: v_max_f32_e32 v31, v31, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v30, v30, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:120
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v29, v29, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:116
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v28, v28, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:112
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v27, v27, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:108
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v26, v26, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:104
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v25, v25, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:100
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v24, v24, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:96
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v23, v23, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:92
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v22, v22, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v21, v21, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v20, v20, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:80
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v19, v19, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v18, v18, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:72
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v17, v17, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v16, v16, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v15, v15, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v14, v14, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v13, v13, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v12, v12, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v11, v11, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v10, v10, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v9, v9, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v8, v8, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:32
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v7, v7, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v6, v6, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:24
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v5, v5, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v4, v4, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v3, v3, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v2, v2, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v1, v1, v32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_and_b32_e32 v32, 0xffff0000, v32
; GFX7-NEXT: v_max_f32_e32 v0, v0, v32
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_maxnum_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX8-NEXT: v_max_f32_e32 v31, v32, v31
; GFX8-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v32, vcc, v32, v31
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX8-NEXT: v_add_u32_e32 v32, vcc, s4, v32
; GFX8-NEXT: v_max_f32_e32 v14, v14, v30
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX8-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX8-NEXT: v_add_u32_e32 v30, vcc, v30, v14
; GFX8-NEXT: v_add_u32_e32 v30, vcc, s4, v30
; GFX8-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX8-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX8-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX8-NEXT: v_max_f32_e32 v32, v32, v30
; GFX8-NEXT: buffer_load_dword v30, off, s[0:3], s32
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX8-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX8-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX8-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX8-NEXT: v_max_f32_e32 v13, v13, v29
; GFX8-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX8-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX8-NEXT: v_alignbit_b32 v14, v14, v31, 16
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v34, 16, v30
; GFX8-NEXT: v_max_f32_e32 v33, v33, v34
; GFX8-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX8-NEXT: v_max_f32_e32 v30, v15, v30
; GFX8-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX8-NEXT: v_add_u32_e32 v15, vcc, v15, v33
; GFX8-NEXT: v_add_u32_e32 v15, vcc, s4, v15
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX8-NEXT: v_bfe_u32 v33, v30, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v30
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v30
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX8-NEXT: v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX8-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v32
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX8-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v29, vcc, v29, v13
; GFX8-NEXT: v_add_u32_e32 v29, vcc, s4, v29
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v13
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX8-NEXT: v_cndmask_b32_e32 v13, v29, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v12
; GFX8-NEXT: v_max_f32_e32 v29, v33, v29
; GFX8-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v29
; GFX8-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX8-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v12, v12, v28
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX8-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v28, vcc, v28, v12
; GFX8-NEXT: v_add_u32_e32 v28, vcc, s4, v28
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX8-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX8-NEXT: v_max_f32_e32 v28, v33, v28
; GFX8-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v28
; GFX8-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX8-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v11, v11, v27
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX8-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v27, vcc, v27, v11
; GFX8-NEXT: v_add_u32_e32 v27, vcc, s4, v27
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX8-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX8-NEXT: v_max_f32_e32 v27, v33, v27
; GFX8-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v27
; GFX8-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX8-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v10, v10, v26
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX8-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v26, vcc, v26, v10
; GFX8-NEXT: v_add_u32_e32 v26, vcc, s4, v26
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX8-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX8-NEXT: v_max_f32_e32 v26, v33, v26
; GFX8-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v26
; GFX8-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX8-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v9, v9, v25
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX8-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v25, vcc, v25, v9
; GFX8-NEXT: v_add_u32_e32 v25, vcc, s4, v25
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX8-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX8-NEXT: v_max_f32_e32 v25, v33, v25
; GFX8-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v25
; GFX8-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX8-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v8, v8, v24
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX8-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v24, vcc, v24, v8
; GFX8-NEXT: v_add_u32_e32 v24, vcc, s4, v24
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX8-NEXT: v_max_f32_e32 v24, v33, v24
; GFX8-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v24
; GFX8-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX8-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v7, v7, v23
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX8-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v23, vcc, v23, v7
; GFX8-NEXT: v_add_u32_e32 v23, vcc, s4, v23
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX8-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX8-NEXT: v_max_f32_e32 v23, v33, v23
; GFX8-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v23
; GFX8-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v6, v6, v22
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX8-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v22, vcc, v22, v6
; GFX8-NEXT: v_add_u32_e32 v22, vcc, s4, v22
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX8-NEXT: v_max_f32_e32 v22, v33, v22
; GFX8-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v22
; GFX8-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v5, v5, v21
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX8-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v21, vcc, v21, v5
; GFX8-NEXT: v_add_u32_e32 v21, vcc, s4, v21
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX8-NEXT: v_max_f32_e32 v21, v33, v21
; GFX8-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v21
; GFX8-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v4, v4, v20
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX8-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v20, vcc, v20, v4
; GFX8-NEXT: v_add_u32_e32 v20, vcc, s4, v20
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX8-NEXT: v_max_f32_e32 v20, v33, v20
; GFX8-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v20
; GFX8-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v3, v3, v19
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX8-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v19, vcc, v19, v3
; GFX8-NEXT: v_add_u32_e32 v19, vcc, s4, v19
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX8-NEXT: v_max_f32_e32 v19, v33, v19
; GFX8-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v19
; GFX8-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v2, v2, v18
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX8-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v18, vcc, v18, v2
; GFX8-NEXT: v_add_u32_e32 v18, vcc, s4, v18
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX8-NEXT: v_max_f32_e32 v18, v33, v18
; GFX8-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v18
; GFX8-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v1, v1, v17
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX8-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v17, vcc, v17, v1
; GFX8-NEXT: v_add_u32_e32 v17, vcc, s4, v17
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX8-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX8-NEXT: v_max_f32_e32 v17, v33, v17
; GFX8-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX8-NEXT: v_add_u32_e32 v33, vcc, v33, v17
; GFX8-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v33, vcc, s4, v33
; GFX8-NEXT: v_max_f32_e32 v0, v0, v16
; GFX8-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX8-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX8-NEXT: v_add_u32_e32 v16, vcc, v16, v0
; GFX8-NEXT: v_add_u32_e32 v16, vcc, s4, v16
; GFX8-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v30
; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX8-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX8-NEXT: v_alignbit_b32 v0, v0, v17, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v18, 16
; GFX8-NEXT: v_alignbit_b32 v2, v2, v19, 16
; GFX8-NEXT: v_alignbit_b32 v3, v3, v20, 16
; GFX8-NEXT: v_alignbit_b32 v4, v4, v21, 16
; GFX8-NEXT: v_alignbit_b32 v5, v5, v22, 16
; GFX8-NEXT: v_alignbit_b32 v6, v6, v23, 16
; GFX8-NEXT: v_alignbit_b32 v7, v7, v24, 16
; GFX8-NEXT: v_alignbit_b32 v8, v8, v25, 16
; GFX8-NEXT: v_alignbit_b32 v9, v9, v26, 16
; GFX8-NEXT: v_alignbit_b32 v10, v10, v27, 16
; GFX8-NEXT: v_alignbit_b32 v11, v11, v28, 16
; GFX8-NEXT: v_alignbit_b32 v12, v12, v29, 16
; GFX8-NEXT: v_alignbit_b32 v13, v13, v32, 16
; GFX8-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_maxnum_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v31, 16, v30
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT: v_max_f32_e32 v31, v32, v31
; GFX9-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v32, v31, 16, 1
; GFX9-NEXT: v_max_f32_e32 v14, v14, v30
; GFX9-NEXT: v_add3_u32 v32, v32, v31, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v31
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT: v_bfe_u32 v30, v14, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v30, v30, v14, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v14
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT: v_cndmask_b32_e32 v14, v30, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v30, 16, v29
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT: v_max_f32_e32 v30, v32, v30
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT: v_bfe_u32 v32, v30, 16, 1
; GFX9-NEXT: v_max_f32_e32 v13, v13, v29
; GFX9-NEXT: v_add3_u32 v32, v32, v30, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v30
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT: v_bfe_u32 v29, v13, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc
; GFX9-NEXT: v_add3_u32 v29, v29, v13, s4
; GFX9-NEXT: v_or_b32_e32 v32, 0x400000, v13
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT: v_cndmask_b32_e32 v13, v29, v32, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v29, 16, v28
; GFX9-NEXT: v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT: v_max_f32_e32 v32, v32, v29
; GFX9-NEXT: buffer_load_dword v29, off, s[0:3], s32
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v15
; GFX9-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT: v_max_f32_e32 v12, v12, v28
; GFX9-NEXT: v_bfe_u32 v28, v12, 16, 1
; GFX9-NEXT: v_add3_u32 v28, v28, v12, s4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v34, 16, v29
; GFX9-NEXT: v_max_f32_e32 v33, v33, v34
; GFX9-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT: v_max_f32_e32 v29, v15, v29
; GFX9-NEXT: v_bfe_u32 v15, v33, 16, 1
; GFX9-NEXT: v_add3_u32 v15, v15, v33, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT: v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v15, v15, v34, vcc
; GFX9-NEXT: v_add3_u32 v33, v33, v29, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT: v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT: v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT: v_add3_u32 v33, v33, v32, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT: v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v12
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT: v_cndmask_b32_e32 v12, v28, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v28, 16, v27
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v11
; GFX9-NEXT: v_max_f32_e32 v28, v33, v28
; GFX9-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT: v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT: v_max_f32_e32 v11, v11, v27
; GFX9-NEXT: v_add3_u32 v33, v33, v28, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT: v_bfe_u32 v27, v11, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v27, v27, v11, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v11
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT: v_cndmask_b32_e32 v11, v27, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v27, 16, v26
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v10
; GFX9-NEXT: v_max_f32_e32 v27, v33, v27
; GFX9-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT: v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT: v_max_f32_e32 v10, v10, v26
; GFX9-NEXT: v_add3_u32 v33, v33, v27, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT: v_bfe_u32 v26, v10, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v26, v26, v10, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v10
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT: v_cndmask_b32_e32 v10, v26, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v26, 16, v25
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v9
; GFX9-NEXT: v_max_f32_e32 v26, v33, v26
; GFX9-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT: v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT: v_max_f32_e32 v9, v9, v25
; GFX9-NEXT: v_add3_u32 v33, v33, v26, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT: v_bfe_u32 v25, v9, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v25, v25, v9, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v9
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT: v_cndmask_b32_e32 v9, v25, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v25, 16, v24
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v8
; GFX9-NEXT: v_max_f32_e32 v25, v33, v25
; GFX9-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT: v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT: v_max_f32_e32 v8, v8, v24
; GFX9-NEXT: v_add3_u32 v33, v33, v25, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT: v_bfe_u32 v24, v8, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v24, v24, v8, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_cndmask_b32_e32 v8, v24, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v24, 16, v23
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v7
; GFX9-NEXT: v_max_f32_e32 v24, v33, v24
; GFX9-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT: v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT: v_max_f32_e32 v7, v7, v23
; GFX9-NEXT: v_add3_u32 v33, v33, v24, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT: v_bfe_u32 v23, v7, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v23, v23, v7, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v7
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT: v_cndmask_b32_e32 v7, v23, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v23, 16, v22
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v6
; GFX9-NEXT: v_max_f32_e32 v23, v33, v23
; GFX9-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT: v_max_f32_e32 v6, v6, v22
; GFX9-NEXT: v_add3_u32 v33, v33, v23, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT: v_bfe_u32 v22, v6, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v22, v22, v6, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v22, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v22, 16, v21
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v5
; GFX9-NEXT: v_max_f32_e32 v22, v33, v22
; GFX9-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT: v_max_f32_e32 v5, v5, v21
; GFX9-NEXT: v_add3_u32 v33, v33, v22, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT: v_bfe_u32 v21, v5, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v21, v21, v5, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v5
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v5, v21, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v21, 16, v20
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v4
; GFX9-NEXT: v_max_f32_e32 v21, v33, v21
; GFX9-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT: v_max_f32_e32 v4, v4, v20
; GFX9-NEXT: v_add3_u32 v33, v33, v21, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT: v_bfe_u32 v20, v4, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v20, v20, v4, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v4, v20, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v20, 16, v19
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v3
; GFX9-NEXT: v_max_f32_e32 v20, v33, v20
; GFX9-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT: v_max_f32_e32 v3, v3, v19
; GFX9-NEXT: v_add3_u32 v33, v33, v20, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT: v_bfe_u32 v19, v3, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v19, v19, v3, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v19, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v19, 16, v18
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v2
; GFX9-NEXT: v_max_f32_e32 v19, v33, v19
; GFX9-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT: v_max_f32_e32 v2, v2, v18
; GFX9-NEXT: v_add3_u32 v33, v33, v19, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT: v_bfe_u32 v18, v2, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v18, v18, v2, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v18, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v18, 16, v17
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v1
; GFX9-NEXT: v_max_f32_e32 v18, v33, v18
; GFX9-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT: v_max_f32_e32 v1, v1, v17
; GFX9-NEXT: v_add3_u32 v33, v33, v18, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT: v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v17, v17, v1, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v33, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v17, 16, v16
; GFX9-NEXT: v_lshlrev_b32_e32 v33, 16, v0
; GFX9-NEXT: v_max_f32_e32 v17, v33, v17
; GFX9-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT: v_max_f32_e32 v0, v0, v16
; GFX9-NEXT: v_add3_u32 v33, v33, v17, s4
; GFX9-NEXT: v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT: v_bfe_u32 v16, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT: v_add3_u32 v16, v16, v0, s4
; GFX9-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v16, v33, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v17, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v18, s4
; GFX9-NEXT: v_perm_b32 v2, v2, v19, s4
; GFX9-NEXT: v_perm_b32 v3, v3, v20, s4
; GFX9-NEXT: v_perm_b32 v4, v4, v21, s4
; GFX9-NEXT: v_perm_b32 v5, v5, v22, s4
; GFX9-NEXT: v_perm_b32 v6, v6, v23, s4
; GFX9-NEXT: v_perm_b32 v7, v7, v24, s4
; GFX9-NEXT: v_perm_b32 v8, v8, v25, s4
; GFX9-NEXT: v_perm_b32 v9, v9, v26, s4
; GFX9-NEXT: v_perm_b32 v10, v10, v27, s4
; GFX9-NEXT: v_perm_b32 v11, v11, v28, s4
; GFX9-NEXT: v_perm_b32 v12, v12, v32, s4
; GFX9-NEXT: v_perm_b32 v13, v13, v30, s4
; GFX9-NEXT: v_perm_b32 v14, v14, v31, s4
; GFX9-NEXT: v_perm_b32 v15, v29, v15, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_maxnum_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX10-NEXT: v_lshlrev_b32_e32 v39, 16, v27
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX10-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX10-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX10-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX10-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX10-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX10-NEXT: v_lshlrev_b32_e32 v37, 16, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v12
; GFX10-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX10-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX10-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX10-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX10-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX10-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX10-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX10-NEXT: v_lshlrev_b32_e32 v54, 16, v8
; GFX10-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX10-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX10-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX10-NEXT: v_lshlrev_b32_e32 v64, 16, v7
; GFX10-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX10-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX10-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX10-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX10-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX10-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX10-NEXT: v_max_f32_e32 v39, v48, v39
; GFX10-NEXT: v_max_f32_e32 v11, v11, v27
; GFX10-NEXT: v_max_f32_e32 v49, v50, v49
; GFX10-NEXT: v_max_f32_e32 v10, v10, v26
; GFX10-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX10-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX10-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX10-NEXT: v_max_f32_e32 v37, v38, v37
; GFX10-NEXT: v_lshlrev_b32_e32 v38, 16, v18
; GFX10-NEXT: v_max_f32_e32 v12, v12, v28
; GFX10-NEXT: v_lshlrev_b32_e32 v28, 16, v2
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v48, 16, v17
; GFX10-NEXT: v_lshlrev_b32_e32 v27, 16, v1
; GFX10-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v50, 16, v16
; GFX10-NEXT: v_lshlrev_b32_e32 v26, 16, v0
; GFX10-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_max_f32_e32 v9, v9, v25
; GFX10-NEXT: v_max_f32_e32 v25, v54, v53
; GFX10-NEXT: v_max_f32_e32 v8, v8, v24
; GFX10-NEXT: v_max_f32_e32 v24, v64, v55
; GFX10-NEXT: v_max_f32_e32 v7, v7, v23
; GFX10-NEXT: v_max_f32_e32 v23, v66, v65
; GFX10-NEXT: v_max_f32_e32 v6, v6, v22
; GFX10-NEXT: v_max_f32_e32 v22, v68, v67
; GFX10-NEXT: v_bfe_u32 v53, v39, 16, 1
; GFX10-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX10-NEXT: v_bfe_u32 v65, v49, 16, 1
; GFX10-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX10-NEXT: v_lshlrev_b32_e32 v33, 16, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v14
; GFX10-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX10-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX10-NEXT: v_max_f32_e32 v35, v36, v35
; GFX10-NEXT: v_lshlrev_b32_e32 v36, 16, v19
; GFX10-NEXT: v_max_f32_e32 v13, v13, v29
; GFX10-NEXT: v_lshlrev_b32_e32 v29, 16, v3
; GFX10-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_max_f32_e32 v2, v2, v18
; GFX10-NEXT: v_max_f32_e32 v18, v27, v48
; GFX10-NEXT: v_max_f32_e32 v1, v1, v17
; GFX10-NEXT: v_max_f32_e32 v17, v26, v50
; GFX10-NEXT: v_max_f32_e32 v0, v0, v16
; GFX10-NEXT: v_or_b32_e32 v54, 0x400000, v39
; GFX10-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX10-NEXT: v_or_b32_e32 v66, 0x400000, v49
; GFX10-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v39, v39
; GFX10-NEXT: v_add3_u32 v39, v53, v39, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v11, v11
; GFX10-NEXT: v_add3_u32 v11, v55, v11, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v49, v49
; GFX10-NEXT: v_add3_u32 v49, v65, v49, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v10, v10
; GFX10-NEXT: v_add3_u32 v10, v67, v10, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_max_f32_e32 v33, v34, v33
; GFX10-NEXT: v_lshlrev_b32_e32 v34, 16, v20
; GFX10-NEXT: v_max_f32_e32 v14, v14, v30
; GFX10-NEXT: v_lshlrev_b32_e32 v30, 16, v4
; GFX10-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_max_f32_e32 v3, v3, v19
; GFX10-NEXT: v_max_f32_e32 v19, v28, v38
; GFX10-NEXT: v_bfe_u32 v38, v37, 16, 1
; GFX10-NEXT: v_bfe_u32 v50, v12, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v39, v39, v54, s9
; GFX10-NEXT: v_bfe_u32 v54, v18, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v11, v11, v64, s10
; GFX10-NEXT: v_bfe_u32 v64, v1, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v49, v49, v66, s11
; GFX10-NEXT: v_bfe_u32 v66, v17, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v10, v10, v68, s12
; GFX10-NEXT: v_bfe_u32 v68, v0, 16, 1
; GFX10-NEXT: v_max_f32_e32 v51, v52, v51
; GFX10-NEXT: v_max_f32_e32 v5, v5, v21
; GFX10-NEXT: v_max_f32_e32 v21, v30, v34
; GFX10-NEXT: v_max_f32_e32 v4, v4, v20
; GFX10-NEXT: v_max_f32_e32 v20, v29, v36
; GFX10-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX10-NEXT: v_bfe_u32 v27, v14, 16, 1
; GFX10-NEXT: v_bfe_u32 v29, v35, 16, 1
; GFX10-NEXT: v_bfe_u32 v34, v13, 16, 1
; GFX10-NEXT: v_or_b32_e32 v48, 0x400000, v37
; GFX10-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v37, v37
; GFX10-NEXT: v_add3_u32 v37, v38, v37, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v12, v12
; GFX10-NEXT: v_add3_u32 v12, v50, v12, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s10, v18, v18
; GFX10-NEXT: v_add3_u32 v54, v54, v18, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v18, 0x400000, v18
; GFX10-NEXT: v_cmp_u_f32_e64 s11, v1, v1
; GFX10-NEXT: v_add3_u32 v64, v64, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e64 s12, v17, v17
; GFX10-NEXT: v_add3_u32 v66, v66, v17, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v17, 0x400000, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s22, v0, v0
; GFX10-NEXT: v_add3_u32 v68, v68, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v0, 0x400000, v0
; GFX10-NEXT: v_or_b32_e32 v26, 0x400000, v33
; GFX10-NEXT: v_or_b32_e32 v28, 0x400000, v14
; GFX10-NEXT: v_or_b32_e32 v30, 0x400000, v35
; GFX10-NEXT: v_or_b32_e32 v36, 0x400000, v13
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX10-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX10-NEXT: v_bfe_u32 v33, v51, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v14, v14
; GFX10-NEXT: v_add3_u32 v14, v27, v14, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v35, v35
; GFX10-NEXT: v_add3_u32 v29, v29, v35, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v13, v13
; GFX10-NEXT: v_add3_u32 v13, v34, v13, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v24, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v37, v37, v48, s7
; GFX10-NEXT: v_bfe_u32 v48, v19, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v12, v12, v52, s8
; GFX10-NEXT: v_bfe_u32 v52, v2, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v18, v54, v18, s10
; GFX10-NEXT: v_cndmask_b32_e64 v17, v66, v17, s12
; GFX10-NEXT: v_cndmask_b32_e64 v0, v68, v0, s22
; GFX10-NEXT: v_cndmask_b32_e64 v1, v64, v1, s11
; GFX10-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX10-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX10-NEXT: v_or_b32_e32 v27, 0x400000, v51
; GFX10-NEXT: v_bfe_u32 v35, v9, 16, 1
; GFX10-NEXT: v_bfe_u32 v38, v25, 16, 1
; GFX10-NEXT: v_or_b32_e32 v67, 0x400000, v24
; GFX10-NEXT: v_cmp_u_f32_e64 s13, v51, v51
; GFX10-NEXT: v_add3_u32 v33, v33, v51, 0x7fff
; GFX10-NEXT: v_bfe_u32 v51, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s17, v24, v24
; GFX10-NEXT: v_add3_u32 v24, v65, v24, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v6, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v16, v16, v26, vcc_lo
; GFX10-NEXT: v_bfe_u32 v26, v21, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v14, v14, v28, s4
; GFX10-NEXT: v_bfe_u32 v28, v4, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v29, v29, v30, s5
; GFX10-NEXT: v_bfe_u32 v30, v20, 16, 1
; GFX10-NEXT: v_cndmask_b32_e64 v13, v13, v36, s6
; GFX10-NEXT: v_bfe_u32 v36, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s8, v19, v19
; GFX10-NEXT: v_add3_u32 v48, v48, v19, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v19, 0x400000, v19
; GFX10-NEXT: v_cmp_u_f32_e64 s9, v2, v2
; GFX10-NEXT: v_add3_u32 v52, v52, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v2
; GFX10-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX10-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX10-NEXT: v_or_b32_e32 v34, 0x400000, v9
; GFX10-NEXT: v_or_b32_e32 v50, 0x400000, v25
; GFX10-NEXT: v_bfe_u32 v53, v8, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s14, v9, v9
; GFX10-NEXT: v_add3_u32 v9, v35, v9, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v35, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e64 s15, v25, v25
; GFX10-NEXT: v_add3_u32 v25, v38, v25, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v23, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s18, v7, v7
; GFX10-NEXT: v_add3_u32 v7, v51, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v51, 0x400000, v6
; GFX10-NEXT: v_cmp_u_f32_e64 s20, v6, v6
; GFX10-NEXT: v_add3_u32 v6, v65, v6, 0x7fff
; GFX10-NEXT: v_bfe_u32 v65, v5, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v21, v21
; GFX10-NEXT: v_add3_u32 v26, v26, v21, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v21
; GFX10-NEXT: v_cmp_u_f32_e64 s5, v4, v4
; GFX10-NEXT: v_add3_u32 v28, v28, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v4
; GFX10-NEXT: v_cmp_u_f32_e64 s6, v20, v20
; GFX10-NEXT: v_add3_u32 v30, v30, v20, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v20
; GFX10-NEXT: v_cmp_u_f32_e64 s7, v3, v3
; GFX10-NEXT: v_add3_u32 v36, v36, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e64 v19, v48, v19, s8
; GFX10-NEXT: v_cndmask_b32_e64 v2, v52, v2, s9
; GFX10-NEXT: v_or_b32_e32 v55, 0x400000, v8
; GFX10-NEXT: v_cmp_u_f32_e64 s16, v8, v8
; GFX10-NEXT: v_add3_u32 v8, v53, v8, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v53, 0x400000, v23
; GFX10-NEXT: v_cmp_u_f32_e64 s19, v23, v23
; GFX10-NEXT: v_add3_u32 v23, v38, v23, 0x7fff
; GFX10-NEXT: v_bfe_u32 v38, v22, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_add3_u32 v65, v65, v5, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e64 v21, v26, v21, s4
; GFX10-NEXT: v_cndmask_b32_e64 v4, v28, v4, s5
; GFX10-NEXT: v_cndmask_b32_e64 v20, v30, v20, s6
; GFX10-NEXT: v_cndmask_b32_e64 v3, v36, v3, s7
; GFX10-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX10-NEXT: v_cmp_u_f32_e64 s21, v22, v22
; GFX10-NEXT: v_add3_u32 v38, v38, v22, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v22, 0x400000, v22
; GFX10-NEXT: v_cndmask_b32_e32 v5, v65, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX10-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e64 v27, v33, v27, s13
; GFX10-NEXT: v_cndmask_b32_e64 v9, v9, v34, s14
; GFX10-NEXT: v_cndmask_b32_e64 v25, v25, v50, s15
; GFX10-NEXT: v_cndmask_b32_e64 v8, v8, v55, s16
; GFX10-NEXT: v_cndmask_b32_e64 v24, v24, v67, s17
; GFX10-NEXT: v_cndmask_b32_e64 v7, v7, v35, s18
; GFX10-NEXT: v_cndmask_b32_e64 v23, v23, v53, s19
; GFX10-NEXT: v_cndmask_b32_e64 v6, v6, v51, s20
; GFX10-NEXT: v_cndmask_b32_e64 v22, v38, v22, s21
; GFX10-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX10-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX10-NEXT: v_perm_b32 v9, v9, v27, 0x7060302
; GFX10-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX10-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX10-NEXT: v_perm_b32 v10, v10, v49, 0x7060302
; GFX10-NEXT: v_perm_b32 v11, v11, v39, 0x7060302
; GFX10-NEXT: v_perm_b32 v12, v12, v37, 0x7060302
; GFX10-NEXT: v_perm_b32 v13, v13, v29, 0x7060302
; GFX10-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX10-NEXT: v_and_b32_e32 v18, 0xffff0000, v32
; GFX10-NEXT: v_max_f32_e32 v17, v31, v17
; GFX10-NEXT: v_max_f32_e32 v15, v15, v18
; GFX10-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX10-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX10-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX10-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX10-NEXT: v_cmp_u_f32_e64 s4, v15, v15
; GFX10-NEXT: v_add3_u32 v17, v18, v17, 0x7fff
; GFX10-NEXT: v_add3_u32 v15, v19, v15, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v15, v15, v21, s4
; GFX10-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_maxnum_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: scratch_load_b32 v32, off, s32
; GFX11-NEXT: v_lshlrev_b32_e32 v67, 16, v21
; GFX11-NEXT: v_lshlrev_b32_e32 v68, 16, v5
; GFX11-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v83, 16, v17
; GFX11-NEXT: v_lshlrev_b32_e32 v84, 16, v1
; GFX11-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v49, 16, v26
; GFX11-NEXT: v_dual_max_f32 v5, v5, v21 :: v_dual_and_b32 v26, 0xffff0000, v26
; GFX11-NEXT: v_lshlrev_b32_e32 v53, 16, v24
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_max_f32 v1, v1, v17 :: v_dual_and_b32 v24, 0xffff0000, v24
; GFX11-NEXT: v_lshlrev_b32_e32 v71, 16, v19
; GFX11-NEXT: v_bfe_u32 v103, v5, 16, 1
; GFX11-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v81, 16, v18
; GFX11-NEXT: v_bfe_u32 v135, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v112, 0x400000, v5
; GFX11-NEXT: v_or_b32_e32 v144, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v103, v103, v5, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v80, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_add3_u32 v135, v135, v1, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v82, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v52, 16, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_max_f32 v3, v3, v19 :: v_dual_lshlrev_b32 v54, 16, v8
; GFX11-NEXT: v_lshlrev_b32_e32 v85, 16, v16
; GFX11-NEXT: v_dual_max_f32 v19, v82, v81 :: v_dual_lshlrev_b32 v64, 16, v7
; GFX11-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-NEXT: v_lshlrev_b32_e32 v65, 16, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v66, 16, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v129, v19, 16, 1
; GFX11-NEXT: v_or_b32_e32 v130, 0x400000, v19
; GFX11-NEXT: v_lshlrev_b32_e32 v48, 16, v11
; GFX11-NEXT: v_bfe_u32 v119, v3, 16, 1
; GFX11-NEXT: v_lshlrev_b32_e32 v51, 16, v25
; GFX11-NEXT: v_add3_u32 v129, v129, v19, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v86, 16, v0
; GFX11-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_dual_max_f32 v17, v86, v85 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: v_dual_max_f32 v8, v8, v24 :: v_dual_lshlrev_b32 v39, 16, v27
; GFX11-NEXT: v_or_b32_e32 v128, 0x400000, v3
; GFX11-NEXT: v_add3_u32 v119, v119, v3, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v145, v17, 16, 1
; GFX11-NEXT: v_or_b32_e32 v146, 0x400000, v17
; GFX11-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-NEXT: v_lshlrev_b32_e32 v70, 16, v4
; GFX11-NEXT: v_add3_u32 v145, v145, v17, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-NEXT: v_lshlrev_b32_e32 v55, 16, v23
; GFX11-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-NEXT: v_lshlrev_b32_e32 v50, 16, v10
; GFX11-NEXT: v_max_f32_e32 v2, v2, v18
; GFX11-NEXT: v_max_f32_e32 v0, v0, v16
; GFX11-NEXT: v_dual_max_f32 v24, v64, v55 :: v_dual_lshlrev_b32 v37, 16, v28
; GFX11-NEXT: v_max_f32_e32 v7, v7, v23
; GFX11-NEXT: v_dual_max_f32 v23, v66, v65 :: v_dual_max_f32 v18, v84, v83
; GFX11-NEXT: v_dual_max_f32 v9, v9, v25 :: v_dual_and_b32 v28, 0xffff0000, v28
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v85, v24, 16, 1
; GFX11-NEXT: v_bfe_u32 v97, v23, 16, 1
; GFX11-NEXT: v_or_b32_e32 v86, 0x400000, v24
; GFX11-NEXT: v_or_b32_e32 v98, 0x400000, v23
; GFX11-NEXT: v_bfe_u32 v87, v7, 16, 1
; GFX11-NEXT: v_add3_u32 v85, v85, v24, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v69, 16, v20
; GFX11-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-NEXT: v_add3_u32 v97, v97, v23, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-NEXT: v_or_b32_e32 v96, 0x400000, v7
; GFX11-NEXT: v_add3_u32 v87, v87, v7, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-NEXT: v_max_f32_e32 v4, v4, v20
; GFX11-NEXT: v_max_f32_e32 v20, v80, v71
; GFX11-NEXT: v_bfe_u32 v71, v9, 16, 1
; GFX11-NEXT: v_or_b32_e32 v80, 0x400000, v9
; GFX11-NEXT: v_lshlrev_b32_e32 v35, 16, v29
; GFX11-NEXT: v_dual_max_f32 v21, v70, v69 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v71, v71, v9, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-NEXT: v_dual_max_f32 v10, v10, v26 :: v_dual_and_b32 v29, 0xffff0000, v29
; GFX11-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-NEXT: v_max_f32_e32 v26, v52, v51
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_max_f32_e32 v6, v6, v22
; GFX11-NEXT: v_lshlrev_b32_e32 v36, 16, v13
; GFX11-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-NEXT: v_dual_max_f32 v11, v11, v27 :: v_dual_lshlrev_b32 v34, 16, v14
; GFX11-NEXT: v_dual_max_f32 v22, v68, v67 :: v_dual_lshlrev_b32 v33, 16, v30
; GFX11-NEXT: v_dual_max_f32 v27, v50, v49 :: v_dual_lshlrev_b32 v38, 16, v12
; GFX11-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-NEXT: v_dual_max_f32 v25, v54, v53 :: v_dual_and_b32 v12, 0xffff0000, v12
; GFX11-NEXT: v_dual_max_f32 v13, v13, v29 :: v_dual_and_b32 v30, 0xffff0000, v30
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_max_f32_e32 v29, v38, v37
; GFX11-NEXT: v_lshlrev_b32_e32 v31, 16, v15
; GFX11-NEXT: v_dual_max_f32 v12, v12, v28 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_max_f32_e32 v14, v14, v30
; GFX11-NEXT: v_max_f32_e32 v28, v48, v39
; GFX11-NEXT: v_dual_max_f32 v30, v36, v35 :: v_dual_max_f32 v33, v34, v33
; GFX11-NEXT: v_bfe_u32 v39, v13, 16, 1
; GFX11-NEXT: v_bfe_u32 v35, v14, 16, 1
; GFX11-NEXT: v_or_b32_e32 v36, 0x400000, v14
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v37, v30, 16, 1
; GFX11-NEXT: v_bfe_u32 v16, v33, 16, 1
; GFX11-NEXT: v_or_b32_e32 v34, 0x400000, v33
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-NEXT: v_add3_u32 v35, v35, v14, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v38, 0x400000, v30
; GFX11-NEXT: v_add3_u32 v16, v16, v33, 0x7fff
; GFX11-NEXT: v_add3_u32 v37, v37, v30, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v48, 0x400000, v13
; GFX11-NEXT: v_bfe_u32 v49, v29, 16, 1
; GFX11-NEXT: v_add3_u32 v39, v39, v13, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v16, v16, v34, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-NEXT: v_or_b32_e32 v50, 0x400000, v29
; GFX11-NEXT: v_bfe_u32 v51, v12, 16, 1
; GFX11-NEXT: v_add3_u32 v49, v49, v29, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v52, 0x400000, v12
; GFX11-NEXT: v_cndmask_b32_e32 v14, v35, v36, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-NEXT: v_bfe_u32 v53, v28, 16, 1
; GFX11-NEXT: v_add3_u32 v51, v51, v12, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v54, 0x400000, v28
; GFX11-NEXT: v_bfe_u32 v55, v11, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v30, v37, v38, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-NEXT: v_add3_u32 v53, v53, v28, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v64, 0x400000, v11
; GFX11-NEXT: v_bfe_u32 v65, v27, 16, 1
; GFX11-NEXT: v_add3_u32 v55, v55, v11, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v13, v39, v48, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-NEXT: v_or_b32_e32 v66, 0x400000, v27
; GFX11-NEXT: v_bfe_u32 v67, v10, 16, 1
; GFX11-NEXT: v_add3_u32 v65, v65, v27, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v68, 0x400000, v10
; GFX11-NEXT: v_cndmask_b32_e32 v29, v49, v50, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-NEXT: v_bfe_u32 v69, v26, 16, 1
; GFX11-NEXT: v_add3_u32 v67, v67, v10, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v70, 0x400000, v26
; GFX11-NEXT: v_bfe_u32 v81, v25, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v12, v51, v52, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-NEXT: v_add3_u32 v69, v69, v26, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v82, 0x400000, v25
; GFX11-NEXT: v_bfe_u32 v83, v8, 16, 1
; GFX11-NEXT: v_add3_u32 v81, v81, v25, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v28, v53, v54, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-NEXT: v_or_b32_e32 v84, 0x400000, v8
; GFX11-NEXT: v_add3_u32 v83, v83, v8, 0x7fff
; GFX11-NEXT: v_bfe_u32 v99, v6, 16, 1
; GFX11-NEXT: v_or_b32_e32 v100, 0x400000, v6
; GFX11-NEXT: v_cndmask_b32_e32 v11, v55, v64, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-NEXT: v_bfe_u32 v101, v22, 16, 1
; GFX11-NEXT: v_add3_u32 v99, v99, v6, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v102, 0x400000, v22
; GFX11-NEXT: v_bfe_u32 v113, v21, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v27, v65, v66, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-NEXT: v_add3_u32 v101, v101, v22, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v114, 0x400000, v21
; GFX11-NEXT: v_bfe_u32 v115, v4, 16, 1
; GFX11-NEXT: v_add3_u32 v113, v113, v21, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v10, v67, v68, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-NEXT: v_or_b32_e32 v116, 0x400000, v4
; GFX11-NEXT: v_bfe_u32 v117, v20, 16, 1
; GFX11-NEXT: v_add3_u32 v115, v115, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v118, 0x400000, v20
; GFX11-NEXT: v_cndmask_b32_e32 v26, v69, v70, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-NEXT: v_add3_u32 v117, v117, v20, 0x7fff
; GFX11-NEXT: v_bfe_u32 v133, v18, 16, 1
; GFX11-NEXT: v_or_b32_e32 v134, 0x400000, v18
; GFX11-NEXT: v_bfe_u32 v147, v0, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v9, v71, v80, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-NEXT: v_add3_u32 v133, v133, v18, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v33, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v147, v147, v0, 0x7fff
; GFX11-NEXT: v_bfe_u32 v131, v2, 16, 1
; GFX11-NEXT: v_cndmask_b32_e32 v25, v81, v82, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-NEXT: v_or_b32_e32 v132, 0x400000, v2
; GFX11-NEXT: v_perm_b32 v9, v9, v26, 0x7060302
; GFX11-NEXT: v_add3_u32 v131, v131, v2, 0x7fff
; GFX11-NEXT: v_perm_b32 v10, v10, v27, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v8, v83, v84, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-NEXT: v_perm_b32 v11, v11, v28, 0x7060302
; GFX11-NEXT: v_perm_b32 v12, v12, v29, 0x7060302
; GFX11-NEXT: v_perm_b32 v13, v13, v30, 0x7060302
; GFX11-NEXT: v_perm_b32 v8, v8, v25, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v24, v85, v86, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_perm_b32 v14, v14, v16, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v7, v87, v96, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v7, v7, v24, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v23, v97, v98, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_cndmask_b32_e32 v6, v99, v100, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-NEXT: v_perm_b32 v6, v6, v23, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v22, v101, v102, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_cndmask_b32_e32 v5, v103, v112, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v5, v5, v22, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v21, v113, v114, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_cndmask_b32_e32 v4, v115, v116, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-NEXT: v_perm_b32 v4, v4, v21, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v20, v117, v118, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-NEXT: v_cndmask_b32_e32 v19, v129, v130, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-NEXT: v_cndmask_b32_e32 v18, v133, v134, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v135, v144, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v1, v1, v18, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v17, v145, v146, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v147, v33, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_perm_b32 v0, v0, v17, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v2, v131, v132, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v2, v2, v19, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v119, v128, vcc_lo
; GFX11-NEXT: v_perm_b32 v3, v3, v20, 0x7060302
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v17, 16, v32
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_max_f32 v17, v31, v17 :: v_dual_and_b32 v18, 0xffff0000, v32
; GFX11-NEXT: v_max_f32_e32 v15, v15, v18
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v18, v17, 16, 1
; GFX11-NEXT: v_bfe_u32 v19, v15, 16, 1
; GFX11-NEXT: v_or_b32_e32 v20, 0x400000, v17
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-NEXT: v_or_b32_e32 v21, 0x400000, v15
; GFX11-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
; GFX11-NEXT: v_add3_u32 v19, v19, v15, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v17, v18, v20, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-NEXT: v_cndmask_b32_e32 v15, v19, v21, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v15, v15, v17, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <32 x bfloat> @llvm.maxnum.v32bf16(<32 x bfloat> %a, <32 x bfloat> %b)
ret <32 x bfloat> %op
}
declare bfloat @llvm.sqrt.bf16(bfloat)
define bfloat @v_sqrt_bf16(bfloat %a) {
; GCN-LABEL: v_sqrt_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0xf800000
; GCN-NEXT: v_mov_b32_e32 v1, 0x260
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v2, 0x4f800000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc
; GCN-NEXT: v_sqrt_f32_e32 v2, v0
; GCN-NEXT: v_add_i32_e64 v3, s[4:5], -1, v2
; GCN-NEXT: v_add_i32_e64 v4, s[4:5], 1, v2
; GCN-NEXT: v_fma_f32 v5, -v3, v2, v0
; GCN-NEXT: v_fma_f32 v6, -v4, v2, v0
; GCN-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v5
; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[4:5]
; GCN-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v6
; GCN-NEXT: v_cndmask_b32_e64 v2, v2, v4, s[4:5]
; GCN-NEXT: v_mul_f32_e32 v3, 0x37800000, v2
; GCN-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GCN-NEXT: v_cmp_class_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sqrt_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0xf800000
; GFX7-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX7-NEXT: v_sqrt_f32_e32 v1, v0
; GFX7-NEXT: v_add_i32_e64 v2, s[4:5], -1, v1
; GFX7-NEXT: v_fma_f32 v3, -v2, v1, v0
; GFX7-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3
; GFX7-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5]
; GFX7-NEXT: v_add_i32_e64 v3, s[4:5], 1, v1
; GFX7-NEXT: v_fma_f32 v1, -v3, v1, v0
; GFX7-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[4:5]
; GFX7-NEXT: v_mul_f32_e32 v2, 0x37800000, v1
; GFX7-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX7-NEXT: v_mov_b32_e32 v2, 0x260
; GFX7-NEXT: v_cmp_class_f32_e32 vcc, v0, v2
; GFX7-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sqrt_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_mov_b32 s4, 0xf800000
; GFX8-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT: v_sqrt_f32_e32 v1, v0
; GFX8-NEXT: v_add_u32_e64 v2, s[4:5], -1, v1
; GFX8-NEXT: v_fma_f32 v3, -v2, v1, v0
; GFX8-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3
; GFX8-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5]
; GFX8-NEXT: v_add_u32_e64 v3, s[4:5], 1, v1
; GFX8-NEXT: v_fma_f32 v1, -v3, v1, v0
; GFX8-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[4:5]
; GFX8-NEXT: v_mul_f32_e32 v2, 0x37800000, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX8-NEXT: v_mov_b32_e32 v2, 0x260
; GFX8-NEXT: v_cmp_class_f32_e32 vcc, v0, v2
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sqrt_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_mov_b32 s4, 0xf800000
; GFX9-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX9-NEXT: v_sqrt_f32_e32 v1, v0
; GFX9-NEXT: v_add_u32_e32 v2, -1, v1
; GFX9-NEXT: v_fma_f32 v3, -v2, v1, v0
; GFX9-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3
; GFX9-NEXT: v_add_u32_e32 v3, 1, v1
; GFX9-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5]
; GFX9-NEXT: v_fma_f32 v1, -v3, v1, v0
; GFX9-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[4:5]
; GFX9-NEXT: v_mul_f32_e32 v2, 0x37800000, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX9-NEXT: v_mov_b32_e32 v2, 0x260
; GFX9-NEXT: v_cmp_class_f32_e32 vcc, v0, v2
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sqrt_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xf800000, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT: v_sqrt_f32_e32 v1, v0
; GFX10-NEXT: v_add_nc_u32_e32 v2, -1, v1
; GFX10-NEXT: v_add_nc_u32_e32 v3, 1, v1
; GFX10-NEXT: v_fma_f32 v4, -v2, v1, v0
; GFX10-NEXT: v_fma_f32 v5, -v3, v1, v0
; GFX10-NEXT: v_cmp_ge_f32_e64 s4, 0, v4
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v2, s4
; GFX10-NEXT: v_cmp_lt_f32_e64 s4, 0, v5
; GFX10-NEXT: v_cndmask_b32_e64 v1, v1, v3, s4
; GFX10-NEXT: v_mul_f32_e32 v2, 0x37800000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc_lo
; GFX10-NEXT: v_cmp_class_f32_e64 vcc_lo, v0, 0x260
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc_lo
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sqrt_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xf800000, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_sqrt_f32_e32 v1, v0
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_add_nc_u32_e32 v2, -1, v1
; GFX11-NEXT: v_add_nc_u32_e32 v3, 1, v1
; GFX11-NEXT: v_fma_f32 v4, -v2, v1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_fma_f32 v5, -v3, v1, v0
; GFX11-NEXT: v_cmp_ge_f32_e64 s0, 0, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e64 v1, v1, v2, s0
; GFX11-NEXT: v_cmp_lt_f32_e64 s0, 0, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e64 v1, v1, v3, s0
; GFX11-NEXT: v_mul_f32_e32 v2, 0x37800000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc_lo
; GFX11-NEXT: v_cmp_class_f32_e64 vcc_lo, v0, 0x260
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.sqrt.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.ldexp.bf16.i32(bfloat, i32)
define bfloat @v_ldexp_bf16_i32(bfloat %a, i32 %b) {
; GCN-LABEL: v_ldexp_bf16_i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_ldexp_bf16_i32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_ldexp_bf16_i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_ldexp_f32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_ldexp_bf16_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_ldexp_f32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_ldexp_bf16_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_ldexp_f32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_ldexp_bf16_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.ldexp.bf16.i32(bfloat %a, i32 %b)
ret bfloat %op
}
declare { bfloat, i16 } @llvm.frexp.bf16.i16(bfloat)
define { bfloat, i16 } @v_frexp_bf16_i16(bfloat %a) {
; GCN-LABEL: v_frexp_bf16_i16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x7f800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_frexp_mant_f32_e32 v1, v0
; GCN-NEXT: v_frexp_exp_i32_f32_e32 v2, v0
; GCN-NEXT: v_cmp_lt_f32_e64 vcc, |v0|, s4
; GCN-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_frexp_bf16_i16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_frexp_exp_i32_f32_e32 v1, v0
; GFX7-NEXT: v_frexp_mant_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_frexp_bf16_i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX8-NEXT: v_frexp_mant_f32_e32 v0, v1
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_frexp_bf16_i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX9-NEXT: v_frexp_mant_f32_e32 v0, v1
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_frexp_bf16_i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX10-NEXT: v_frexp_mant_f32_e32 v0, v1
; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_frexp_bf16_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_frexp_mant_f32_e32 v0, v1
; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc_lo
; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call { bfloat, i16 } @llvm.frexp.bf16.i16(bfloat %a)
ret { bfloat, i16 } %op
}
declare bfloat @llvm.log.bf16(bfloat)
declare bfloat @llvm.log2.bf16(bfloat)
declare bfloat @llvm.log10.bf16(bfloat)
define bfloat @v_log_bf16(bfloat %a) {
; GCN-LABEL: v_log_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x800000
; GCN-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GCN-NEXT: s_mov_b32 s5, 0x7f800000
; GCN-NEXT: v_mov_b32_e32 v2, 0x41b17218
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_log_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xfffff000, v0
; GCN-NEXT: v_sub_f32_e32 v3, v0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 0x3805fdf4, v1
; GCN-NEXT: v_mul_f32_e32 v1, 0x3f317000, v1
; GCN-NEXT: v_mul_f32_e32 v5, 0x3f317000, v3
; GCN-NEXT: v_mul_f32_e32 v3, 0x3805fdf4, v3
; GCN-NEXT: v_add_f32_e32 v3, v4, v3
; GCN-NEXT: v_add_f32_e32 v3, v5, v3
; GCN-NEXT: v_add_f32_e32 v1, v1, v3
; GCN-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s5
; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
; GCN-NEXT: v_sub_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_log_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0x800000
; GFX7-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_log_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x3f317217
; GFX7-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0
; GFX7-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX7-NEXT: s_mov_b32 s4, 0x3377d1cf
; GFX7-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX7-NEXT: s_mov_b32 s4, 0x7f800000
; GFX7-NEXT: v_add_f32_e32 v1, v1, v2
; GFX7-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX7-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX7-NEXT: v_mov_b32_e32 v1, 0x41b17218
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_log_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_mov_b32 s4, 0x800000
; GFX8-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_log_f32_e32 v0, v0
; GFX8-NEXT: s_mov_b32 s4, 0x7f800000
; GFX8-NEXT: v_and_b32_e32 v1, 0xfffff000, v0
; GFX8-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX8-NEXT: v_mul_f32_e32 v3, 0x3f317000, v2
; GFX8-NEXT: v_mul_f32_e32 v2, 0x3805fdf4, v2
; GFX8-NEXT: v_mul_f32_e32 v4, 0x3805fdf4, v1
; GFX8-NEXT: v_add_f32_e32 v2, v4, v2
; GFX8-NEXT: v_add_f32_e32 v2, v3, v2
; GFX8-NEXT: v_mul_f32_e32 v1, 0x3f317000, v1
; GFX8-NEXT: v_add_f32_e32 v1, v1, v2
; GFX8-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX8-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX8-NEXT: v_mov_b32_e32 v1, 0x41b17218
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_log_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_mov_b32 s4, 0x800000
; GFX9-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_log_f32_e32 v0, v0
; GFX9-NEXT: s_mov_b32 s4, 0x3f317217
; GFX9-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0
; GFX9-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX9-NEXT: s_mov_b32 s4, 0x3377d1cf
; GFX9-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX9-NEXT: s_mov_b32 s4, 0x7f800000
; GFX9-NEXT: v_add_f32_e32 v1, v1, v2
; GFX9-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX9-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX9-NEXT: v_mov_b32_e32 v1, 0x41b17218
; GFX9-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_log_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX10-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x4f800000, vcc_lo
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_log_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0
; GFX10-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1
; GFX10-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v2
; GFX10-NEXT: v_cndmask_b32_e64 v2, 0, 0x41b17218, vcc_lo
; GFX10-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_log_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX11-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x4f800000, vcc_lo
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_log_f32_e32 v0, v0
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_mul_f32_e32 v1, 0x3f317217, v0
; GFX11-NEXT: v_fma_f32 v2, 0x3f317217, v0, -v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fmamk_f32 v2, v0, 0x3377d1cf, v2
; GFX11-NEXT: v_add_f32_e32 v1, v1, v2
; GFX11-NEXT: v_cndmask_b32_e64 v2, 0, 0x41b17218, vcc_lo
; GFX11-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.log.bf16(bfloat %a)
ret bfloat %op
}
define bfloat @v_log2_bf16(bfloat %a) {
; GCN-LABEL: v_log2_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x800000
; GCN-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GCN-NEXT: v_mov_b32_e32 v2, 0x42000000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_log_f32_e32 v0, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
; GCN-NEXT: v_sub_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_log2_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0x800000
; GFX7-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_log_f32_e32 v0, v0
; GFX7-NEXT: v_mov_b32_e32 v1, 0x42000000
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_log2_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_mov_b32 s4, 0x800000
; GFX8-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_log_f32_e32 v0, v0
; GFX8-NEXT: v_mov_b32_e32 v1, 0x42000000
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_log2_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_mov_b32 s4, 0x800000
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_mov_b32_e32 v2, 0x4f800000
; GFX9-NEXT: v_cndmask_b32_e32 v2, 1.0, v2, vcc
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX9-NEXT: v_log_f32_e32 v0, v0
; GFX9-NEXT: v_mov_b32_e32 v1, 0x42000000
; GFX9-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_log2_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX10-NEXT: v_cndmask_b32_e64 v2, 1.0, 0x4f800000, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX10-NEXT: v_log_f32_e32 v0, v0
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_log2_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX11-NEXT: v_cndmask_b32_e64 v2, 1.0, 0x4f800000, vcc_lo
; GFX11-NEXT: v_cndmask_b32_e64 v1, 0, 0x42000000, vcc_lo
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_log_f32_e32 v0, v0
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.log2.bf16(bfloat %a)
ret bfloat %op
}
define bfloat @v_log10_bf16(bfloat %a) {
; GCN-LABEL: v_log10_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x800000
; GCN-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GCN-NEXT: s_mov_b32 s5, 0x7f800000
; GCN-NEXT: v_mov_b32_e32 v2, 0x411a209b
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_log_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xfffff000, v0
; GCN-NEXT: v_sub_f32_e32 v3, v0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 0x369a84fb, v1
; GCN-NEXT: v_mul_f32_e32 v1, 0x3e9a2000, v1
; GCN-NEXT: v_mul_f32_e32 v5, 0x3e9a2000, v3
; GCN-NEXT: v_mul_f32_e32 v3, 0x369a84fb, v3
; GCN-NEXT: v_add_f32_e32 v3, v4, v3
; GCN-NEXT: v_add_f32_e32 v3, v5, v3
; GCN-NEXT: v_add_f32_e32 v1, v1, v3
; GCN-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s5
; GCN-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v2, vcc
; GCN-NEXT: v_sub_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_log10_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0x800000
; GFX7-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_log_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x3e9a209a
; GFX7-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0
; GFX7-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX7-NEXT: s_mov_b32 s4, 0x3284fbcf
; GFX7-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX7-NEXT: s_mov_b32 s4, 0x7f800000
; GFX7-NEXT: v_add_f32_e32 v1, v1, v2
; GFX7-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX7-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX7-NEXT: v_mov_b32_e32 v1, 0x411a209b
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_log10_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_mov_b32 s4, 0x800000
; GFX8-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_log_f32_e32 v0, v0
; GFX8-NEXT: s_mov_b32 s4, 0x7f800000
; GFX8-NEXT: v_and_b32_e32 v1, 0xfffff000, v0
; GFX8-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX8-NEXT: v_mul_f32_e32 v3, 0x3e9a2000, v2
; GFX8-NEXT: v_mul_f32_e32 v2, 0x369a84fb, v2
; GFX8-NEXT: v_mul_f32_e32 v4, 0x369a84fb, v1
; GFX8-NEXT: v_add_f32_e32 v2, v4, v2
; GFX8-NEXT: v_add_f32_e32 v2, v3, v2
; GFX8-NEXT: v_mul_f32_e32 v1, 0x3e9a2000, v1
; GFX8-NEXT: v_add_f32_e32 v1, v1, v2
; GFX8-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX8-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX8-NEXT: v_mov_b32_e32 v1, 0x411a209b
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_log10_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_mov_b32 s4, 0x800000
; GFX9-NEXT: v_mov_b32_e32 v1, 0x4f800000
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_log_f32_e32 v0, v0
; GFX9-NEXT: s_mov_b32 s4, 0x3e9a209a
; GFX9-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0
; GFX9-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX9-NEXT: s_mov_b32 s4, 0x3284fbcf
; GFX9-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX9-NEXT: s_mov_b32 s4, 0x7f800000
; GFX9-NEXT: v_add_f32_e32 v1, v1, v2
; GFX9-NEXT: v_cmp_lt_f32_e64 s[4:5], |v0|, s4
; GFX9-NEXT: v_cndmask_b32_e64 v0, v0, v1, s[4:5]
; GFX9-NEXT: v_mov_b32_e32 v1, 0x411a209b
; GFX9-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX9-NEXT: v_sub_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_log10_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX10-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x4f800000, vcc_lo
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_log_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0
; GFX10-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1
; GFX10-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v2
; GFX10-NEXT: v_cndmask_b32_e64 v2, 0, 0x411a209b, vcc_lo
; GFX10-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_log10_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0x800000, v0
; GFX11-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x4f800000, vcc_lo
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_log_f32_e32 v0, v0
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_mul_f32_e32 v1, 0x3e9a209a, v0
; GFX11-NEXT: v_fma_f32 v2, 0x3e9a209a, v0, -v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fmamk_f32 v2, v0, 0x3284fbcf, v2
; GFX11-NEXT: v_add_f32_e32 v1, v1, v2
; GFX11-NEXT: v_cndmask_b32_e64 v2, 0, 0x411a209b, vcc_lo
; GFX11-NEXT: v_cmp_gt_f32_e64 vcc_lo, 0x7f800000, |v0|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-NEXT: v_sub_f32_e32 v0, v0, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.log10.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.exp.bf16(bfloat)
declare bfloat @llvm.exp2.bf16(bfloat)
declare bfloat @llvm.exp10.bf16(bfloat)
define bfloat @v_exp_bf16(bfloat %a) {
; GCN-LABEL: v_exp_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0xc2ce8ed0
; GCN-NEXT: s_mov_b32 s5, 0x42b17218
; GCN-NEXT: v_mov_b32_e32 v1, 0x7f800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v2, 0x3fb8a000, v0
; GCN-NEXT: v_sub_f32_e32 v3, v0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 0x39a3b295, v0
; GCN-NEXT: v_rndne_f32_e32 v5, v2
; GCN-NEXT: v_mul_f32_e32 v6, 0x39a3b295, v3
; GCN-NEXT: v_mul_f32_e32 v3, 0x3fb8a000, v3
; GCN-NEXT: v_sub_f32_e32 v2, v2, v5
; GCN-NEXT: v_add_f32_e32 v3, v3, v6
; GCN-NEXT: v_cvt_i32_f32_e32 v5, v5
; GCN-NEXT: v_add_f32_e32 v3, v4, v3
; GCN-NEXT: v_add_f32_e32 v2, v2, v3
; GCN-NEXT: v_exp_f32_e32 v2, v2
; GCN-NEXT: v_ldexp_f32_e32 v2, v2, v5
; GCN-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
; GCN-NEXT: v_cmp_nlt_f32_e32 vcc, s5, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_exp_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0x3fb8aa3b
; GFX7-NEXT: v_mul_f32_e32 v1, 0x3fb8aa3b, v0
; GFX7-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX7-NEXT: s_mov_b32 s4, 0x32a5705f
; GFX7-NEXT: v_rndne_f32_e32 v3, v1
; GFX7-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX7-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX7-NEXT: v_add_f32_e32 v1, v1, v2
; GFX7-NEXT: v_exp_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v2, v3
; GFX7-NEXT: s_mov_b32 s4, 0xc2ce8ed0
; GFX7-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX7-NEXT: s_mov_b32 s4, 0x42b17218
; GFX7-NEXT: v_ldexp_f32_e32 v1, v1, v2
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX7-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_exp_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v3, v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x3fb8a000, v0
; GFX8-NEXT: v_mul_f32_e32 v4, 0x39a3b295, v3
; GFX8-NEXT: v_mul_f32_e32 v3, 0x3fb8a000, v3
; GFX8-NEXT: v_rndne_f32_e32 v2, v1
; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
; GFX8-NEXT: v_mul_f32_e32 v4, 0x39a3b295, v0
; GFX8-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX8-NEXT: v_add_f32_e32 v3, v4, v3
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_exp_f32_e32 v1, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX8-NEXT: s_mov_b32 s4, 0xc2ce8ed0
; GFX8-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX8-NEXT: s_mov_b32 s4, 0x42b17218
; GFX8-NEXT: v_ldexp_f32 v1, v1, v2
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX8-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_exp_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x3fb8aa3b, v0
; GFX9-NEXT: s_mov_b32 s4, 0x3fb8aa3b
; GFX9-NEXT: v_rndne_f32_e32 v2, v1
; GFX9-NEXT: v_sub_f32_e32 v3, v1, v2
; GFX9-NEXT: v_fma_f32 v1, v0, s4, -v1
; GFX9-NEXT: s_mov_b32 s4, 0x32a5705f
; GFX9-NEXT: v_fma_f32 v1, v0, s4, v1
; GFX9-NEXT: v_add_f32_e32 v1, v3, v1
; GFX9-NEXT: v_exp_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX9-NEXT: s_mov_b32 s4, 0xc2ce8ed0
; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX9-NEXT: s_mov_b32 s4, 0x42b17218
; GFX9-NEXT: v_ldexp_f32 v1, v1, v2
; GFX9-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX9-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX9-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_exp_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x3fb8aa3b, v0
; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc2ce8ed0, v0
; GFX10-NEXT: v_rndne_f32_e32 v2, v1
; GFX10-NEXT: v_fma_f32 v3, 0x3fb8aa3b, v0, -v1
; GFX10-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX10-NEXT: v_fmamk_f32 v3, v0, 0x32a5705f, v3
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_exp_f32_e32 v1, v1
; GFX10-NEXT: v_ldexp_f32 v1, v1, v2
; GFX10-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo
; GFX10-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x42b17218, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_exp_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v1, 0x3fb8aa3b, v0
; GFX11-NEXT: v_rndne_f32_e32 v2, v1
; GFX11-NEXT: v_fma_f32 v3, 0x3fb8aa3b, v0, -v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX11-NEXT: v_fmamk_f32 v3, v0, 0x32a5705f, v3
; GFX11-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc2ce8ed0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v1, v1, v3
; GFX11-NEXT: v_exp_f32_e32 v1, v1
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_ldexp_f32 v1, v1, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo
; GFX11-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x42b17218, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.exp.bf16(bfloat %a)
ret bfloat %op
}
define bfloat @v_exp2_bf16(bfloat %a) {
; GCN-LABEL: v_exp2_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0xc2fc0000
; GCN-NEXT: v_mov_b32_e32 v1, 0x42800000
; GCN-NEXT: v_mov_b32_e32 v2, 0x1f800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GCN-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-NEXT: v_exp_f32_e32 v0, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, 1.0, v2, vcc
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_exp2_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0xc2fc0000
; GFX7-NEXT: v_mov_b32_e32 v1, 0x42800000
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_add_f32_e32 v0, v0, v1
; GFX7-NEXT: v_exp_f32_e32 v0, v0
; GFX7-NEXT: v_mov_b32_e32 v1, 0x1f800000
; GFX7-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_exp2_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_mov_b32 s4, 0xc2fc0000
; GFX8-NEXT: v_mov_b32_e32 v1, 0x42800000
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_exp_f32_e32 v0, v0
; GFX8-NEXT: v_mov_b32_e32 v1, 0x1f800000
; GFX8-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_exp2_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_mov_b32 s4, 0xc2fc0000
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_mov_b32_e32 v2, 0x42800000
; GFX9-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_exp_f32_e32 v0, v0
; GFX9-NEXT: v_mov_b32_e32 v1, 0x1f800000
; GFX9-NEXT: v_cndmask_b32_e32 v1, 1.0, v1, vcc
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_exp2_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0
; GFX10-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x1f800000, vcc_lo
; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
; GFX10-NEXT: v_exp_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_exp2_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, 0xc2fc0000, v0
; GFX11-NEXT: v_cndmask_b32_e64 v2, 0, 0x42800000, vcc_lo
; GFX11-NEXT: v_cndmask_b32_e64 v1, 1.0, 0x1f800000, vcc_lo
; GFX11-NEXT: v_add_f32_e32 v0, v0, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_exp_f32_e32 v0, v0
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.exp2.bf16(bfloat %a)
ret bfloat %op
}
define bfloat @v_exp10_bf16(bfloat %a) {
; GCN-LABEL: v_exp10_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0xc23369f4
; GCN-NEXT: s_mov_b32 s5, 0x421a209b
; GCN-NEXT: v_mov_b32_e32 v1, 0x7f800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v2, 0x40549000, v0
; GCN-NEXT: v_sub_f32_e32 v3, v0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 0x3a2784bc, v0
; GCN-NEXT: v_rndne_f32_e32 v5, v2
; GCN-NEXT: v_mul_f32_e32 v6, 0x3a2784bc, v3
; GCN-NEXT: v_mul_f32_e32 v3, 0x40549000, v3
; GCN-NEXT: v_sub_f32_e32 v2, v2, v5
; GCN-NEXT: v_add_f32_e32 v3, v3, v6
; GCN-NEXT: v_cvt_i32_f32_e32 v5, v5
; GCN-NEXT: v_add_f32_e32 v3, v4, v3
; GCN-NEXT: v_add_f32_e32 v2, v2, v3
; GCN-NEXT: v_exp_f32_e32 v2, v2
; GCN-NEXT: v_ldexp_f32_e32 v2, v2, v5
; GCN-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GCN-NEXT: v_cndmask_b32_e32 v2, 0, v2, vcc
; GCN-NEXT: v_cmp_nlt_f32_e32 vcc, s5, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_exp10_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_mov_b32 s4, 0x40549a78
; GFX7-NEXT: v_mul_f32_e32 v1, 0x40549a78, v0
; GFX7-NEXT: v_fma_f32 v2, v0, s4, -v1
; GFX7-NEXT: s_mov_b32 s4, 0x33979a37
; GFX7-NEXT: v_rndne_f32_e32 v3, v1
; GFX7-NEXT: v_fma_f32 v2, v0, s4, v2
; GFX7-NEXT: v_sub_f32_e32 v1, v1, v3
; GFX7-NEXT: v_add_f32_e32 v1, v1, v2
; GFX7-NEXT: v_exp_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v2, v3
; GFX7-NEXT: s_mov_b32 s4, 0xc23369f4
; GFX7-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX7-NEXT: s_mov_b32 s4, 0x421a209b
; GFX7-NEXT: v_ldexp_f32_e32 v1, v1, v2
; GFX7-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX7-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX7-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_exp10_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_sub_f32_e32 v3, v0, v0
; GFX8-NEXT: v_mul_f32_e32 v1, 0x40549000, v0
; GFX8-NEXT: v_mul_f32_e32 v4, 0x3a2784bc, v3
; GFX8-NEXT: v_mul_f32_e32 v3, 0x40549000, v3
; GFX8-NEXT: v_rndne_f32_e32 v2, v1
; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
; GFX8-NEXT: v_mul_f32_e32 v4, 0x3a2784bc, v0
; GFX8-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX8-NEXT: v_add_f32_e32 v3, v4, v3
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_exp_f32_e32 v1, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX8-NEXT: s_mov_b32 s4, 0xc23369f4
; GFX8-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX8-NEXT: s_mov_b32 s4, 0x421a209b
; GFX8-NEXT: v_ldexp_f32 v1, v1, v2
; GFX8-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX8-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX8-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_exp10_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v1, 0x40549a78, v0
; GFX9-NEXT: s_mov_b32 s4, 0x40549a78
; GFX9-NEXT: v_rndne_f32_e32 v2, v1
; GFX9-NEXT: v_sub_f32_e32 v3, v1, v2
; GFX9-NEXT: v_fma_f32 v1, v0, s4, -v1
; GFX9-NEXT: s_mov_b32 s4, 0x33979a37
; GFX9-NEXT: v_fma_f32 v1, v0, s4, v1
; GFX9-NEXT: v_add_f32_e32 v1, v3, v1
; GFX9-NEXT: v_exp_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX9-NEXT: s_mov_b32 s4, 0xc23369f4
; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, s4, v0
; GFX9-NEXT: s_mov_b32 s4, 0x421a209b
; GFX9-NEXT: v_ldexp_f32 v1, v1, v2
; GFX9-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc
; GFX9-NEXT: v_mov_b32_e32 v2, 0x7f800000
; GFX9-NEXT: v_cmp_nlt_f32_e32 vcc, s4, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_exp10_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v1, 0x40549a78, v0
; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc23369f4, v0
; GFX10-NEXT: v_rndne_f32_e32 v2, v1
; GFX10-NEXT: v_fma_f32 v3, 0x40549a78, v0, -v1
; GFX10-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX10-NEXT: v_fmamk_f32 v3, v0, 0x33979a37, v3
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_exp_f32_e32 v1, v1
; GFX10-NEXT: v_ldexp_f32 v1, v1, v2
; GFX10-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo
; GFX10-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x421a209b, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_exp10_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v1, 0x40549a78, v0
; GFX11-NEXT: v_rndne_f32_e32 v2, v1
; GFX11-NEXT: v_fma_f32 v3, 0x40549a78, v0, -v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_sub_f32_e32 v1, v1, v2
; GFX11-NEXT: v_fmamk_f32 v3, v0, 0x33979a37, v3
; GFX11-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11-NEXT: v_cmp_ngt_f32_e32 vcc_lo, 0xc23369f4, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_f32_e32 v1, v1, v3
; GFX11-NEXT: v_exp_f32_e32 v1, v1
; GFX11-NEXT: s_waitcnt_depctr 0xfff
; GFX11-NEXT: v_ldexp_f32 v1, v1, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v1, 0, v1, vcc_lo
; GFX11-NEXT: v_cmp_nlt_f32_e32 vcc_lo, 0x421a209b, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x7f800000, v1, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.exp10.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.ceil.bf16(bfloat)
define bfloat @v_ceil_bf16(bfloat %a) {
; GCN-LABEL: v_ceil_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_ceil_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_ceil_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_ceil_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_ceil_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_ceil_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_ceil_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_ceil_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_ceil_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_ceil_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_ceil_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_ceil_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.ceil.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.trunc.bf16(bfloat)
define bfloat @v_trunc_bf16(bfloat %a) {
; GCN-LABEL: v_trunc_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_trunc_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_trunc_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_trunc_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_trunc_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_trunc_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_trunc_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_trunc_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_trunc_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.trunc.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.rint.bf16(bfloat)
define bfloat @v_rint_bf16(bfloat %a) {
; GCN-LABEL: v_rint_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_rndne_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_rint_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_rndne_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_rint_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_rndne_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_rint_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_rndne_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_rint_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_rndne_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_rint_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_rndne_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.rint.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.nearbyint.bf16(bfloat)
define bfloat @v_nearbyint_bf16(bfloat %a) {
; GCN-LABEL: v_nearbyint_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_rndne_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_nearbyint_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_rndne_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_nearbyint_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_rndne_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_nearbyint_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_rndne_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_nearbyint_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_rndne_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_nearbyint_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_rndne_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.nearbyint.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.round.bf16(bfloat)
define bfloat @v_round_bf16(bfloat %a) {
; GCN-LABEL: v_round_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_trunc_f32_e32 v1, v0
; GCN-NEXT: v_sub_f32_e32 v2, v0, v1
; GCN-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
; GCN-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
; GCN-NEXT: s_brev_b32 s4, -2
; GCN-NEXT: v_bfi_b32 v0, s4, v2, v0
; GCN-NEXT: v_add_f32_e32 v0, v1, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_round_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v1, v0
; GFX7-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX7-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
; GFX7-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
; GFX7-NEXT: s_brev_b32 s4, -2
; GFX7-NEXT: v_bfi_b32 v0, s4, v2, v0
; GFX7-NEXT: v_add_f32_e32 v0, v1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_round_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v1, v0
; GFX8-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX8-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
; GFX8-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
; GFX8-NEXT: s_brev_b32 s4, -2
; GFX8-NEXT: v_bfi_b32 v0, s4, v2, v0
; GFX8-NEXT: v_add_f32_e32 v0, v1, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_round_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v1, v0
; GFX9-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX9-NEXT: v_cmp_ge_f32_e64 s[4:5], |v2|, 0.5
; GFX9-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s[4:5]
; GFX9-NEXT: s_brev_b32 s4, -2
; GFX9-NEXT: v_bfi_b32 v0, s4, v2, v0
; GFX9-NEXT: v_add_f32_e32 v0, v1, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_round_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_trunc_f32_e32 v1, v0
; GFX10-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX10-NEXT: v_cmp_ge_f32_e64 s4, |v2|, 0.5
; GFX10-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s4
; GFX10-NEXT: v_bfi_b32 v0, 0x7fffffff, v2, v0
; GFX10-NEXT: v_add_f32_e32 v0, v1, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_round_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_trunc_f32_e32 v1, v0
; GFX11-NEXT: v_sub_f32_e32 v2, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cmp_ge_f32_e64 s0, |v2|, 0.5
; GFX11-NEXT: v_cndmask_b32_e64 v2, 0, 1.0, s0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fffffff, v2, v0
; GFX11-NEXT: v_add_f32_e32 v0, v1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.round.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.roundeven.bf16(bfloat)
define bfloat @v_roundeven_bf16(bfloat %a) {
; GCN-LABEL: v_roundeven_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_rndne_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_roundeven_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_rndne_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_roundeven_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_rndne_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_roundeven_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_rndne_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_roundeven_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_rndne_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_roundeven_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_rndne_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.roundeven.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.floor.bf16(bfloat)
define bfloat @v_floor_bf16(bfloat %a) {
; GCN-LABEL: v_floor_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_floor_f32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_floor_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_floor_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_floor_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_floor_f32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_floor_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_floor_f32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_floor_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_floor_f32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_floor_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_floor_f32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.floor.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.canonicalize.bf16(bfloat)
define bfloat @v_canonicalize_bf16(bfloat %a) {
; GCN-LABEL: v_canonicalize_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_canonicalize_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_canonicalize_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_canonicalize_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_max_f32_e32 v0, v0, v0
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_canonicalize_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_max_f32_e32 v0, v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_canonicalize_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_max_f32_e32 v0, v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.canonicalize.bf16(bfloat %a)
ret bfloat %op
}
declare bfloat @llvm.arithmetic.fence.bf16(bfloat)
; FIXME: Promotion broken
; define bfloat @v_arithmetic_fence_bf16(bfloat %a) {
; %op = call bfloat @llvm.arithmetic.fence.bf16(bfloat %a)
; ret bfloat %op
; }
define i1 @v_fcmp_false_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_false_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mov_b32_e32 v0, 0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_false_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mov_b32_e32 v0, 0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_false_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, 0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_false_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_false_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_false_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v0, 0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp false bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_oeq_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_oeq_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_eq_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_oeq_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_eq_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_oeq_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_eq_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_oeq_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_oeq_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_oeq_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp oeq bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ogt_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ogt_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_gt_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ogt_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_gt_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ogt_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_gt_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ogt_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_gt_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ogt_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_gt_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ogt_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_gt_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ogt bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_oge_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_oge_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_ge_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_oge_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_ge_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_oge_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_ge_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_oge_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_ge_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_oge_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_ge_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_oge_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_ge_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp oge bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_olt_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_olt_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_olt_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_olt_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_olt_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_olt_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_lt_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_olt_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_lt_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp olt bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ole_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ole_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_le_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ole_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_le_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ole_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_le_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ole_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_le_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ole_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_le_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ole_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_le_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ole bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_one_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_one_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_lg_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_one_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_lg_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_one_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_lg_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_one_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_lg_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_one_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_lg_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_one_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_lg_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp one bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_uno_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_uno_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_u_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_uno_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_u_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_uno_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_uno_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_uno_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_uno_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp uno bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ueq_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ueq_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_nlg_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ueq_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_nlg_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ueq_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_nlg_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ueq_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_nlg_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ueq_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_nlg_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ueq_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_nlg_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ueq bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ugt_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ugt_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_nle_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ugt_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_nle_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ugt_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_nle_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ugt_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_nle_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ugt_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_nle_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ugt_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_nle_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ugt bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_uge_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_uge_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_uge_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_uge_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_uge_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_uge_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_nlt_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_uge_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_nlt_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp uge bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ult_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ult_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_nge_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ult_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_nge_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ult_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_nge_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ult_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_nge_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ult_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_nge_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ult_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_nge_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ult bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_ule_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_ule_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_ngt_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_ule_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_ngt_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_ule_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_ngt_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_ule_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_ule_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_ule_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp ule bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_une_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_une_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cmp_neq_f32_e32 vcc, v0, v1
; GCN-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_une_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cmp_neq_f32_e32 vcc, v0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_une_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cmp_neq_f32_e32 vcc, v0, v1
; GFX8-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_une_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cmp_neq_f32_e32 vcc, v0, v1
; GFX9-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_une_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cmp_neq_f32_e32 vcc_lo, v0, v1
; GFX10-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_une_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_neq_f32_e32 vcc_lo, v0, v1
; GFX11-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp une bfloat %a, %b
ret i1 %op
}
define i1 @v_fcmp_true_bf16(bfloat %a, bfloat %b) {
; GCN-LABEL: v_fcmp_true_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mov_b32_e32 v0, 1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fcmp_true_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mov_b32_e32 v0, 1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fcmp_true_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_mov_b32_e32 v0, 1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fcmp_true_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, 1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fcmp_true_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v0, 1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fcmp_true_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_mov_b32_e32 v0, 1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fcmp true bfloat %a, %b
ret i1 %op
}
declare bfloat @llvm.copysign.bf16(bfloat, bfloat)
define bfloat @v_copysign_bf16_bf16(bfloat %mag, bfloat %sign) {
; GCN-LABEL: v_copysign_bf16_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x80000000, v1
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x80000000, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_bf16_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_bf16_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_bf16_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_bf16_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define bfloat @v_copysign_bf16_s_bf16(bfloat %mag, bfloat inreg %sign) {
; GCN-LABEL: v_copysign_bf16_s_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_and_b32 s4, s6, 0x80000000
; GCN-NEXT: s_lshr_b32 s4, s4, 16
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, s4, v0
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_s_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_and_b32 s4, s6, 0x80000000
; GFX7-NEXT: s_lshr_b32 s4, s4, 16
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, s4, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_bf16_s_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v1, s6
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_bf16_s_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v1, s6
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_bf16_s_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, s6
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_bf16_s_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, s0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define bfloat @v_copysign_s_bf16_bf16(bfloat inreg %mag, bfloat %sign) {
; GCN-LABEL: v_copysign_s_bf16_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e64 v1, 1.0, s6
; GCN-NEXT: v_and_b32_e32 v0, 0x80000000, v0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_bfe_u32 v1, v1, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_s_bf16_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s6
; GFX7-NEXT: v_and_b32_e32 v0, 0x80000000, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_bfe_u32 v1, v1, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_s_bf16_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v1, s6
; GFX8-NEXT: v_bfi_b32 v0, s4, v1, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_s_bf16_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v1, s6
; GFX9-NEXT: v_bfi_b32 v0, s4, v1, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_s_bf16_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s6, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_s_bf16_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define bfloat @v_copysign_bf16_f32(bfloat %mag, float %sign.f32) {
; GCN-LABEL: v_copysign_bf16_f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x80000000, v1
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x80000000, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_bf16_f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_bf16_f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_bf16_f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_bf16_f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = fptrunc float %sign.f32 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define bfloat @v_copysign_bf16_f64(bfloat %mag, double %sign.f64) {
; GCN-LABEL: v_copysign_bf16_f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x80000000, v2
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x80000000, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_bf16_f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v2
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_bf16_f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v2
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_bf16_f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v2
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_bf16_f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = fptrunc double %sign.f64 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define bfloat @v_copysign_bf16_f16(bfloat %mag, half %sign.f16) {
; GCN-LABEL: v_copysign_bf16_f16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_cvt_f16_f32_e32 v1, v1
; GCN-NEXT: v_and_b32_e32 v1, 0x8000, v1
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_f16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_and_b32_e32 v1, 0x8000, v1
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_bf16_f16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_bf16_f16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_bf16_f16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_bf16_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = bitcast half %sign.f16 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
ret bfloat %op
}
define amdgpu_ps i32 @s_copysign_bf16_bf16(bfloat inreg %mag, bfloat inreg %sign) {
; GCN-LABEL: s_copysign_bf16_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: s_and_b32 s0, s1, 0x80000000
; GCN-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, s0, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GFX7-NEXT: s_and_b32 s0, s1, 0x80000000
; GFX7-NEXT: s_lshr_b32 s0, s0, 16
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, s0, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_movk_i32 s2, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_movk_i32 s2, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v0, s1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_mov_b32_e32 v0, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define amdgpu_ps i32 @s_copysign_bf16_f32(bfloat inreg %mag, float inreg %sign.f32) {
; GCN-LABEL: s_copysign_bf16_f32:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: s_and_b32 s0, s1, 0x80000000
; GCN-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, s0, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GFX7-NEXT: s_and_b32 s0, s1, 0x80000000
; GFX7-NEXT: s_lshr_b32 s0, s0, 16
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, s0, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_lshrrev_b32_e64 v0, 16, s1
; GFX8-NEXT: s_movk_i32 s1, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v1, s0
; GFX8-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_f32:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_lshrrev_b32_e64 v0, 16, s1
; GFX9-NEXT: s_movk_i32 s1, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v1, s0
; GFX9-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_f32:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_lshrrev_b32_e64 v0, 16, s1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_lshrrev_b32_e64 v0, 16, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = fptrunc float %sign.f32 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define amdgpu_ps i32 @s_copysign_bf16_f64(bfloat inreg %mag, double inreg %sign.f64) {
; GCN-LABEL: s_copysign_bf16_f64:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: s_and_b32 s0, s2, 0x80000000
; GCN-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, s0, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GFX7-NEXT: s_and_b32 s0, s2, 0x80000000
; GFX7-NEXT: s_lshr_b32 s0, s0, 16
; GFX7-NEXT: v_bfe_u32 v0, v0, 16, 15
; GFX7-NEXT: v_or_b32_e32 v0, s0, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_lshrrev_b32_e64 v0, 16, s2
; GFX8-NEXT: s_movk_i32 s1, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v1, s0
; GFX8-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_f64:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_lshrrev_b32_e64 v0, 16, s2
; GFX9-NEXT: s_movk_i32 s1, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v1, s0
; GFX9-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_f64:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_lshrrev_b32_e64 v0, 16, s2
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_lshrrev_b32_e64 v0, 16, s2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = fptrunc double %sign.f64 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define amdgpu_ps i32 @s_copysign_bf16_f16(bfloat inreg %mag, half inreg %sign.f16) {
; GCN-LABEL: s_copysign_bf16_f16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
; GCN-NEXT: v_cvt_f16_f32_e32 v1, s1
; GCN-NEXT: v_and_b32_e32 v1, 0x8000, v1
; GCN-NEXT: v_bfe_u32 v0, v0, 16, 15
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_f16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_cvt_f16_f32_e32 v0, s1
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s0
; GFX7-NEXT: v_bfe_u32 v1, v1, 16, 15
; GFX7-NEXT: v_and_b32_e32 v0, 0x8000, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_f16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_movk_i32 s2, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_f16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_movk_i32 s2, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_f16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v0, s1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_f16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_mov_b32_e32 v0, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = bitcast half %sign.f16 to bfloat
%op = call bfloat @llvm.copysign.bf16(bfloat %mag, bfloat %sign)
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
declare float @llvm.copysign.f32(float, float)
define float @v_copysign_f32_bf16(float %mag, bfloat %sign.bf16) {
; GCN-LABEL: v_copysign_f32_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_brev_b32 s4, -2
; GCN-NEXT: v_bfi_b32 v0, s4, v0, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_f32_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_brev_b32 s4, -2
; GFX7-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_f32_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: s_brev_b32 s4, -2
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_f32_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: s_brev_b32 s4, -2
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_f32_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fffffff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_f32_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fffffff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = fpext bfloat %sign.bf16 to float
%op = call float @llvm.copysign.f32(float %mag, float %sign)
ret float %op
}
define amdgpu_ps i32 @s_copysign_f32_bf16(float inreg %mag, bfloat inreg %sign.bf16) {
; GCN-LABEL: s_copysign_f32_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_brev_b32 s2, -2
; GCN-NEXT: v_mov_b32_e32 v0, s0
; GCN-NEXT: v_mov_b32_e32 v1, s1
; GCN-NEXT: v_bfi_b32 v0, s2, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_f32_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_brev_b32 s2, -2
; GFX7-NEXT: v_mov_b32_e32 v0, s0
; GFX7-NEXT: v_mov_b32_e32 v1, s1
; GFX7-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_f32_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_lshlrev_b32_e64 v0, 16, s1
; GFX8-NEXT: s_brev_b32 s1, -2
; GFX8-NEXT: v_mov_b32_e32 v1, s0
; GFX8-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_f32_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_lshlrev_b32_e64 v0, 16, s1
; GFX9-NEXT: s_brev_b32 s1, -2
; GFX9-NEXT: v_mov_b32_e32 v1, s0
; GFX9-NEXT: v_bfi_b32 v0, s1, v1, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_f32_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_lshlrev_b32_e64 v0, 16, s1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fffffff, s0, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_f32_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_lshlrev_b32_e64 v0, 16, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fffffff, s0, v0
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = fpext bfloat %sign.bf16 to float
%op = call float @llvm.copysign.f32(float %mag, float %sign)
%cast = bitcast float %op to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %cast)
ret i32 %readlane
}
declare half @llvm.copysign.f16(half, half)
define half @v_copysign_f16_bf16(half %mag, bfloat %sign.bf16) {
; GCN-LABEL: v_copysign_f16_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
; GCN-NEXT: s_brev_b32 s4, -2
; GCN-NEXT: v_bfi_b32 v0, s4, v0, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_f16_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
; GFX7-NEXT: s_brev_b32 s4, -2
; GFX7-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_f16_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_f16_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfi_b32 v0, s4, v0, v1
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_f16_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_f16_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = bitcast bfloat %sign.bf16 to half
%op = call half @llvm.copysign.f16(half %mag, half %sign)
ret half %op
}
define amdgpu_ps i32 @s_copysign_f16_bf16(half inreg %mag, bfloat inreg %sign.bf16) {
; GCN-LABEL: s_copysign_f16_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s1
; GCN-NEXT: v_cvt_f16_f32_e32 v1, s0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_cvt_f32_f16_e32 v1, v1
; GCN-NEXT: v_cvt_f32_f16_e32 v0, v0
; GCN-NEXT: s_brev_b32 s0, -2
; GCN-NEXT: v_bfi_b32 v0, s0, v1, v0
; GCN-NEXT: v_cvt_f16_f32_e32 v0, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_f16_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_cvt_f16_f32_e32 v0, s0
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_f16_e32 v0, v0
; GFX7-NEXT: s_brev_b32 s0, -2
; GFX7-NEXT: v_bfi_b32 v0, s0, v0, v1
; GFX7-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_f16_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_movk_i32 s2, 0x7fff
; GFX8-NEXT: v_mov_b32_e32 v0, s0
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_f16_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_movk_i32 s2, 0x7fff
; GFX9-NEXT: v_mov_b32_e32 v0, s0
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_bfi_b32 v0, s2, v0, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_f16_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v0, s1
; GFX10-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_f16_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_mov_b32_e32 v0, s1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fff, s0, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = bitcast bfloat %sign.bf16 to half
%op = call half @llvm.copysign.f16(half %mag, half %sign)
%cast = bitcast half %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
declare double @llvm.copysign.f64(double, double)
define double @v_copysign_f64_bf16(double %mag, bfloat %sign.bf16) {
; GCN-LABEL: v_copysign_f64_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_brev_b32 s4, -2
; GCN-NEXT: v_bfi_b32 v1, s4, v1, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_f64_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_brev_b32 s4, -2
; GFX7-NEXT: v_bfi_b32 v1, s4, v1, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_copysign_f64_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT: s_brev_b32 s4, -2
; GFX8-NEXT: v_bfi_b32 v1, s4, v1, v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_copysign_f64_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT: s_brev_b32 s4, -2
; GFX9-NEXT: v_bfi_b32 v1, s4, v1, v2
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_copysign_f64_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT: v_bfi_b32 v1, 0x7fffffff, v1, v2
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_copysign_f64_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v1, 0x7fffffff, v1, v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%sign = fpext bfloat %sign.bf16 to double
%op = call double @llvm.copysign.f64(double %mag, double %sign)
ret double %op
}
define amdgpu_ps <2 x i32> @s_copysign_f64_bf16(double inreg %mag, bfloat inreg %sign.bf16) {
; GCN-LABEL: s_copysign_f64_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_brev_b32 s3, -2
; GCN-NEXT: v_mov_b32_e32 v0, s1
; GCN-NEXT: v_mov_b32_e32 v1, s2
; GCN-NEXT: v_bfi_b32 v0, s3, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s1, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_f64_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_brev_b32 s3, -2
; GFX7-NEXT: v_mov_b32_e32 v0, s1
; GFX7-NEXT: v_mov_b32_e32 v1, s2
; GFX7-NEXT: v_bfi_b32 v0, s3, v0, v1
; GFX7-NEXT: v_readfirstlane_b32 s1, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_f64_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_lshlrev_b32_e64 v0, 16, s2
; GFX8-NEXT: s_brev_b32 s2, -2
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_bfi_b32 v0, s2, v1, v0
; GFX8-NEXT: v_readfirstlane_b32 s1, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_f64_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_lshlrev_b32_e64 v0, 16, s2
; GFX9-NEXT: s_brev_b32 s2, -2
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_bfi_b32 v0, s2, v1, v0
; GFX9-NEXT: v_readfirstlane_b32 s1, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_f64_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_lshlrev_b32_e64 v0, 16, s2
; GFX10-NEXT: v_bfi_b32 v0, 0x7fffffff, s1, v0
; GFX10-NEXT: v_readfirstlane_b32 s1, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_f64_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_lshlrev_b32_e64 v0, 16, s2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfi_b32 v0, 0x7fffffff, s1, v0
; GFX11-NEXT: v_readfirstlane_b32 s1, v0
; GFX11-NEXT: ; return to shader part epilog
%sign = fpext bfloat %sign.bf16 to double
%op = call double @llvm.copysign.f64(double %mag, double %sign)
%cast = bitcast double %op to <2 x i32>
%cast.0 = extractelement <2 x i32> %cast, i32 0
%cast.1 = extractelement <2 x i32> %cast, i32 1
%readlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %cast.0)
%readlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %cast.1)
%ins.0 = insertelement <2 x i32> poison, i32 %readlane0, i32 0
%ins.1 = insertelement <2 x i32> %ins.0, i32 %readlane1, i32 1
ret <2 x i32> %ins.1
}
define i16 @v_fptosi_bf16_to_i16(bfloat %x) {
; GCN-LABEL: v_fptosi_bf16_to_i16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_bf16_to_i16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_bf16_to_i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_bf16_to_i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_bf16_to_i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_bf16_to_i16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi bfloat %x to i16
ret i16 %op
}
define <2 x i16> @v_fptosi_v2bf16_to_v2i16(<2 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v2bf16_to_v2i16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_or_b32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_or_b32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cvt_i32_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
; GFX8-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fptosi_v2bf16_to_v2i16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <2 x bfloat> %x to <2 x i16>
ret <2 x i16> %op
}
define <3 x i16> @v_fptosi_v3bf16_to_v3i16(<3 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v3bf16_to_v3i16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v3, v2
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v3
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_alignbit_b32 v1, v3, v1, 16
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_cvt_i32_f32_e32 v3, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v3
; GFX7-NEXT: v_alignbit_b32 v1, v3, v1, 16
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX8-NEXT: v_cvt_i32_f32_sdwa v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX8-NEXT: v_or_b32_sdwa v0, v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX9-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v2, v0, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fptosi_v3bf16_to_v3i16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <3 x bfloat> %x to <3 x i16>
ret <3 x i16> %op
}
define <4 x i16> @v_fptosi_v4bf16_to_v4i16(<4 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v4bf16_to_v4i16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v3, v3
; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_or_b32_e32 v2, v2, v4
; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX7-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v3
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_or_b32_e32 v2, v2, v4
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX8-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX8-NEXT: v_cvt_i32_f32_sdwa v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
; GFX8-NEXT: v_cvt_i32_f32_sdwa v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v0, v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX9-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX9-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x5040100
; GFX10-NEXT: v_perm_b32 v1, v1, v2, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX11TRUE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v2, v0, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v1, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fptosi_v4bf16_to_v4i16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11FAKE16-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v3, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v1, v1, v2, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <4 x bfloat> %x to <4 x i16>
ret <4 x i16> %op
}
define i32 @v_fptosi_bf16_to_i32(bfloat %x) {
; GCN-LABEL: v_fptosi_bf16_to_i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_bf16_to_i32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_bf16_to_i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_bf16_to_i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_bf16_to_i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_bf16_to_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi bfloat %x to i32
ret i32 %op
}
define <2 x i32> @v_fptosi_v2bf16_to_v2i32(<2 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v2bf16_to_v2i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v2bf16_to_v2i32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v2bf16_to_v2i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v1, v0
; GFX8-NEXT: v_mov_b32_e32 v0, v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v2bf16_to_v2i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v1, v0
; GFX9-NEXT: v_mov_b32_e32 v0, v2
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v2bf16_to_v2i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v2
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v2bf16_to_v2i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v1
; GFX11-NEXT: v_cvt_i32_f32_e32 v1, v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <2 x bfloat> %x to <2 x i32>
ret <2 x i32> %op
}
define <3 x i32> @v_fptosi_v3bf16_to_v3i32(<3 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v3bf16_to_v3i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v3bf16_to_v3i32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v3bf16_to_v3i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v4, v2
; GFX8-NEXT: v_cvt_i32_f32_e32 v3, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v0
; GFX8-NEXT: v_mov_b32_e32 v0, v4
; GFX8-NEXT: v_mov_b32_e32 v1, v3
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v3bf16_to_v3i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v4, v2
; GFX9-NEXT: v_cvt_i32_f32_e32 v3, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v0
; GFX9-NEXT: v_mov_b32_e32 v0, v4
; GFX9-NEXT: v_mov_b32_e32 v1, v3
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v3bf16_to_v3i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v2
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v3
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v4
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v3bf16_to_v3i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v2
; GFX11-NEXT: v_cvt_i32_f32_e32 v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_cvt_i32_f32_e32 v2, v4
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <3 x bfloat> %x to <3 x i32>
ret <3 x i32> %op
}
define <4 x i32> @v_fptosi_v4bf16_to_v4i32(<4 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v4bf16_to_v4i32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_cvt_i32_f32_e32 v0, v0
; GCN-NEXT: v_cvt_i32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_i32_f32_e32 v2, v2
; GCN-NEXT: v_cvt_i32_f32_e32 v3, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v4bf16_to_v4i32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_cvt_i32_f32_e32 v0, v0
; GFX7-NEXT: v_cvt_i32_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_i32_f32_e32 v2, v2
; GFX7-NEXT: v_cvt_i32_f32_e32 v3, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v4bf16_to_v4i32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_i32_f32_e32 v5, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v4, v2
; GFX8-NEXT: v_cvt_i32_f32_e32 v2, v0
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
; GFX8-NEXT: v_cvt_i32_f32_e32 v3, v0
; GFX8-NEXT: v_mov_b32_e32 v0, v4
; GFX8-NEXT: v_mov_b32_e32 v1, v5
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v4bf16_to_v4i32:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_i32_f32_e32 v5, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v4, v2
; GFX9-NEXT: v_cvt_i32_f32_e32 v2, v0
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v1
; GFX9-NEXT: v_cvt_i32_f32_e32 v3, v0
; GFX9-NEXT: v_mov_b32_e32 v0, v4
; GFX9-NEXT: v_mov_b32_e32 v1, v5
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v4bf16_to_v4i32:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
; GFX10-NEXT: v_cvt_i32_f32_e32 v0, v2
; GFX10-NEXT: v_cvt_i32_f32_e32 v1, v3
; GFX10-NEXT: v_cvt_i32_f32_e32 v2, v4
; GFX10-NEXT: v_cvt_i32_f32_e32 v3, v5
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v4bf16_to_v4i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_i32_f32_e32 v0, v2
; GFX11-NEXT: v_cvt_i32_f32_e32 v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_i32_f32_e32 v2, v4
; GFX11-NEXT: v_cvt_i32_f32_e32 v3, v5
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <4 x bfloat> %x to <4 x i32>
ret <4 x i32> %op
}
define i64 @v_fptosi_bf16_to_i64(bfloat %x) {
; GCN-LABEL: v_fptosi_bf16_to_i64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x2f800000
; GCN-NEXT: s_mov_b32 s5, 0xcf800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_trunc_f32_e32 v0, v0
; GCN-NEXT: v_mul_f32_e64 v1, |v0|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v0
; GCN-NEXT: v_floor_f32_e32 v1, v1
; GCN-NEXT: v_fma_f32 v0, v1, s5, |v0|
; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1
; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0
; GCN-NEXT: v_xor_b32_e32 v1, v1, v2
; GCN-NEXT: v_xor_b32_e32 v0, v0, v2
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v2
; GCN-NEXT: v_subb_u32_e32 v1, vcc, v1, v2, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_bf16_to_i64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x2f800000
; GFX7-NEXT: v_mul_f32_e64 v1, |v0|, s4
; GFX7-NEXT: v_floor_f32_e32 v1, v1
; GFX7-NEXT: s_mov_b32 s4, 0xcf800000
; GFX7-NEXT: v_fma_f32 v2, v1, s4, |v0|
; GFX7-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX7-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GFX7-NEXT: v_xor_b32_e32 v0, v2, v3
; GFX7-NEXT: v_xor_b32_e32 v1, v1, v3
; GFX7-NEXT: v_sub_i32_e32 v0, vcc, v0, v3
; GFX7-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_bf16_to_i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v0, v0
; GFX8-NEXT: s_mov_b32 s4, 0x2f800000
; GFX8-NEXT: v_mul_f32_e64 v1, |v0|, s4
; GFX8-NEXT: v_floor_f32_e32 v1, v1
; GFX8-NEXT: s_mov_b32 s4, 0xcf800000
; GFX8-NEXT: v_fma_f32 v2, v1, s4, |v0|
; GFX8-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX8-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GFX8-NEXT: v_xor_b32_e32 v0, v2, v3
; GFX8-NEXT: v_xor_b32_e32 v1, v1, v3
; GFX8-NEXT: v_sub_u32_e32 v0, vcc, v0, v3
; GFX8-NEXT: v_subb_u32_e32 v1, vcc, v1, v3, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_bf16_to_i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v0, v0
; GFX9-NEXT: s_mov_b32 s4, 0x2f800000
; GFX9-NEXT: v_mul_f32_e64 v1, |v0|, s4
; GFX9-NEXT: v_floor_f32_e32 v1, v1
; GFX9-NEXT: s_mov_b32 s4, 0xcf800000
; GFX9-NEXT: v_cvt_u32_f32_e32 v2, v1
; GFX9-NEXT: v_fma_f32 v1, v1, s4, |v0|
; GFX9-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GFX9-NEXT: v_xor_b32_e32 v2, v2, v3
; GFX9-NEXT: v_xor_b32_e32 v0, v1, v3
; GFX9-NEXT: v_sub_co_u32_e32 v0, vcc, v0, v3
; GFX9-NEXT: v_subb_co_u32_e32 v1, vcc, v2, v3, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_bf16_to_i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_trunc_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v0|
; GFX10-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GFX10-NEXT: v_floor_f32_e32 v1, v1
; GFX10-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v0|
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v2
; GFX10-NEXT: v_xor_b32_e32 v1, v1, v3
; GFX10-NEXT: v_xor_b32_e32 v0, v0, v3
; GFX10-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v3
; GFX10-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v1, v3, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_bf16_to_i64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_trunc_f32_e32 v0, v0
; GFX11-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v0|
; GFX11-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_floor_f32_e32 v1, v1
; GFX11-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v0|
; GFX11-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_u32_f32_e32 v0, v2
; GFX11-NEXT: v_xor_b32_e32 v1, v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_xor_b32_e32 v0, v0, v3
; GFX11-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v1, v3, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi bfloat %x to i64
ret i64 %op
}
define <2 x i64> @v_fptosi_v2bf16_to_v2i64(<2 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v2bf16_to_v2i64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x2f800000
; GCN-NEXT: s_mov_b32 s5, 0xcf800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_trunc_f32_e32 v0, v0
; GCN-NEXT: v_trunc_f32_e32 v1, v1
; GCN-NEXT: v_mul_f32_e64 v2, |v0|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v3, 31, v0
; GCN-NEXT: v_mul_f32_e64 v4, |v1|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v1
; GCN-NEXT: v_floor_f32_e32 v2, v2
; GCN-NEXT: v_floor_f32_e32 v4, v4
; GCN-NEXT: v_fma_f32 v0, v2, s5, |v0|
; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2
; GCN-NEXT: v_fma_f32 v1, v4, s5, |v1|
; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4
; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0
; GCN-NEXT: v_xor_b32_e32 v2, v2, v3
; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1
; GCN-NEXT: v_xor_b32_e32 v4, v4, v5
; GCN-NEXT: v_xor_b32_e32 v0, v0, v3
; GCN-NEXT: v_xor_b32_e32 v6, v1, v5
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v3
; GCN-NEXT: v_subb_u32_e32 v1, vcc, v2, v3, vcc
; GCN-NEXT: v_sub_i32_e32 v2, vcc, v6, v5
; GCN-NEXT: v_subb_u32_e32 v3, vcc, v4, v5, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v2bf16_to_v2i64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x2f800000
; GFX7-NEXT: v_mul_f32_e64 v2, |v0|, s4
; GFX7-NEXT: v_floor_f32_e32 v2, v2
; GFX7-NEXT: s_mov_b32 s5, 0xcf800000
; GFX7-NEXT: v_fma_f32 v3, v2, s5, |v0|
; GFX7-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_xor_b32_e32 v0, v3, v4
; GFX7-NEXT: v_trunc_f32_e32 v3, v1
; GFX7-NEXT: v_mul_f32_e64 v1, |v3|, s4
; GFX7-NEXT: v_floor_f32_e32 v1, v1
; GFX7-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX7-NEXT: v_fma_f32 v5, v1, s5, |v3|
; GFX7-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX7-NEXT: v_cvt_u32_f32_e32 v6, v1
; GFX7-NEXT: v_xor_b32_e32 v2, v2, v4
; GFX7-NEXT: v_sub_i32_e32 v0, vcc, v0, v4
; GFX7-NEXT: v_ashrrev_i32_e32 v3, 31, v3
; GFX7-NEXT: v_subb_u32_e32 v1, vcc, v2, v4, vcc
; GFX7-NEXT: v_xor_b32_e32 v2, v5, v3
; GFX7-NEXT: v_xor_b32_e32 v4, v6, v3
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, v2, v3
; GFX7-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v2bf16_to_v2i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v1, v1
; GFX8-NEXT: s_mov_b32 s4, 0x2f800000
; GFX8-NEXT: v_mul_f32_e64 v2, |v1|, s4
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_floor_f32_e32 v2, v2
; GFX8-NEXT: s_mov_b32 s5, 0xcf800000
; GFX8-NEXT: v_trunc_f32_e32 v4, v0
; GFX8-NEXT: v_fma_f32 v3, v2, s5, |v1|
; GFX8-NEXT: v_mul_f32_e64 v0, |v4|, s4
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX8-NEXT: v_floor_f32_e32 v0, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX8-NEXT: v_fma_f32 v5, v0, s5, |v4|
; GFX8-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX8-NEXT: v_cvt_u32_f32_e32 v6, v0
; GFX8-NEXT: v_xor_b32_e32 v3, v3, v1
; GFX8-NEXT: v_xor_b32_e32 v2, v2, v1
; GFX8-NEXT: v_sub_u32_e32 v0, vcc, v3, v1
; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v4
; GFX8-NEXT: v_subb_u32_e32 v1, vcc, v2, v1, vcc
; GFX8-NEXT: v_xor_b32_e32 v2, v5, v3
; GFX8-NEXT: v_xor_b32_e32 v4, v6, v3
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, v2, v3
; GFX8-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v2bf16_to_v2i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v1, v1
; GFX9-NEXT: s_mov_b32 s4, 0x2f800000
; GFX9-NEXT: v_mul_f32_e64 v2, |v1|, s4
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_floor_f32_e32 v2, v2
; GFX9-NEXT: s_mov_b32 s5, 0xcf800000
; GFX9-NEXT: v_trunc_f32_e32 v4, v0
; GFX9-NEXT: v_fma_f32 v3, v2, s5, |v1|
; GFX9-NEXT: v_mul_f32_e64 v0, |v4|, s4
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX9-NEXT: v_floor_f32_e32 v0, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX9-NEXT: v_fma_f32 v5, v0, s5, |v4|
; GFX9-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_cvt_u32_f32_e32 v6, v0
; GFX9-NEXT: v_xor_b32_e32 v3, v3, v1
; GFX9-NEXT: v_xor_b32_e32 v2, v2, v1
; GFX9-NEXT: v_sub_co_u32_e32 v0, vcc, v3, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 31, v4
; GFX9-NEXT: v_subb_co_u32_e32 v1, vcc, v2, v1, vcc
; GFX9-NEXT: v_xor_b32_e32 v2, v5, v3
; GFX9-NEXT: v_xor_b32_e32 v4, v6, v3
; GFX9-NEXT: v_sub_co_u32_e32 v2, vcc, v2, v3
; GFX9-NEXT: v_subb_co_u32_e32 v3, vcc, v4, v3, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v2bf16_to_v2i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_trunc_f32_e32 v1, v1
; GFX10-NEXT: v_trunc_f32_e32 v0, v0
; GFX10-NEXT: v_mul_f32_e64 v2, 0x2f800000, |v1|
; GFX10-NEXT: v_mul_f32_e64 v3, 0x2f800000, |v0|
; GFX10-NEXT: v_ashrrev_i32_e32 v6, 31, v0
; GFX10-NEXT: v_floor_f32_e32 v2, v2
; GFX10-NEXT: v_floor_f32_e32 v3, v3
; GFX10-NEXT: v_fma_f32 v4, 0xcf800000, v2, |v1|
; GFX10-NEXT: v_fma_f32 v5, 0xcf800000, v3, |v0|
; GFX10-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v4
; GFX10-NEXT: v_cvt_u32_f32_e32 v4, v5
; GFX10-NEXT: v_xor_b32_e32 v2, v2, v1
; GFX10-NEXT: v_xor_b32_e32 v3, v3, v6
; GFX10-NEXT: v_xor_b32_e32 v0, v0, v1
; GFX10-NEXT: v_xor_b32_e32 v4, v4, v6
; GFX10-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v1
; GFX10-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v2, v1, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v2, vcc_lo, v4, v6
; GFX10-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v3, v6, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v2bf16_to_v2i64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_trunc_f32_e32 v1, v1
; GFX11-NEXT: v_trunc_f32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_mul_f32_e64 v2, 0x2f800000, |v1|
; GFX11-NEXT: v_mul_f32_e64 v3, 0x2f800000, |v0|
; GFX11-NEXT: v_ashrrev_i32_e32 v6, 31, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_floor_f32_e32 v2, v2
; GFX11-NEXT: v_floor_f32_e32 v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_fma_f32 v4, 0xcf800000, v2, |v1|
; GFX11-NEXT: v_fma_f32 v5, 0xcf800000, v3, |v0|
; GFX11-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX11-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX11-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX11-NEXT: v_cvt_u32_f32_e32 v0, v4
; GFX11-NEXT: v_cvt_u32_f32_e32 v4, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_xor_b32_e32 v2, v2, v1
; GFX11-NEXT: v_xor_b32_e32 v3, v3, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_xor_b32_e32 v0, v0, v1
; GFX11-NEXT: v_xor_b32_e32 v4, v4, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_sub_co_u32 v0, vcc_lo, v0, v1
; GFX11-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v2, v1, vcc_lo
; GFX11-NEXT: v_sub_co_u32 v2, vcc_lo, v4, v6
; GFX11-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v3, v6, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <2 x bfloat> %x to <2 x i64>
ret <2 x i64> %op
}
define <3 x i64> @v_fptosi_v3bf16_to_v3i64(<3 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v3bf16_to_v3i64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x2f800000
; GCN-NEXT: s_mov_b32 s5, 0xcf800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_trunc_f32_e32 v0, v0
; GCN-NEXT: v_trunc_f32_e32 v1, v1
; GCN-NEXT: v_trunc_f32_e32 v2, v2
; GCN-NEXT: v_mul_f32_e64 v3, |v0|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v4, 31, v0
; GCN-NEXT: v_mul_f32_e64 v5, |v1|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v6, 31, v1
; GCN-NEXT: v_mul_f32_e64 v7, |v2|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v8, 31, v2
; GCN-NEXT: v_floor_f32_e32 v3, v3
; GCN-NEXT: v_floor_f32_e32 v5, v5
; GCN-NEXT: v_floor_f32_e32 v7, v7
; GCN-NEXT: v_fma_f32 v0, v3, s5, |v0|
; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
; GCN-NEXT: v_fma_f32 v1, v5, s5, |v1|
; GCN-NEXT: v_cvt_u32_f32_e32 v5, v5
; GCN-NEXT: v_fma_f32 v2, v7, s5, |v2|
; GCN-NEXT: v_cvt_u32_f32_e32 v7, v7
; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0
; GCN-NEXT: v_xor_b32_e32 v3, v3, v4
; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1
; GCN-NEXT: v_xor_b32_e32 v5, v5, v6
; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2
; GCN-NEXT: v_xor_b32_e32 v7, v7, v8
; GCN-NEXT: v_xor_b32_e32 v0, v0, v4
; GCN-NEXT: v_xor_b32_e32 v9, v1, v6
; GCN-NEXT: v_xor_b32_e32 v10, v2, v8
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v4
; GCN-NEXT: v_subb_u32_e32 v1, vcc, v3, v4, vcc
; GCN-NEXT: v_sub_i32_e32 v2, vcc, v9, v6
; GCN-NEXT: v_subb_u32_e32 v3, vcc, v5, v6, vcc
; GCN-NEXT: v_sub_i32_e32 v4, vcc, v10, v8
; GCN-NEXT: v_subb_u32_e32 v5, vcc, v7, v8, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v3bf16_to_v3i64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x2f800000
; GFX7-NEXT: v_mul_f32_e64 v3, |v0|, s4
; GFX7-NEXT: v_floor_f32_e32 v3, v3
; GFX7-NEXT: s_mov_b32 s5, 0xcf800000
; GFX7-NEXT: v_fma_f32 v4, v3, s5, |v0|
; GFX7-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_xor_b32_e32 v0, v4, v5
; GFX7-NEXT: v_trunc_f32_e32 v4, v1
; GFX7-NEXT: v_mul_f32_e64 v1, |v4|, s4
; GFX7-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX7-NEXT: v_floor_f32_e32 v1, v1
; GFX7-NEXT: v_fma_f32 v6, v1, s5, |v4|
; GFX7-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_xor_b32_e32 v3, v3, v5
; GFX7-NEXT: v_sub_i32_e32 v0, vcc, v0, v5
; GFX7-NEXT: v_cvt_u32_f32_e32 v7, v1
; GFX7-NEXT: v_subb_u32_e32 v1, vcc, v3, v5, vcc
; GFX7-NEXT: v_ashrrev_i32_e32 v3, 31, v4
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_xor_b32_e32 v5, v6, v3
; GFX7-NEXT: v_trunc_f32_e32 v6, v2
; GFX7-NEXT: v_mul_f32_e64 v2, |v6|, s4
; GFX7-NEXT: v_floor_f32_e32 v2, v2
; GFX7-NEXT: v_xor_b32_e32 v4, v7, v3
; GFX7-NEXT: v_fma_f32 v7, v2, s5, |v6|
; GFX7-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX7-NEXT: v_cvt_u32_f32_e32 v8, v2
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, v5, v3
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v6
; GFX7-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX7-NEXT: v_xor_b32_e32 v4, v7, v5
; GFX7-NEXT: v_xor_b32_e32 v6, v8, v5
; GFX7-NEXT: v_sub_i32_e32 v4, vcc, v4, v5
; GFX7-NEXT: v_subb_u32_e32 v5, vcc, v6, v5, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v3bf16_to_v3i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v2, v2
; GFX8-NEXT: s_mov_b32 s4, 0x2f800000
; GFX8-NEXT: v_mul_f32_e64 v3, |v2|, s4
; GFX8-NEXT: v_floor_f32_e32 v3, v3
; GFX8-NEXT: s_mov_b32 s5, 0xcf800000
; GFX8-NEXT: v_fma_f32 v4, v3, s5, |v2|
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX8-NEXT: v_trunc_f32_e32 v5, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX8-NEXT: v_mul_f32_e64 v0, |v5|, s4
; GFX8-NEXT: v_floor_f32_e32 v0, v0
; GFX8-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX8-NEXT: v_fma_f32 v6, v0, s5, |v5|
; GFX8-NEXT: v_xor_b32_e32 v4, v4, v2
; GFX8-NEXT: v_cvt_u32_f32_e32 v7, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_xor_b32_e32 v3, v3, v2
; GFX8-NEXT: v_cvt_u32_f32_e32 v8, v0
; GFX8-NEXT: v_sub_u32_e32 v0, vcc, v4, v2
; GFX8-NEXT: v_trunc_f32_e32 v1, v1
; GFX8-NEXT: v_subb_u32_e32 v6, vcc, v3, v2, vcc
; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v5
; GFX8-NEXT: v_mul_f32_e64 v5, |v1|, s4
; GFX8-NEXT: v_floor_f32_e32 v5, v5
; GFX8-NEXT: v_xor_b32_e32 v2, v7, v3
; GFX8-NEXT: v_fma_f32 v7, v5, s5, |v1|
; GFX8-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX8-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX8-NEXT: v_xor_b32_e32 v4, v8, v3
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, v2, v3
; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX8-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX8-NEXT: v_xor_b32_e32 v4, v7, v1
; GFX8-NEXT: v_xor_b32_e32 v5, v5, v1
; GFX8-NEXT: v_sub_u32_e32 v4, vcc, v4, v1
; GFX8-NEXT: v_subb_u32_e32 v5, vcc, v5, v1, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, v6
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v3bf16_to_v3i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v2, v2
; GFX9-NEXT: s_mov_b32 s4, 0x2f800000
; GFX9-NEXT: v_mul_f32_e64 v3, |v2|, s4
; GFX9-NEXT: v_floor_f32_e32 v3, v3
; GFX9-NEXT: s_mov_b32 s5, 0xcf800000
; GFX9-NEXT: v_fma_f32 v4, v3, s5, |v2|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX9-NEXT: v_trunc_f32_e32 v5, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX9-NEXT: v_mul_f32_e64 v0, |v5|, s4
; GFX9-NEXT: v_floor_f32_e32 v0, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX9-NEXT: v_fma_f32 v6, v0, s5, |v5|
; GFX9-NEXT: v_xor_b32_e32 v4, v4, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v7, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_xor_b32_e32 v3, v3, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v8, v0
; GFX9-NEXT: v_sub_co_u32_e32 v0, vcc, v4, v2
; GFX9-NEXT: v_trunc_f32_e32 v1, v1
; GFX9-NEXT: v_subb_co_u32_e32 v6, vcc, v3, v2, vcc
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 31, v5
; GFX9-NEXT: v_mul_f32_e64 v5, |v1|, s4
; GFX9-NEXT: v_floor_f32_e32 v5, v5
; GFX9-NEXT: v_xor_b32_e32 v2, v7, v3
; GFX9-NEXT: v_fma_f32 v7, v5, s5, |v1|
; GFX9-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX9-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX9-NEXT: v_xor_b32_e32 v4, v8, v3
; GFX9-NEXT: v_sub_co_u32_e32 v2, vcc, v2, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_subb_co_u32_e32 v3, vcc, v4, v3, vcc
; GFX9-NEXT: v_xor_b32_e32 v4, v7, v1
; GFX9-NEXT: v_xor_b32_e32 v5, v5, v1
; GFX9-NEXT: v_sub_co_u32_e32 v4, vcc, v4, v1
; GFX9-NEXT: v_subb_co_u32_e32 v5, vcc, v5, v1, vcc
; GFX9-NEXT: v_mov_b32_e32 v1, v6
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v3bf16_to_v3i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_trunc_f32_e32 v2, v2
; GFX10-NEXT: v_trunc_f32_e32 v0, v0
; GFX10-NEXT: v_trunc_f32_e32 v1, v1
; GFX10-NEXT: v_mul_f32_e64 v3, 0x2f800000, |v2|
; GFX10-NEXT: v_mul_f32_e64 v4, 0x2f800000, |v0|
; GFX10-NEXT: v_mul_f32_e64 v6, 0x2f800000, |v1|
; GFX10-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; GFX10-NEXT: v_ashrrev_i32_e32 v7, 31, v0
; GFX10-NEXT: v_floor_f32_e32 v3, v3
; GFX10-NEXT: v_floor_f32_e32 v4, v4
; GFX10-NEXT: v_floor_f32_e32 v6, v6
; GFX10-NEXT: v_ashrrev_i32_e32 v8, 31, v1
; GFX10-NEXT: v_fma_f32 v2, 0xcf800000, v3, |v2|
; GFX10-NEXT: v_fma_f32 v0, 0xcf800000, v4, |v0|
; GFX10-NEXT: v_fma_f32 v1, 0xcf800000, v6, |v1|
; GFX10-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX10-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_xor_b32_e32 v3, v3, v5
; GFX10-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX10-NEXT: v_xor_b32_e32 v2, v2, v5
; GFX10-NEXT: v_xor_b32_e32 v9, v0, v7
; GFX10-NEXT: v_xor_b32_e32 v4, v4, v7
; GFX10-NEXT: v_xor_b32_e32 v10, v1, v8
; GFX10-NEXT: v_xor_b32_e32 v6, v6, v8
; GFX10-NEXT: v_sub_co_u32 v0, vcc_lo, v2, v5
; GFX10-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v3, v5, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v2, vcc_lo, v9, v7
; GFX10-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v4, v7, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v4, vcc_lo, v10, v8
; GFX10-NEXT: v_sub_co_ci_u32_e32 v5, vcc_lo, v6, v8, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v3bf16_to_v3i64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_trunc_f32_e32 v2, v2
; GFX11-NEXT: v_trunc_f32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_trunc_f32_e32 v1, v1
; GFX11-NEXT: v_mul_f32_e64 v3, 0x2f800000, |v2|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_mul_f32_e64 v4, 0x2f800000, |v0|
; GFX11-NEXT: v_mul_f32_e64 v6, 0x2f800000, |v1|
; GFX11-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; GFX11-NEXT: v_ashrrev_i32_e32 v7, 31, v0
; GFX11-NEXT: v_floor_f32_e32 v3, v3
; GFX11-NEXT: v_floor_f32_e32 v4, v4
; GFX11-NEXT: v_floor_f32_e32 v6, v6
; GFX11-NEXT: v_ashrrev_i32_e32 v8, 31, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_fma_f32 v2, 0xcf800000, v3, |v2|
; GFX11-NEXT: v_fma_f32 v0, 0xcf800000, v4, |v0|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_fma_f32 v1, 0xcf800000, v6, |v1|
; GFX11-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX11-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX11-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX11-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX11-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX11-NEXT: v_xor_b32_e32 v3, v3, v5
; GFX11-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX11-NEXT: v_xor_b32_e32 v2, v2, v5
; GFX11-NEXT: v_xor_b32_e32 v9, v0, v7
; GFX11-NEXT: v_xor_b32_e32 v4, v4, v7
; GFX11-NEXT: v_xor_b32_e32 v10, v1, v8
; GFX11-NEXT: v_xor_b32_e32 v6, v6, v8
; GFX11-NEXT: v_sub_co_u32 v0, vcc_lo, v2, v5
; GFX11-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v3, v5, vcc_lo
; GFX11-NEXT: v_sub_co_u32 v2, vcc_lo, v9, v7
; GFX11-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v4, v7, vcc_lo
; GFX11-NEXT: v_sub_co_u32 v4, vcc_lo, v10, v8
; GFX11-NEXT: v_sub_co_ci_u32_e32 v5, vcc_lo, v6, v8, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <3 x bfloat> %x to <3 x i64>
ret <3 x i64> %op
}
define <4 x i64> @v_fptosi_v4bf16_to_v4i64(<4 x bfloat> %x) {
; GCN-LABEL: v_fptosi_v4bf16_to_v4i64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: s_mov_b32 s4, 0x2f800000
; GCN-NEXT: s_mov_b32 s5, 0xcf800000
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_trunc_f32_e32 v0, v0
; GCN-NEXT: v_trunc_f32_e32 v1, v1
; GCN-NEXT: v_trunc_f32_e32 v2, v2
; GCN-NEXT: v_trunc_f32_e32 v3, v3
; GCN-NEXT: v_mul_f32_e64 v4, |v0|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v0
; GCN-NEXT: v_mul_f32_e64 v6, |v1|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v1
; GCN-NEXT: v_mul_f32_e64 v8, |v2|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v9, 31, v2
; GCN-NEXT: v_mul_f32_e64 v10, |v3|, s4
; GCN-NEXT: v_ashrrev_i32_e32 v11, 31, v3
; GCN-NEXT: v_floor_f32_e32 v4, v4
; GCN-NEXT: v_floor_f32_e32 v6, v6
; GCN-NEXT: v_floor_f32_e32 v8, v8
; GCN-NEXT: v_floor_f32_e32 v10, v10
; GCN-NEXT: v_fma_f32 v0, v4, s5, |v0|
; GCN-NEXT: v_cvt_u32_f32_e32 v4, v4
; GCN-NEXT: v_fma_f32 v1, v6, s5, |v1|
; GCN-NEXT: v_cvt_u32_f32_e32 v6, v6
; GCN-NEXT: v_fma_f32 v2, v8, s5, |v2|
; GCN-NEXT: v_cvt_u32_f32_e32 v8, v8
; GCN-NEXT: v_fma_f32 v3, v10, s5, |v3|
; GCN-NEXT: v_cvt_u32_f32_e32 v10, v10
; GCN-NEXT: v_cvt_u32_f32_e32 v0, v0
; GCN-NEXT: v_xor_b32_e32 v4, v4, v5
; GCN-NEXT: v_cvt_u32_f32_e32 v1, v1
; GCN-NEXT: v_xor_b32_e32 v6, v6, v7
; GCN-NEXT: v_cvt_u32_f32_e32 v2, v2
; GCN-NEXT: v_xor_b32_e32 v8, v8, v9
; GCN-NEXT: v_cvt_u32_f32_e32 v3, v3
; GCN-NEXT: v_xor_b32_e32 v10, v10, v11
; GCN-NEXT: v_xor_b32_e32 v0, v0, v5
; GCN-NEXT: v_xor_b32_e32 v12, v1, v7
; GCN-NEXT: v_xor_b32_e32 v13, v2, v9
; GCN-NEXT: v_xor_b32_e32 v14, v3, v11
; GCN-NEXT: v_sub_i32_e32 v0, vcc, v0, v5
; GCN-NEXT: v_subb_u32_e32 v1, vcc, v4, v5, vcc
; GCN-NEXT: v_sub_i32_e32 v2, vcc, v12, v7
; GCN-NEXT: v_subb_u32_e32 v3, vcc, v6, v7, vcc
; GCN-NEXT: v_sub_i32_e32 v4, vcc, v13, v9
; GCN-NEXT: v_subb_u32_e32 v5, vcc, v8, v9, vcc
; GCN-NEXT: v_sub_i32_e32 v6, vcc, v14, v11
; GCN-NEXT: v_subb_u32_e32 v7, vcc, v10, v11, vcc
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fptosi_v4bf16_to_v4i64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_trunc_f32_e32 v0, v0
; GFX7-NEXT: s_mov_b32 s4, 0x2f800000
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v3
; GFX7-NEXT: v_mul_f32_e64 v3, |v0|, s4
; GFX7-NEXT: v_floor_f32_e32 v3, v3
; GFX7-NEXT: s_mov_b32 s5, 0xcf800000
; GFX7-NEXT: v_fma_f32 v5, v3, s5, |v0|
; GFX7-NEXT: v_cvt_u32_f32_e32 v5, v5
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v6, 31, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_xor_b32_e32 v0, v5, v6
; GFX7-NEXT: v_trunc_f32_e32 v5, v1
; GFX7-NEXT: v_mul_f32_e64 v1, |v5|, s4
; GFX7-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX7-NEXT: v_floor_f32_e32 v1, v1
; GFX7-NEXT: v_fma_f32 v7, v1, s5, |v5|
; GFX7-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_xor_b32_e32 v3, v3, v6
; GFX7-NEXT: v_sub_i32_e32 v0, vcc, v0, v6
; GFX7-NEXT: v_cvt_u32_f32_e32 v8, v1
; GFX7-NEXT: v_subb_u32_e32 v1, vcc, v3, v6, vcc
; GFX7-NEXT: v_ashrrev_i32_e32 v3, 31, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_xor_b32_e32 v6, v7, v3
; GFX7-NEXT: v_trunc_f32_e32 v7, v2
; GFX7-NEXT: v_mul_f32_e64 v2, |v7|, s4
; GFX7-NEXT: v_floor_f32_e32 v2, v2
; GFX7-NEXT: v_xor_b32_e32 v5, v8, v3
; GFX7-NEXT: v_fma_f32 v8, v2, s5, |v7|
; GFX7-NEXT: v_cvt_u32_f32_e32 v8, v8
; GFX7-NEXT: v_cvt_u32_f32_e32 v9, v2
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, v6, v3
; GFX7-NEXT: v_subb_u32_e32 v3, vcc, v5, v3, vcc
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v7
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_xor_b32_e32 v7, v8, v5
; GFX7-NEXT: v_trunc_f32_e32 v8, v4
; GFX7-NEXT: v_mul_f32_e64 v4, |v8|, s4
; GFX7-NEXT: v_floor_f32_e32 v4, v4
; GFX7-NEXT: v_xor_b32_e32 v6, v9, v5
; GFX7-NEXT: v_fma_f32 v9, v4, s5, |v8|
; GFX7-NEXT: v_cvt_u32_f32_e32 v9, v9
; GFX7-NEXT: v_cvt_u32_f32_e32 v10, v4
; GFX7-NEXT: v_sub_i32_e32 v4, vcc, v7, v5
; GFX7-NEXT: v_ashrrev_i32_e32 v7, 31, v8
; GFX7-NEXT: v_subb_u32_e32 v5, vcc, v6, v5, vcc
; GFX7-NEXT: v_xor_b32_e32 v6, v9, v7
; GFX7-NEXT: v_xor_b32_e32 v8, v10, v7
; GFX7-NEXT: v_sub_i32_e32 v6, vcc, v6, v7
; GFX7-NEXT: v_subb_u32_e32 v7, vcc, v8, v7, vcc
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fptosi_v4bf16_to_v4i64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_trunc_f32_e32 v2, v2
; GFX8-NEXT: s_mov_b32 s4, 0x2f800000
; GFX8-NEXT: v_mul_f32_e64 v3, |v2|, s4
; GFX8-NEXT: v_floor_f32_e32 v3, v3
; GFX8-NEXT: s_mov_b32 s5, 0xcf800000
; GFX8-NEXT: v_fma_f32 v4, v3, s5, |v2|
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX8-NEXT: v_trunc_f32_e32 v5, v0
; GFX8-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX8-NEXT: v_mul_f32_e64 v0, |v5|, s4
; GFX8-NEXT: v_floor_f32_e32 v0, v0
; GFX8-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX8-NEXT: v_fma_f32 v6, v0, s5, |v5|
; GFX8-NEXT: v_xor_b32_e32 v4, v4, v2
; GFX8-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX8-NEXT: v_xor_b32_e32 v3, v3, v2
; GFX8-NEXT: v_cvt_u32_f32_e32 v7, v0
; GFX8-NEXT: v_sub_u32_e32 v0, vcc, v4, v2
; GFX8-NEXT: v_subb_u32_e32 v8, vcc, v3, v2, vcc
; GFX8-NEXT: v_ashrrev_i32_e32 v3, 31, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX8-NEXT: v_trunc_f32_e32 v5, v5
; GFX8-NEXT: v_xor_b32_e32 v2, v6, v3
; GFX8-NEXT: v_mul_f32_e64 v6, |v5|, s4
; GFX8-NEXT: v_floor_f32_e32 v6, v6
; GFX8-NEXT: v_xor_b32_e32 v4, v7, v3
; GFX8-NEXT: v_fma_f32 v7, v6, s5, |v5|
; GFX8-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, v2, v3
; GFX8-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX8-NEXT: v_trunc_f32_e32 v1, v1
; GFX8-NEXT: v_subb_u32_e32 v3, vcc, v4, v3, vcc
; GFX8-NEXT: v_xor_b32_e32 v4, v7, v5
; GFX8-NEXT: v_mul_f32_e64 v7, |v1|, s4
; GFX8-NEXT: v_floor_f32_e32 v7, v7
; GFX8-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX8-NEXT: v_fma_f32 v9, v7, s5, |v1|
; GFX8-NEXT: v_cvt_u32_f32_e32 v9, v9
; GFX8-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX8-NEXT: v_xor_b32_e32 v6, v6, v5
; GFX8-NEXT: v_sub_u32_e32 v4, vcc, v4, v5
; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX8-NEXT: v_subb_u32_e32 v5, vcc, v6, v5, vcc
; GFX8-NEXT: v_xor_b32_e32 v6, v9, v1
; GFX8-NEXT: v_xor_b32_e32 v7, v7, v1
; GFX8-NEXT: v_sub_u32_e32 v6, vcc, v6, v1
; GFX8-NEXT: v_subb_u32_e32 v7, vcc, v7, v1, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, v8
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fptosi_v4bf16_to_v4i64:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX9-NEXT: v_trunc_f32_e32 v2, v2
; GFX9-NEXT: s_mov_b32 s4, 0x2f800000
; GFX9-NEXT: v_mul_f32_e64 v3, |v2|, s4
; GFX9-NEXT: v_floor_f32_e32 v3, v3
; GFX9-NEXT: s_mov_b32 s5, 0xcf800000
; GFX9-NEXT: v_fma_f32 v4, v3, s5, |v2|
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v4, v4
; GFX9-NEXT: v_trunc_f32_e32 v5, v0
; GFX9-NEXT: v_cvt_u32_f32_e32 v3, v3
; GFX9-NEXT: v_mul_f32_e64 v0, |v5|, s4
; GFX9-NEXT: v_floor_f32_e32 v0, v0
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX9-NEXT: v_fma_f32 v6, v0, s5, |v5|
; GFX9-NEXT: v_xor_b32_e32 v4, v4, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX9-NEXT: v_xor_b32_e32 v3, v3, v2
; GFX9-NEXT: v_cvt_u32_f32_e32 v7, v0
; GFX9-NEXT: v_sub_co_u32_e32 v0, vcc, v4, v2
; GFX9-NEXT: v_subb_co_u32_e32 v8, vcc, v3, v2, vcc
; GFX9-NEXT: v_ashrrev_i32_e32 v3, 31, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v1
; GFX9-NEXT: v_trunc_f32_e32 v5, v5
; GFX9-NEXT: v_xor_b32_e32 v2, v6, v3
; GFX9-NEXT: v_mul_f32_e64 v6, |v5|, s4
; GFX9-NEXT: v_floor_f32_e32 v6, v6
; GFX9-NEXT: v_xor_b32_e32 v4, v7, v3
; GFX9-NEXT: v_fma_f32 v7, v6, s5, |v5|
; GFX9-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_sub_co_u32_e32 v2, vcc, v2, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX9-NEXT: v_trunc_f32_e32 v1, v1
; GFX9-NEXT: v_subb_co_u32_e32 v3, vcc, v4, v3, vcc
; GFX9-NEXT: v_xor_b32_e32 v4, v7, v5
; GFX9-NEXT: v_mul_f32_e64 v7, |v1|, s4
; GFX9-NEXT: v_floor_f32_e32 v7, v7
; GFX9-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX9-NEXT: v_fma_f32 v9, v7, s5, |v1|
; GFX9-NEXT: v_cvt_u32_f32_e32 v9, v9
; GFX9-NEXT: v_cvt_u32_f32_e32 v7, v7
; GFX9-NEXT: v_xor_b32_e32 v6, v6, v5
; GFX9-NEXT: v_sub_co_u32_e32 v4, vcc, v4, v5
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_subb_co_u32_e32 v5, vcc, v6, v5, vcc
; GFX9-NEXT: v_xor_b32_e32 v6, v9, v1
; GFX9-NEXT: v_xor_b32_e32 v7, v7, v1
; GFX9-NEXT: v_sub_co_u32_e32 v6, vcc, v6, v1
; GFX9-NEXT: v_subb_co_u32_e32 v7, vcc, v7, v1, vcc
; GFX9-NEXT: v_mov_b32_e32 v1, v8
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fptosi_v4bf16_to_v4i64:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_trunc_f32_e32 v2, v2
; GFX10-NEXT: v_trunc_f32_e32 v0, v0
; GFX10-NEXT: v_trunc_f32_e32 v3, v3
; GFX10-NEXT: v_trunc_f32_e32 v4, v1
; GFX10-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v2|
; GFX10-NEXT: v_mul_f32_e64 v6, 0x2f800000, |v0|
; GFX10-NEXT: v_mul_f32_e64 v8, 0x2f800000, |v3|
; GFX10-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; GFX10-NEXT: v_mul_f32_e64 v9, 0x2f800000, |v4|
; GFX10-NEXT: v_floor_f32_e32 v1, v1
; GFX10-NEXT: v_floor_f32_e32 v6, v6
; GFX10-NEXT: v_floor_f32_e32 v8, v8
; GFX10-NEXT: v_ashrrev_i32_e32 v7, 31, v0
; GFX10-NEXT: v_floor_f32_e32 v9, v9
; GFX10-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v2|
; GFX10-NEXT: v_fma_f32 v0, 0xcf800000, v6, |v0|
; GFX10-NEXT: v_ashrrev_i32_e32 v10, 31, v3
; GFX10-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX10-NEXT: v_fma_f32 v3, 0xcf800000, v8, |v3|
; GFX10-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX10-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX10-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX10-NEXT: v_fma_f32 v11, 0xcf800000, v9, |v4|
; GFX10-NEXT: v_xor_b32_e32 v1, v1, v5
; GFX10-NEXT: v_xor_b32_e32 v2, v2, v5
; GFX10-NEXT: v_cvt_u32_f32_e32 v12, v3
; GFX10-NEXT: v_xor_b32_e32 v3, v0, v7
; GFX10-NEXT: v_cvt_u32_f32_e32 v8, v8
; GFX10-NEXT: v_xor_b32_e32 v6, v6, v7
; GFX10-NEXT: v_cvt_u32_f32_e32 v11, v11
; GFX10-NEXT: v_sub_co_u32 v0, vcc_lo, v2, v5
; GFX10-NEXT: v_ashrrev_i32_e32 v13, 31, v4
; GFX10-NEXT: v_cvt_u32_f32_e32 v9, v9
; GFX10-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v1, v5, vcc_lo
; GFX10-NEXT: v_xor_b32_e32 v4, v12, v10
; GFX10-NEXT: v_sub_co_u32 v2, vcc_lo, v3, v7
; GFX10-NEXT: v_xor_b32_e32 v5, v8, v10
; GFX10-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v6, v7, vcc_lo
; GFX10-NEXT: v_xor_b32_e32 v6, v11, v13
; GFX10-NEXT: v_xor_b32_e32 v7, v9, v13
; GFX10-NEXT: v_sub_co_u32 v4, vcc_lo, v4, v10
; GFX10-NEXT: v_sub_co_ci_u32_e32 v5, vcc_lo, v5, v10, vcc_lo
; GFX10-NEXT: v_sub_co_u32 v6, vcc_lo, v6, v13
; GFX10-NEXT: v_sub_co_ci_u32_e32 v7, vcc_lo, v7, v13, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fptosi_v4bf16_to_v4i64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_trunc_f32_e32 v2, v2
; GFX11-NEXT: v_trunc_f32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_trunc_f32_e32 v3, v3
; GFX11-NEXT: v_trunc_f32_e32 v4, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e64 v1, 0x2f800000, |v2|
; GFX11-NEXT: v_mul_f32_e64 v6, 0x2f800000, |v0|
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_mul_f32_e64 v8, 0x2f800000, |v3|
; GFX11-NEXT: v_ashrrev_i32_e32 v5, 31, v2
; GFX11-NEXT: v_mul_f32_e64 v9, 0x2f800000, |v4|
; GFX11-NEXT: v_floor_f32_e32 v1, v1
; GFX11-NEXT: v_floor_f32_e32 v6, v6
; GFX11-NEXT: v_floor_f32_e32 v8, v8
; GFX11-NEXT: v_ashrrev_i32_e32 v7, 31, v0
; GFX11-NEXT: v_floor_f32_e32 v9, v9
; GFX11-NEXT: v_fma_f32 v2, 0xcf800000, v1, |v2|
; GFX11-NEXT: v_fma_f32 v0, 0xcf800000, v6, |v0|
; GFX11-NEXT: v_ashrrev_i32_e32 v10, 31, v3
; GFX11-NEXT: v_cvt_u32_f32_e32 v1, v1
; GFX11-NEXT: v_fma_f32 v3, 0xcf800000, v8, |v3|
; GFX11-NEXT: v_cvt_u32_f32_e32 v2, v2
; GFX11-NEXT: v_cvt_u32_f32_e32 v0, v0
; GFX11-NEXT: v_cvt_u32_f32_e32 v6, v6
; GFX11-NEXT: v_fma_f32 v11, 0xcf800000, v9, |v4|
; GFX11-NEXT: v_xor_b32_e32 v1, v1, v5
; GFX11-NEXT: v_xor_b32_e32 v2, v2, v5
; GFX11-NEXT: v_cvt_u32_f32_e32 v12, v3
; GFX11-NEXT: v_xor_b32_e32 v3, v0, v7
; GFX11-NEXT: v_cvt_u32_f32_e32 v8, v8
; GFX11-NEXT: v_xor_b32_e32 v6, v6, v7
; GFX11-NEXT: v_cvt_u32_f32_e32 v11, v11
; GFX11-NEXT: v_sub_co_u32 v0, vcc_lo, v2, v5
; GFX11-NEXT: v_ashrrev_i32_e32 v13, 31, v4
; GFX11-NEXT: v_cvt_u32_f32_e32 v9, v9
; GFX11-NEXT: v_sub_co_ci_u32_e32 v1, vcc_lo, v1, v5, vcc_lo
; GFX11-NEXT: v_xor_b32_e32 v4, v12, v10
; GFX11-NEXT: v_sub_co_u32 v2, vcc_lo, v3, v7
; GFX11-NEXT: v_xor_b32_e32 v5, v8, v10
; GFX11-NEXT: v_sub_co_ci_u32_e32 v3, vcc_lo, v6, v7, vcc_lo
; GFX11-NEXT: v_xor_b32_e32 v6, v11, v13
; GFX11-NEXT: v_xor_b32_e32 v7, v9, v13
; GFX11-NEXT: v_sub_co_u32 v4, vcc_lo, v4, v10
; GFX11-NEXT: v_sub_co_ci_u32_e32 v5, vcc_lo, v5, v10, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_sub_co_u32 v6, vcc_lo, v6, v13
; GFX11-NEXT: v_sub_co_ci_u32_e32 v7, vcc_lo, v7, v13, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = fptosi <4 x bfloat> %x to <4 x i64>
ret <4 x i64> %op
}
define bfloat @v_sitofp_i16_to_bf16(i16 %x) {
; GCN-LABEL: v_sitofp_i16_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 16
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_i16_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_i16_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v1, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_i16_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_i16_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_i16_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp i16 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_sitofp_v2i16_to_v2bf16(<2 x i16> %x) {
; GCN-LABEL: v_sitofp_v2i16_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 16
; GCN-NEXT: v_bfe_i32 v1, v1, 0, 16
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v2i16_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX7-NEXT: v_bfe_i32 v1, v1, 0, 16
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v2i16_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_sdwa v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v2i16_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_sdwa v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT: v_add3_u32 v2, v2, v1, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v2i16_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_sdwa v1, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v2i16_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfe_i32 v1, v0, 0, 16
; GFX11-NEXT: v_ashrrev_i32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <2 x i16> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_sitofp_v3i16_to_v3bf16(<3 x i16> %x) {
; GCN-LABEL: v_sitofp_v3i16_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 16
; GCN-NEXT: v_bfe_i32 v1, v1, 0, 16
; GCN-NEXT: v_bfe_i32 v2, v2, 0, 16
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX7-NEXT: v_bfe_i32 v1, v1, 0, 16
; GFX7-NEXT: v_bfe_i32 v2, v2, 0, 16
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_i32_sdwa v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX8-NEXT: v_bfe_u32 v3, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v4
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v2, vcc
; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_i32_sdwa v2, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_sdwa v2, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_bfe_i32 v2, v0, 0, 16
; GFX11TRUE16-NEXT: v_bfe_i32 v1, v1, 0, 16
; GFX11TRUE16-NEXT: v_ashrrev_i32_e32 v0, 16, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_sitofp_v3i16_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_bfe_i32 v2, v0, 0, 16
; GFX11FAKE16-NEXT: v_bfe_i32 v1, v1, 0, 16
; GFX11FAKE16-NEXT: v_ashrrev_i32_e32 v0, 16, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <3 x i16> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_sitofp_v4i16_to_v4bf16(<4 x i16> %x) {
; GCN-LABEL: v_sitofp_v4i16_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_bfe_i32 v0, v0, 0, 16
; GCN-NEXT: v_bfe_i32 v1, v1, 0, 16
; GCN-NEXT: v_bfe_i32 v2, v2, 0, 16
; GCN-NEXT: v_bfe_i32 v3, v3, 0, 16
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v3
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v4i16_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX7-NEXT: v_bfe_i32 v1, v1, 0, 16
; GFX7-NEXT: v_bfe_i32 v2, v2, 0, 16
; GFX7-NEXT: v_bfe_i32 v3, v3, 0, 16
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v4i16_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_sdwa v2, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_cvt_f32_i32_sdwa v5, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_bfe_u32 v4, v5, 16, 1
; GFX8-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v5
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v4i16_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_sdwa v2, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_cvt_f32_i32_sdwa v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_bfe_u32 v3, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v4, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v4i16_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_sdwa v2, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_i32_sdwa v3, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_i32_sdwa v0, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_cvt_f32_i32_sdwa v1, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v8, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v10, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v8, v8, v3, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v0
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v10, v10, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v4i16_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_bfe_i32 v2, v1, 0, 16
; GFX11-NEXT: v_ashrrev_i32_e32 v1, 16, v1
; GFX11-NEXT: v_bfe_i32 v3, v0, 0, 16
; GFX11-NEXT: v_ashrrev_i32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v8, v3, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX11-NEXT: v_bfe_u32 v10, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v6, v6, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v8, v8, v3, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v10, v10, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <4 x i16> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_sitofp_i32_to_bf16(i32 %x) {
; GCN-LABEL: v_sitofp_i32_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_i32_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_i32_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v1, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_i32_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_i32_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_i32_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp i32 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_sitofp_v2i32_to_v2bf16(<2 x i32> %x) {
; GCN-LABEL: v_sitofp_v2i32_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v2i32_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v2i32_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v2i32_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v1, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v2i32_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v2i32_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <2 x i32> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_sitofp_v3i32_to_v3bf16(<3 x i32> %x) {
; GCN-LABEL: v_sitofp_v3i32_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: v_mov_b32_e32 v1, v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v2, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v2, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v2, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_sitofp_v3i32_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v2, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <3 x i32> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_sitofp_v4i32_to_v4bf16(<4 x i32> %x) {
; GCN-LABEL: v_sitofp_v4i32_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v3
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v4i32_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v4i32_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX8-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX8-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: v_alignbit_b32 v1, v3, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v4i32_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX9-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v2, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_perm_b32 v1, v3, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v4i32_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v9, v9, v1, 0x7fff
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v4, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v4i32_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: v_cvt_f32_i32_e32 v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v9, v9, v1, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v4, vcc_lo
; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <4 x i32> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_sitofp_i64_to_bf16(i64 %x) {
; GCN-LABEL: v_sitofp_i64_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_xor_b32_e32 v2, v0, v1
; GCN-NEXT: v_ffbh_i32_e32 v3, v1
; GCN-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GCN-NEXT: v_add_i32_e32 v3, vcc, -1, v3
; GCN-NEXT: v_add_i32_e32 v2, vcc, 32, v2
; GCN-NEXT: v_min_u32_e32 v2, v3, v2
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_i64_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_xor_b32_e32 v2, v0, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX7-NEXT: v_ffbh_i32_e32 v3, v1
; GFX7-NEXT: v_add_i32_e32 v2, vcc, 32, v2
; GFX7-NEXT: v_add_i32_e32 v3, vcc, -1, v3
; GFX7-NEXT: v_min_u32_e32 v2, v3, v2
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_i64_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_xor_b32_e32 v2, v0, v1
; GFX8-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX8-NEXT: v_ffbh_i32_e32 v3, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 32, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, -1, v3
; GFX8-NEXT: v_min_u32_e32 v2, v3, v2
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
; GFX8-NEXT: v_ldexp_f32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_i64_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_xor_b32_e32 v2, v0, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX9-NEXT: v_ffbh_i32_e32 v3, v1
; GFX9-NEXT: v_add_u32_e32 v2, 32, v2
; GFX9-NEXT: v_add_u32_e32 v3, -1, v3
; GFX9-NEXT: v_min_u32_e32 v2, v3, v2
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v2
; GFX9-NEXT: v_ldexp_f32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_i64_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_xor_b32_e32 v2, v0, v1
; GFX10-NEXT: v_ffbh_i32_e32 v3, v1
; GFX10-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX10-NEXT: v_add_nc_u32_e32 v3, -1, v3
; GFX10-NEXT: v_add_nc_u32_e32 v2, 32, v2
; GFX10-NEXT: v_min_u32_e32 v2, v3, v2
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, 32, v2
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_ldexp_f32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_i64_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_xor_b32_e32 v2, v0, v1
; GFX11-NEXT: v_cls_i32_e32 v3, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_ashrrev_i32_e32 v2, 31, v2
; GFX11-NEXT: v_add_nc_u32_e32 v3, -1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add_nc_u32_e32 v2, 32, v2
; GFX11-NEXT: v_min_u32_e32 v2, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: v_sub_nc_u32_e32 v1, 32, v2
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp i64 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_sitofp_v2i64_to_v2bf16(<2 x i64> %x) {
; GCN-LABEL: v_sitofp_v2i64_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_i32_e32 v4, v3
; GCN-NEXT: v_xor_b32_e32 v5, v2, v3
; GCN-NEXT: v_ffbh_i32_e32 v6, v1
; GCN-NEXT: v_xor_b32_e32 v7, v0, v1
; GCN-NEXT: v_add_i32_e32 v4, vcc, -1, v4
; GCN-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GCN-NEXT: v_add_i32_e32 v6, vcc, -1, v6
; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GCN-NEXT: v_add_i32_e32 v5, vcc, 32, v5
; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v7
; GCN-NEXT: v_min_u32_e32 v4, v4, v5
; GCN-NEXT: v_min_u32_e32 v5, v6, v7
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v4
; GCN-NEXT: v_sub_i32_e32 v4, vcc, 32, v4
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v5
; GCN-NEXT: v_sub_i32_e32 v5, vcc, 32, v5
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v1, v1, v4
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v5
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v2i64_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_xor_b32_e32 v5, v2, v3
; GFX7-NEXT: v_ffbh_i32_e32 v4, v3
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX7-NEXT: v_add_i32_e32 v4, vcc, -1, v4
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 32, v5
; GFX7-NEXT: v_min_u32_e32 v4, v4, v5
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v4
; GFX7-NEXT: v_xor_b32_e32 v5, v0, v1
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_i32_e32 v3, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX7-NEXT: v_add_i32_e32 v3, vcc, -1, v3
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 32, v5
; GFX7-NEXT: v_min_u32_e32 v3, v3, v5
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v4, vcc, 32, v4
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v4
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v2i64_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_xor_b32_e32 v5, v0, v1
; GFX8-NEXT: v_ffbh_i32_e32 v4, v1
; GFX8-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX8-NEXT: v_add_u32_e32 v4, vcc, -1, v4
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 32, v5
; GFX8-NEXT: v_min_u32_e32 v4, v4, v5
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v4
; GFX8-NEXT: v_ldexp_f32 v4, v0, v1
; GFX8-NEXT: v_bfe_u32 v0, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v4
; GFX8-NEXT: v_xor_b32_e32 v1, v2, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v0
; GFX8-NEXT: v_ffbh_i32_e32 v0, v3
; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, -1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 32, v1
; GFX8-NEXT: v_min_u32_e32 v6, v0, v1
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v6, v[2:3]
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v4
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, 32, v6
; GFX8-NEXT: v_ldexp_f32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v2i64_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_xor_b32_e32 v5, v0, v1
; GFX9-NEXT: v_ffbh_i32_e32 v4, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX9-NEXT: v_add_u32_e32 v4, -1, v4
; GFX9-NEXT: v_add_u32_e32 v5, 32, v5
; GFX9-NEXT: v_min_u32_e32 v4, v4, v5
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v4
; GFX9-NEXT: v_ldexp_f32 v4, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v4, 16, 1
; GFX9-NEXT: v_xor_b32_e32 v1, v2, v3
; GFX9-NEXT: v_add3_u32 v5, v0, v4, s4
; GFX9-NEXT: v_ffbh_i32_e32 v0, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_add_u32_e32 v0, -1, v0
; GFX9-NEXT: v_add_u32_e32 v1, 32, v1
; GFX9-NEXT: v_min_u32_e32 v6, v0, v1
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v6, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v6
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v2i64_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_xor_b32_e32 v4, v0, v1
; GFX10-NEXT: v_xor_b32_e32 v5, v2, v3
; GFX10-NEXT: v_ffbh_i32_e32 v6, v1
; GFX10-NEXT: v_ffbh_i32_e32 v7, v3
; GFX10-NEXT: v_ashrrev_i32_e32 v4, 31, v4
; GFX10-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX10-NEXT: v_add_nc_u32_e32 v6, -1, v6
; GFX10-NEXT: v_add_nc_u32_e32 v7, -1, v7
; GFX10-NEXT: v_add_nc_u32_e32 v4, 32, v4
; GFX10-NEXT: v_add_nc_u32_e32 v5, 32, v5
; GFX10-NEXT: v_min_u32_e32 v4, v6, v4
; GFX10-NEXT: v_min_u32_e32 v5, v7, v5
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v1, v3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v2, 32, v4
; GFX10-NEXT: v_sub_nc_u32_e32 v3, 32, v5
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_ldexp_f32 v0, v0, v2
; GFX10-NEXT: v_ldexp_f32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v2i64_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_xor_b32_e32 v4, v0, v1
; GFX11-NEXT: v_xor_b32_e32 v5, v2, v3
; GFX11-NEXT: v_cls_i32_e32 v6, v1
; GFX11-NEXT: v_cls_i32_e32 v7, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_ashrrev_i32_e32 v4, 31, v4
; GFX11-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_nc_u32_e32 v6, -1, v6
; GFX11-NEXT: v_add_nc_u32_e32 v7, -1, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_add_nc_u32_e32 v4, 32, v4
; GFX11-NEXT: v_add_nc_u32_e32 v5, 32, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_u32_e32 v4, v6, v4
; GFX11-NEXT: v_min_u32_e32 v5, v7, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX11-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
; GFX11-NEXT: v_sub_nc_u32_e32 v2, 32, v4
; GFX11-NEXT: v_sub_nc_u32_e32 v3, 32, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_ldexp_f32 v0, v0, v2
; GFX11-NEXT: v_ldexp_f32 v1, v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <2 x i64> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_sitofp_v3i64_to_v3bf16(<3 x i64> %x) {
; GCN-LABEL: v_sitofp_v3i64_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_i32_e32 v6, v5
; GCN-NEXT: v_xor_b32_e32 v7, v4, v5
; GCN-NEXT: v_ffbh_i32_e32 v8, v3
; GCN-NEXT: v_xor_b32_e32 v9, v2, v3
; GCN-NEXT: v_ffbh_i32_e32 v10, v1
; GCN-NEXT: v_xor_b32_e32 v11, v0, v1
; GCN-NEXT: v_add_i32_e32 v6, vcc, -1, v6
; GCN-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GCN-NEXT: v_add_i32_e32 v8, vcc, -1, v8
; GCN-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GCN-NEXT: v_add_i32_e32 v10, vcc, -1, v10
; GCN-NEXT: v_ashrrev_i32_e32 v11, 31, v11
; GCN-NEXT: v_add_i32_e32 v7, vcc, 32, v7
; GCN-NEXT: v_add_i32_e32 v9, vcc, 32, v9
; GCN-NEXT: v_add_i32_e32 v11, vcc, 32, v11
; GCN-NEXT: v_min_u32_e32 v6, v6, v7
; GCN-NEXT: v_min_u32_e32 v7, v8, v9
; GCN-NEXT: v_min_u32_e32 v8, v10, v11
; GCN-NEXT: v_lshl_b64 v[4:5], v[4:5], v6
; GCN-NEXT: v_sub_i32_e32 v6, vcc, 32, v6
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v7
; GCN-NEXT: v_sub_i32_e32 v7, vcc, 32, v7
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v8
; GCN-NEXT: v_sub_i32_e32 v8, vcc, 32, v8
; GCN-NEXT: v_min_u32_e32 v4, 1, v4
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v4, v5, v4
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v4
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v3, v1, v6
; GCN-NEXT: v_ldexp_f32_e32 v1, v2, v7
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX7-NEXT: v_ffbh_i32_e32 v6, v5
; GFX7-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX7-NEXT: v_add_i32_e32 v6, vcc, -1, v6
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 32, v7
; GFX7-NEXT: v_min_u32_e32 v6, v6, v7
; GFX7-NEXT: v_lshl_b64 v[4:5], v[4:5], v6
; GFX7-NEXT: v_xor_b32_e32 v7, v2, v3
; GFX7-NEXT: v_min_u32_e32 v4, 1, v4
; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v6
; GFX7-NEXT: v_ffbh_i32_e32 v6, v3
; GFX7-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX7-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX7-NEXT: v_add_i32_e32 v6, vcc, -1, v6
; GFX7-NEXT: v_add_i32_e32 v7, vcc, 32, v7
; GFX7-NEXT: v_min_u32_e32 v6, v6, v7
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v6
; GFX7-NEXT: v_ldexp_f32_e32 v4, v4, v5
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_xor_b32_e32 v5, v0, v1
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_i32_e32 v3, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX7-NEXT: v_add_i32_e32 v3, vcc, -1, v3
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 32, v5
; GFX7-NEXT: v_min_u32_e32 v3, v3, v5
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v6
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v5
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX8-NEXT: v_ffbh_i32_e32 v6, v5
; GFX8-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX8-NEXT: v_add_u32_e32 v6, vcc, -1, v6
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 32, v7
; GFX8-NEXT: v_min_u32_e32 v6, v6, v7
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX8-NEXT: v_xor_b32_e32 v8, v0, v1
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX8-NEXT: v_sub_u32_e32 v5, vcc, 32, v6
; GFX8-NEXT: v_ffbh_i32_e32 v7, v1
; GFX8-NEXT: v_ldexp_f32 v4, v4, v5
; GFX8-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v7, vcc, -1, v7
; GFX8-NEXT: v_add_u32_e32 v8, vcc, 32, v8
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_min_u32_e32 v7, v7, v8
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v7, v[0:1]
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_xor_b32_e32 v6, v2, v3
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_ffbh_i32_e32 v5, v3
; GFX8-NEXT: v_ashrrev_i32_e32 v6, 31, v6
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, -1, v5
; GFX8-NEXT: v_add_u32_e32 v6, vcc, 32, v6
; GFX8-NEXT: v_min_u32_e32 v5, v5, v6
; GFX8-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v4
; GFX8-NEXT: v_sub_u32_e32 v4, vcc, 32, v7
; GFX8-NEXT: v_ldexp_f32 v0, v0, v4
; GFX8-NEXT: v_min_u32_e32 v2, 1, v2
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v6, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, 32, v5
; GFX8-NEXT: v_ldexp_f32 v2, v2, v3
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_alignbit_b32 v0, v2, v0, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX9-NEXT: v_ffbh_i32_e32 v6, v5
; GFX9-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX9-NEXT: v_add_u32_e32 v6, -1, v6
; GFX9-NEXT: v_add_u32_e32 v7, 32, v7
; GFX9-NEXT: v_min_u32_e32 v6, v6, v7
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX9-NEXT: v_xor_b32_e32 v7, v0, v1
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_sub_u32_e32 v5, 32, v6
; GFX9-NEXT: v_ffbh_i32_e32 v6, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX9-NEXT: v_add_u32_e32 v6, -1, v6
; GFX9-NEXT: v_add_u32_e32 v7, 32, v7
; GFX9-NEXT: v_min_u32_e32 v6, v6, v7
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v6, v[0:1]
; GFX9-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_ldexp_f32 v4, v4, v5
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v6
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc
; GFX9-NEXT: v_ldexp_f32 v5, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v5, 16, 1
; GFX9-NEXT: v_xor_b32_e32 v1, v2, v3
; GFX9-NEXT: v_add3_u32 v6, v0, v5, s4
; GFX9-NEXT: v_ffbh_i32_e32 v0, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_add_u32_e32 v0, -1, v0
; GFX9-NEXT: v_add_u32_e32 v1, 32, v1
; GFX9-NEXT: v_min_u32_e32 v7, v0, v1
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v7, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v5
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v7
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v4, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_xor_b32_e32 v8, v0, v1
; GFX10-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX10-NEXT: v_xor_b32_e32 v9, v2, v3
; GFX10-NEXT: v_ffbh_i32_e32 v10, v1
; GFX10-NEXT: v_ffbh_i32_e32 v6, v5
; GFX10-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX10-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX10-NEXT: v_ffbh_i32_e32 v11, v3
; GFX10-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX10-NEXT: v_add_nc_u32_e32 v10, -1, v10
; GFX10-NEXT: v_add_nc_u32_e32 v8, 32, v8
; GFX10-NEXT: v_add_nc_u32_e32 v6, -1, v6
; GFX10-NEXT: v_add_nc_u32_e32 v7, 32, v7
; GFX10-NEXT: v_add_nc_u32_e32 v11, -1, v11
; GFX10-NEXT: v_add_nc_u32_e32 v9, 32, v9
; GFX10-NEXT: v_min_u32_e32 v8, v10, v8
; GFX10-NEXT: v_min_u32_e32 v6, v6, v7
; GFX10-NEXT: v_min_u32_e32 v7, v11, v9
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v8, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX10-NEXT: v_sub_nc_u32_e32 v6, 32, v6
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v7, v[2:3]
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v4, 1, v4
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v1, v5, v4
; GFX10-NEXT: v_sub_nc_u32_e32 v4, 32, v7
; GFX10-NEXT: v_or_b32_e32 v2, v3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v3, 32, v8
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX10-NEXT: v_ldexp_f32 v0, v0, v3
; GFX10-NEXT: v_ldexp_f32 v1, v1, v6
; GFX10-NEXT: v_ldexp_f32 v2, v2, v4
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX10-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX10-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_xor_b32_e32 v8, v0, v1
; GFX11TRUE16-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX11TRUE16-NEXT: v_xor_b32_e32 v9, v2, v3
; GFX11TRUE16-NEXT: v_cls_i32_e32 v10, v1
; GFX11TRUE16-NEXT: v_cls_i32_e32 v6, v5
; GFX11TRUE16-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX11TRUE16-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX11TRUE16-NEXT: v_cls_i32_e32 v11, v3
; GFX11TRUE16-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v10, -1, v10
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v8, 32, v8
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v6, -1, v6
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v7, 32, v7
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v11, -1, v11
; GFX11TRUE16-NEXT: v_add_nc_u32_e32 v9, 32, v9
; GFX11TRUE16-NEXT: v_min_u32_e32 v8, v10, v8
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_min_u32_e32 v6, v6, v7
; GFX11TRUE16-NEXT: v_min_u32_e32 v7, v11, v9
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[0:1], v8, v[0:1]
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v6, 32, v6
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[2:3], v7, v[2:3]
; GFX11TRUE16-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11TRUE16-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v1, v5, v4
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v4, 32, v7
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v3, 32, v8
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11TRUE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v3
; GFX11TRUE16-NEXT: v_ldexp_f32 v1, v1, v6
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_ldexp_f32 v2, v2, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_sitofp_v3i64_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_xor_b32_e32 v8, v0, v1
; GFX11FAKE16-NEXT: v_xor_b32_e32 v7, v4, v5
; GFX11FAKE16-NEXT: v_xor_b32_e32 v9, v2, v3
; GFX11FAKE16-NEXT: v_cls_i32_e32 v10, v1
; GFX11FAKE16-NEXT: v_cls_i32_e32 v6, v5
; GFX11FAKE16-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX11FAKE16-NEXT: v_ashrrev_i32_e32 v7, 31, v7
; GFX11FAKE16-NEXT: v_cls_i32_e32 v11, v3
; GFX11FAKE16-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v10, -1, v10
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v8, 32, v8
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v6, -1, v6
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v7, 32, v7
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v11, -1, v11
; GFX11FAKE16-NEXT: v_add_nc_u32_e32 v9, 32, v9
; GFX11FAKE16-NEXT: v_min_u32_e32 v8, v10, v8
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_min_u32_e32 v6, v6, v7
; GFX11FAKE16-NEXT: v_min_u32_e32 v7, v11, v9
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[0:1], v8, v[0:1]
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v6, 32, v6
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[2:3], v7, v[2:3]
; GFX11FAKE16-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11FAKE16-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v1, v5, v4
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v4, 32, v7
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v3, 32, v8
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11FAKE16-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_ldexp_f32 v0, v0, v3
; GFX11FAKE16-NEXT: v_ldexp_f32 v1, v1, v6
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_ldexp_f32 v2, v2, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <3 x i64> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_sitofp_v4i64_to_v4bf16(<4 x i64> %x) {
; GCN-LABEL: v_sitofp_v4i64_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_i32_e32 v8, v7
; GCN-NEXT: v_xor_b32_e32 v9, v6, v7
; GCN-NEXT: v_ffbh_i32_e32 v10, v5
; GCN-NEXT: v_xor_b32_e32 v11, v4, v5
; GCN-NEXT: v_ffbh_i32_e32 v12, v3
; GCN-NEXT: v_xor_b32_e32 v13, v2, v3
; GCN-NEXT: v_ffbh_i32_e32 v14, v1
; GCN-NEXT: v_xor_b32_e32 v15, v0, v1
; GCN-NEXT: v_add_i32_e32 v8, vcc, -1, v8
; GCN-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GCN-NEXT: v_add_i32_e32 v10, vcc, -1, v10
; GCN-NEXT: v_ashrrev_i32_e32 v11, 31, v11
; GCN-NEXT: v_add_i32_e32 v12, vcc, -1, v12
; GCN-NEXT: v_ashrrev_i32_e32 v13, 31, v13
; GCN-NEXT: v_add_i32_e32 v14, vcc, -1, v14
; GCN-NEXT: v_ashrrev_i32_e32 v15, 31, v15
; GCN-NEXT: v_add_i32_e32 v9, vcc, 32, v9
; GCN-NEXT: v_add_i32_e32 v11, vcc, 32, v11
; GCN-NEXT: v_add_i32_e32 v13, vcc, 32, v13
; GCN-NEXT: v_add_i32_e32 v15, vcc, 32, v15
; GCN-NEXT: v_min_u32_e32 v8, v8, v9
; GCN-NEXT: v_min_u32_e32 v9, v10, v11
; GCN-NEXT: v_min_u32_e32 v10, v12, v13
; GCN-NEXT: v_min_u32_e32 v11, v14, v15
; GCN-NEXT: v_lshl_b64 v[6:7], v[6:7], v8
; GCN-NEXT: v_sub_i32_e32 v8, vcc, 32, v8
; GCN-NEXT: v_lshl_b64 v[4:5], v[4:5], v9
; GCN-NEXT: v_sub_i32_e32 v9, vcc, 32, v9
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v10
; GCN-NEXT: v_sub_i32_e32 v10, vcc, 32, v10
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v11
; GCN-NEXT: v_sub_i32_e32 v11, vcc, 32, v11
; GCN-NEXT: v_min_u32_e32 v6, 1, v6
; GCN-NEXT: v_min_u32_e32 v4, 1, v4
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v6, v7, v6
; GCN-NEXT: v_or_b32_e32 v4, v5, v4
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_i32_e32 v1, v6
; GCN-NEXT: v_cvt_f32_i32_e32 v3, v4
; GCN-NEXT: v_cvt_f32_i32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_i32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v4, v1, v8
; GCN-NEXT: v_ldexp_f32_e32 v3, v3, v9
; GCN-NEXT: v_ldexp_f32_e32 v1, v2, v10
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v11
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_sitofp_v4i64_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_xor_b32_e32 v9, v6, v7
; GFX7-NEXT: v_ffbh_i32_e32 v8, v7
; GFX7-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX7-NEXT: v_add_i32_e32 v8, vcc, -1, v8
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 32, v9
; GFX7-NEXT: v_min_u32_e32 v8, v8, v9
; GFX7-NEXT: v_lshl_b64 v[6:7], v[6:7], v8
; GFX7-NEXT: v_xor_b32_e32 v9, v4, v5
; GFX7-NEXT: v_min_u32_e32 v6, 1, v6
; GFX7-NEXT: v_or_b32_e32 v6, v7, v6
; GFX7-NEXT: v_sub_i32_e32 v7, vcc, 32, v8
; GFX7-NEXT: v_ffbh_i32_e32 v8, v5
; GFX7-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX7-NEXT: v_add_i32_e32 v8, vcc, -1, v8
; GFX7-NEXT: v_add_i32_e32 v9, vcc, 32, v9
; GFX7-NEXT: v_min_u32_e32 v8, v8, v9
; GFX7-NEXT: v_cvt_f32_i32_e32 v6, v6
; GFX7-NEXT: v_lshl_b64 v[4:5], v[4:5], v8
; GFX7-NEXT: v_min_u32_e32 v4, 1, v4
; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v8
; GFX7-NEXT: v_xor_b32_e32 v8, v2, v3
; GFX7-NEXT: v_ldexp_f32_e32 v6, v6, v7
; GFX7-NEXT: v_ffbh_i32_e32 v7, v3
; GFX7-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX7-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX7-NEXT: v_add_i32_e32 v7, vcc, -1, v7
; GFX7-NEXT: v_add_i32_e32 v8, vcc, 32, v8
; GFX7-NEXT: v_min_u32_e32 v7, v7, v8
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v7
; GFX7-NEXT: v_ldexp_f32_e32 v4, v4, v5
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_xor_b32_e32 v5, v0, v1
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_i32_e32 v3, v1
; GFX7-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX7-NEXT: v_add_i32_e32 v3, vcc, -1, v3
; GFX7-NEXT: v_add_i32_e32 v5, vcc, 32, v5
; GFX7-NEXT: v_min_u32_e32 v3, v3, v5
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_cvt_f32_i32_e32 v2, v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v7
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v5
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_sitofp_v4i64_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_xor_b32_e32 v9, v4, v5
; GFX8-NEXT: v_ffbh_i32_e32 v8, v5
; GFX8-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX8-NEXT: v_add_u32_e32 v8, vcc, -1, v8
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 32, v9
; GFX8-NEXT: v_min_u32_e32 v8, v8, v9
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX8-NEXT: v_sub_u32_e32 v5, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v8, v4, v5
; GFX8-NEXT: v_bfe_u32 v4, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v8
; GFX8-NEXT: v_xor_b32_e32 v5, v6, v7
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v4
; GFX8-NEXT: v_ffbh_i32_e32 v4, v7
; GFX8-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX8-NEXT: v_add_u32_e32 v4, vcc, -1, v4
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 32, v5
; GFX8-NEXT: v_min_u32_e32 v10, v4, v5
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v10, v[6:7]
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v8
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v6, vcc
; GFX8-NEXT: v_xor_b32_e32 v9, v0, v1
; GFX8-NEXT: v_ffbh_i32_e32 v8, v1
; GFX8-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX8-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX8-NEXT: v_add_u32_e32 v8, vcc, -1, v8
; GFX8-NEXT: v_add_u32_e32 v9, vcc, 32, v9
; GFX8-NEXT: v_min_u32_e32 v8, v8, v9
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v8, v[0:1]
; GFX8-NEXT: v_sub_u32_e32 v6, vcc, 32, v10
; GFX8-NEXT: v_ldexp_f32 v4, v4, v6
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v6, v0, v1
; GFX8-NEXT: v_bfe_u32 v0, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v6
; GFX8-NEXT: v_xor_b32_e32 v1, v2, v3
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v0
; GFX8-NEXT: v_ffbh_i32_e32 v0, v3
; GFX8-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, -1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 32, v1
; GFX8-NEXT: v_min_u32_e32 v8, v0, v1
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v8, v[2:3]
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v6
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v1, v7, v2, vcc
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: v_alignbit_b32 v1, v4, v5, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_sitofp_v4i64_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_xor_b32_e32 v9, v4, v5
; GFX9-NEXT: v_ffbh_i32_e32 v8, v5
; GFX9-NEXT: v_ashrrev_i32_e32 v9, 31, v9
; GFX9-NEXT: v_add_u32_e32 v8, -1, v8
; GFX9-NEXT: v_add_u32_e32 v9, 32, v9
; GFX9-NEXT: v_min_u32_e32 v8, v8, v9
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX9-NEXT: v_sub_u32_e32 v5, 32, v8
; GFX9-NEXT: v_ldexp_f32 v8, v4, v5
; GFX9-NEXT: v_bfe_u32 v4, v8, 16, 1
; GFX9-NEXT: v_xor_b32_e32 v5, v6, v7
; GFX9-NEXT: v_add3_u32 v9, v4, v8, s4
; GFX9-NEXT: v_ffbh_i32_e32 v4, v7
; GFX9-NEXT: v_ashrrev_i32_e32 v5, 31, v5
; GFX9-NEXT: v_add_u32_e32 v4, -1, v4
; GFX9-NEXT: v_add_u32_e32 v5, 32, v5
; GFX9-NEXT: v_min_u32_e32 v10, v4, v5
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v10, v[6:7]
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_xor_b32_e32 v8, v0, v1
; GFX9-NEXT: v_ffbh_i32_e32 v7, v1
; GFX9-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX9-NEXT: v_add_u32_e32 v7, -1, v7
; GFX9-NEXT: v_add_u32_e32 v8, 32, v8
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_min_u32_e32 v7, v7, v8
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v7, v[0:1]
; GFX9-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v6, vcc
; GFX9-NEXT: v_sub_u32_e32 v6, 32, v10
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_ldexp_f32 v4, v4, v6
; GFX9-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v6, v6, v4, s4
; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v7
; GFX9-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc
; GFX9-NEXT: v_ldexp_f32 v6, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v6, 16, 1
; GFX9-NEXT: v_xor_b32_e32 v1, v2, v3
; GFX9-NEXT: v_add3_u32 v7, v0, v6, s4
; GFX9-NEXT: v_ffbh_i32_e32 v0, v3
; GFX9-NEXT: v_ashrrev_i32_e32 v1, 31, v1
; GFX9-NEXT: v_add_u32_e32 v0, -1, v0
; GFX9-NEXT: v_add_u32_e32 v1, 32, v1
; GFX9-NEXT: v_min_u32_e32 v8, v0, v1
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v8, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v6
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v1, v7, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v8
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: v_perm_b32 v1, v4, v5, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_sitofp_v4i64_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_xor_b32_e32 v8, v4, v5
; GFX10-NEXT: v_ffbh_i32_e32 v9, v5
; GFX10-NEXT: v_xor_b32_e32 v11, v6, v7
; GFX10-NEXT: v_xor_b32_e32 v13, v0, v1
; GFX10-NEXT: v_ffbh_i32_e32 v10, v7
; GFX10-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX10-NEXT: v_add_nc_u32_e32 v9, -1, v9
; GFX10-NEXT: v_ffbh_i32_e32 v12, v1
; GFX10-NEXT: v_xor_b32_e32 v14, v2, v3
; GFX10-NEXT: v_ashrrev_i32_e32 v11, 31, v11
; GFX10-NEXT: v_add_nc_u32_e32 v8, 32, v8
; GFX10-NEXT: v_add_nc_u32_e32 v10, -1, v10
; GFX10-NEXT: v_add_nc_u32_e32 v12, -1, v12
; GFX10-NEXT: v_ashrrev_i32_e32 v14, 31, v14
; GFX10-NEXT: v_add_nc_u32_e32 v11, 32, v11
; GFX10-NEXT: v_min_u32_e32 v8, v9, v8
; GFX10-NEXT: v_ashrrev_i32_e32 v9, 31, v13
; GFX10-NEXT: v_ffbh_i32_e32 v13, v3
; GFX10-NEXT: v_add_nc_u32_e32 v14, 32, v14
; GFX10-NEXT: v_min_u32_e32 v10, v10, v11
; GFX10-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX10-NEXT: v_add_nc_u32_e32 v9, 32, v9
; GFX10-NEXT: v_add_nc_u32_e32 v13, -1, v13
; GFX10-NEXT: v_lshlrev_b64 v[6:7], v10, v[6:7]
; GFX10-NEXT: v_min_u32_e32 v9, v12, v9
; GFX10-NEXT: v_min_u32_e32 v11, v13, v14
; GFX10-NEXT: v_min_u32_e32 v4, 1, v4
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v9, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v11, v[2:3]
; GFX10-NEXT: v_or_b32_e32 v4, v5, v4
; GFX10-NEXT: v_min_u32_e32 v5, 1, v6
; GFX10-NEXT: v_sub_nc_u32_e32 v6, 32, v8
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX10-NEXT: v_or_b32_e32 v5, v7, v5
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v1, v3, v2
; GFX10-NEXT: v_ldexp_f32 v2, v4, v6
; GFX10-NEXT: v_cvt_f32_i32_e32 v3, v5
; GFX10-NEXT: v_sub_nc_u32_e32 v4, 32, v10
; GFX10-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v5, 32, v9
; GFX10-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX10-NEXT: v_sub_nc_u32_e32 v6, 32, v11
; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
; GFX10-NEXT: v_ldexp_f32 v3, v3, v4
; GFX10-NEXT: v_ldexp_f32 v0, v0, v5
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_ldexp_f32 v1, v1, v6
; GFX10-NEXT: v_add3_u32 v4, v7, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v6, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v5, v7, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v7, v8, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v4, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_sitofp_v4i64_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_xor_b32_e32 v8, v4, v5
; GFX11-NEXT: v_cls_i32_e32 v9, v5
; GFX11-NEXT: v_xor_b32_e32 v11, v6, v7
; GFX11-NEXT: v_xor_b32_e32 v13, v0, v1
; GFX11-NEXT: v_cls_i32_e32 v10, v7
; GFX11-NEXT: v_ashrrev_i32_e32 v8, 31, v8
; GFX11-NEXT: v_add_nc_u32_e32 v9, -1, v9
; GFX11-NEXT: v_cls_i32_e32 v12, v1
; GFX11-NEXT: v_xor_b32_e32 v14, v2, v3
; GFX11-NEXT: v_ashrrev_i32_e32 v11, 31, v11
; GFX11-NEXT: v_add_nc_u32_e32 v8, 32, v8
; GFX11-NEXT: v_add_nc_u32_e32 v10, -1, v10
; GFX11-NEXT: v_add_nc_u32_e32 v12, -1, v12
; GFX11-NEXT: v_ashrrev_i32_e32 v14, 31, v14
; GFX11-NEXT: v_add_nc_u32_e32 v11, 32, v11
; GFX11-NEXT: v_min_u32_e32 v8, v9, v8
; GFX11-NEXT: v_ashrrev_i32_e32 v9, 31, v13
; GFX11-NEXT: v_cls_i32_e32 v13, v3
; GFX11-NEXT: v_add_nc_u32_e32 v14, 32, v14
; GFX11-NEXT: v_min_u32_e32 v10, v10, v11
; GFX11-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX11-NEXT: v_add_nc_u32_e32 v9, 32, v9
; GFX11-NEXT: v_add_nc_u32_e32 v13, -1, v13
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_lshlrev_b64 v[6:7], v10, v[6:7]
; GFX11-NEXT: v_min_u32_e32 v9, v12, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_min_u32_e32 v11, v13, v14
; GFX11-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v9, v[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_lshlrev_b64 v[2:3], v11, v[2:3]
; GFX11-NEXT: v_or_b32_e32 v4, v5, v4
; GFX11-NEXT: v_min_u32_e32 v5, 1, v6
; GFX11-NEXT: v_sub_nc_u32_e32 v6, 32, v8
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11-NEXT: v_cvt_f32_i32_e32 v4, v4
; GFX11-NEXT: v_or_b32_e32 v5, v7, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_ldexp_f32 v2, v4, v6
; GFX11-NEXT: v_cvt_f32_i32_e32 v3, v5
; GFX11-NEXT: v_sub_nc_u32_e32 v4, 32, v10
; GFX11-NEXT: v_cvt_f32_i32_e32 v0, v0
; GFX11-NEXT: v_sub_nc_u32_e32 v5, 32, v9
; GFX11-NEXT: v_cvt_f32_i32_e32 v1, v1
; GFX11-NEXT: v_sub_nc_u32_e32 v6, 32, v11
; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
; GFX11-NEXT: v_ldexp_f32 v3, v3, v4
; GFX11-NEXT: v_ldexp_f32 v0, v0, v5
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_ldexp_f32 v1, v1, v6
; GFX11-NEXT: v_add3_u32 v4, v7, v2, 0x7fff
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX11-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX11-NEXT: v_add3_u32 v4, v6, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v7, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v7, v8, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v5, v6, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v4, v9, vcc_lo
; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = sitofp <4 x i64> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_uitofp_i16_to_bf16(i16 %x) {
; GCN-LABEL: v_uitofp_i16_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_i16_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_i16_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_i16_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_i16_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_i16_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp i16 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_uitofp_v2i16_to_v2bf16(<2 x i16> %x) {
; GCN-LABEL: v_uitofp_v2i16_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v2i16_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v2i16_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_sdwa v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v2i16_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_sdwa v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT: v_add3_u32 v2, v2, v1, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v2i16_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_sdwa v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v2i16_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v0
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <2 x i16> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_uitofp_v3i16_to_v3bf16(<3 x i16> %x) {
; GCN-LABEL: v_uitofp_v3i16_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_u32_sdwa v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX8-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v4
; GFX8-NEXT: v_add_u32_e32 v2, vcc, s4, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc
; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_u32_sdwa v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v2, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_sdwa v2, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff, v0
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_uitofp_v3i16_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v0
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <3 x i16> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_uitofp_v4i16_to_v4bf16(<4 x i16> %x) {
; GCN-LABEL: v_uitofp_v4i16_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GCN-NEXT: v_cvt_f32_u32_e32 v3, v3
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0x7fff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v4i16_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff, v3
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0x7fff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v4i16_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_sdwa v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_cvt_f32_u32_sdwa v5, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_bfe_u32 v3, v5, 16, 1
; GFX8-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v5
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v5
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT: v_cndmask_b32_e32 v3, v3, v4, vcc
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v4i16_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_sdwa v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_cvt_f32_u32_sdwa v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX9-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_bfe_u32 v3, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v4, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v3, v5, vcc
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v4i16_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_sdwa v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_u32_sdwa v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
; GFX10-NEXT: v_cvt_f32_u32_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_cvt_f32_u32_sdwa v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v8, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v10, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v8, v8, v3, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v11, 0x400000, v0
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add3_u32 v10, v10, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v10, v11, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v4i16_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff, v1
; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff, v0
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: v_add3_u32 v6, v6, v1, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX11-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_bfe_u32 v8, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v10, v0, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_or_b32_e32 v11, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v8, v8, v3, 0x7fff
; GFX11-NEXT: v_add3_u32 v10, v10, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v8, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v10, v11, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v3, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v6, v7, vcc_lo
; GFX11-NEXT: v_perm_b32 v1, v1, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <4 x i16> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_uitofp_i32_to_bf16(i32 %x) {
; GCN-LABEL: v_uitofp_i32_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_i32_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_i32_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v1, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_i32_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_i32_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_i32_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp i32 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_uitofp_v2i32_to_v2bf16(<2 x i32> %x) {
; GCN-LABEL: v_uitofp_v2i32_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v2i32_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v2i32_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX8-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v0
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v2i32_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v1, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v2i32_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v2i32_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <2 x i32> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_uitofp_v3i32_to_v3bf16(<3 x i32> %x) {
; GCN-LABEL: v_uitofp_v3i32_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: v_mov_b32_e32 v1, v2
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX9-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v2, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v2, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v2, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v2, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_uitofp_v3i32_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v1, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v2, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <3 x i32> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_uitofp_v4i32_to_v4bf16(<4 x i32> %x) {
; GCN-LABEL: v_uitofp_v4i32_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_cvt_f32_u32_e32 v3, v3
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v1
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0x7fff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v4i32_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0x7fff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0x7fff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v4i32_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX8-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v2
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX8-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v4, vcc
; GFX8-NEXT: v_bfe_u32 v5, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v5
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v1, v0, 16
; GFX8-NEXT: v_alignbit_b32 v1, v3, v2, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v4i32_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX9-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v2, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc
; GFX9-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v4, v4, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_perm_b32 v1, v3, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v4i32_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX10-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v9, v9, v1, 0x7fff
; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v4, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v4i32_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: v_cvt_f32_u32_e32 v3, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v4, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v9, v1, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v9, v9, v1, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_bfe_u32 v6, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v1, v9, v10, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v4, vcc_lo
; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <4 x i32> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_uitofp_i64_to_bf16(i64 %x) {
; GCN-LABEL: v_uitofp_i64_to_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_u32_e32 v2, v1
; GCN-NEXT: v_min_u32_e32 v2, 32, v2
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_i64_to_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_ffbh_u32_e32 v2, v1
; GFX7-NEXT: v_min_u32_e32 v2, 32, v2
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_i64_to_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_ffbh_u32_e32 v2, v1
; GFX8-NEXT: v_min_u32_e32 v2, 32, v2
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
; GFX8-NEXT: v_ldexp_f32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_i64_to_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_ffbh_u32_e32 v2, v1
; GFX9-NEXT: v_min_u32_e32 v2, 32, v2
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v2
; GFX9-NEXT: v_ldexp_f32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_i64_to_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_ffbh_u32_e32 v2, v1
; GFX10-NEXT: v_min_u32_e32 v2, 32, v2
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_sub_nc_u32_e32 v1, 32, v2
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_ldexp_f32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_i64_to_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_clz_i32_u32_e32 v2, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_u32_e32 v2, 32, v2
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: v_sub_nc_u32_e32 v1, 32, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp i64 %x to bfloat
ret bfloat %op
}
define <2 x bfloat> @v_uitofp_v2i64_to_v2bf16(<2 x i64> %x) {
; GCN-LABEL: v_uitofp_v2i64_to_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_u32_e32 v4, v3
; GCN-NEXT: v_ffbh_u32_e32 v5, v1
; GCN-NEXT: v_min_u32_e32 v4, 32, v4
; GCN-NEXT: v_min_u32_e32 v5, 32, v5
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v4
; GCN-NEXT: v_sub_i32_e32 v4, vcc, 32, v4
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v5
; GCN-NEXT: v_sub_i32_e32 v5, vcc, 32, v5
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v1, v1, v4
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v5
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v2i64_to_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_ffbh_u32_e32 v4, v3
; GFX7-NEXT: v_min_u32_e32 v4, 32, v4
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v4
; GFX7-NEXT: v_sub_i32_e32 v4, vcc, 32, v4
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_u32_e32 v3, v1
; GFX7-NEXT: v_min_u32_e32 v3, 32, v3
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v4
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v2i64_to_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_ffbh_u32_e32 v4, v1
; GFX8-NEXT: v_min_u32_e32 v4, 32, v4
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v4
; GFX8-NEXT: v_ldexp_f32 v4, v0, v1
; GFX8-NEXT: v_bfe_u32 v0, v4, 16, 1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v4
; GFX8-NEXT: v_add_u32_e32 v5, vcc, 0x7fff, v0
; GFX8-NEXT: v_ffbh_u32_e32 v0, v3
; GFX8-NEXT: v_min_u32_e32 v6, 32, v0
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v6, v[2:3]
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v4
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, 32, v6
; GFX8-NEXT: v_ldexp_f32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v2i64_to_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_ffbh_u32_e32 v4, v1
; GFX9-NEXT: v_min_u32_e32 v4, 32, v4
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v4
; GFX9-NEXT: v_ldexp_f32 v4, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v5, v0, v4, s4
; GFX9-NEXT: v_ffbh_u32_e32 v0, v3
; GFX9-NEXT: v_min_u32_e32 v6, 32, v0
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v6, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v6
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v2i64_to_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_ffbh_u32_e32 v4, v1
; GFX10-NEXT: v_ffbh_u32_e32 v5, v3
; GFX10-NEXT: v_min_u32_e32 v4, 32, v4
; GFX10-NEXT: v_min_u32_e32 v5, 32, v5
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v1, v3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v2, 32, v4
; GFX10-NEXT: v_sub_nc_u32_e32 v3, 32, v5
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_ldexp_f32 v0, v0, v2
; GFX10-NEXT: v_ldexp_f32 v1, v1, v3
; GFX10-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v2i64_to_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_clz_i32_u32_e32 v4, v1
; GFX11-NEXT: v_clz_i32_u32_e32 v5, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_u32_e32 v4, 32, v4
; GFX11-NEXT: v_min_u32_e32 v5, 32, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
; GFX11-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
; GFX11-NEXT: v_sub_nc_u32_e32 v2, 32, v4
; GFX11-NEXT: v_sub_nc_u32_e32 v3, 32, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_ldexp_f32 v0, v0, v2
; GFX11-NEXT: v_ldexp_f32 v1, v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX11-NEXT: v_add3_u32 v2, v2, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <2 x i64> %x to <2 x bfloat>
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_uitofp_v3i64_to_v3bf16(<3 x i64> %x) {
; GCN-LABEL: v_uitofp_v3i64_to_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_u32_e32 v6, v5
; GCN-NEXT: v_ffbh_u32_e32 v7, v3
; GCN-NEXT: v_ffbh_u32_e32 v8, v1
; GCN-NEXT: v_min_u32_e32 v6, 32, v6
; GCN-NEXT: v_min_u32_e32 v7, 32, v7
; GCN-NEXT: v_min_u32_e32 v8, 32, v8
; GCN-NEXT: v_lshl_b64 v[4:5], v[4:5], v6
; GCN-NEXT: v_sub_i32_e32 v6, vcc, 32, v6
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v7
; GCN-NEXT: v_sub_i32_e32 v7, vcc, 32, v7
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v8
; GCN-NEXT: v_sub_i32_e32 v8, vcc, 32, v8
; GCN-NEXT: v_min_u32_e32 v4, 1, v4
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v4, v5, v4
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v4
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v3, v1, v6
; GCN-NEXT: v_ldexp_f32_e32 v1, v2, v7
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_ffbh_u32_e32 v6, v5
; GFX7-NEXT: v_min_u32_e32 v6, 32, v6
; GFX7-NEXT: v_lshl_b64 v[4:5], v[4:5], v6
; GFX7-NEXT: v_min_u32_e32 v4, 1, v4
; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v6
; GFX7-NEXT: v_ffbh_u32_e32 v6, v3
; GFX7-NEXT: v_min_u32_e32 v6, 32, v6
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v6
; GFX7-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_u32_e32 v3, v1
; GFX7-NEXT: v_min_u32_e32 v3, 32, v3
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_ldexp_f32_e32 v4, v4, v5
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v6
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v5
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_ffbh_u32_e32 v6, v5
; GFX8-NEXT: v_min_u32_e32 v6, 32, v6
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX8-NEXT: v_ffbh_u32_e32 v7, v1
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX8-NEXT: v_sub_u32_e32 v5, vcc, 32, v6
; GFX8-NEXT: v_min_u32_e32 v7, 32, v7
; GFX8-NEXT: v_ldexp_f32 v4, v4, v5
; GFX8-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v7, v[0:1]
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v4
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_ffbh_u32_e32 v5, v3
; GFX8-NEXT: v_min_u32_e32 v5, 32, v5
; GFX8-NEXT: v_lshlrev_b64 v[2:3], v5, v[2:3]
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v4
; GFX8-NEXT: v_sub_u32_e32 v4, vcc, 32, v7
; GFX8-NEXT: v_ldexp_f32 v0, v0, v4
; GFX8-NEXT: v_min_u32_e32 v2, 1, v2
; GFX8-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX8-NEXT: v_or_b32_e32 v2, v3, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v6, vcc
; GFX8-NEXT: v_sub_u32_e32 v3, vcc, 32, v5
; GFX8-NEXT: v_ldexp_f32 v2, v2, v3
; GFX8-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v2
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v2
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_alignbit_b32 v0, v2, v0, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_ffbh_u32_e32 v6, v5
; GFX9-NEXT: v_min_u32_e32 v6, 32, v6
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v6, v[4:5]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_sub_u32_e32 v5, 32, v6
; GFX9-NEXT: v_ffbh_u32_e32 v6, v1
; GFX9-NEXT: v_min_u32_e32 v6, 32, v6
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v6, v[0:1]
; GFX9-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_ldexp_f32 v4, v4, v5
; GFX9-NEXT: v_bfe_u32 v5, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v5, v5, v4, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v6
; GFX9-NEXT: v_cndmask_b32_e32 v4, v5, v7, vcc
; GFX9-NEXT: v_ldexp_f32 v5, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v5, 16, 1
; GFX9-NEXT: v_add3_u32 v6, v0, v5, s4
; GFX9-NEXT: v_ffbh_u32_e32 v0, v3
; GFX9-NEXT: v_min_u32_e32 v7, 32, v0
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v7, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v5
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v7
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v4, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_ffbh_u32_e32 v6, v1
; GFX10-NEXT: v_ffbh_u32_e32 v8, v3
; GFX10-NEXT: v_ffbh_u32_e32 v7, v5
; GFX10-NEXT: v_min_u32_e32 v6, 32, v6
; GFX10-NEXT: v_min_u32_e32 v8, 32, v8
; GFX10-NEXT: v_min_u32_e32 v7, 32, v7
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v6, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v8, v[2:3]
; GFX10-NEXT: v_lshlrev_b64 v[4:5], v7, v[4:5]
; GFX10-NEXT: v_sub_nc_u32_e32 v7, 32, v7
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_min_u32_e32 v4, 1, v4
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v2, v3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v3, 32, v6
; GFX10-NEXT: v_or_b32_e32 v1, v5, v4
; GFX10-NEXT: v_sub_nc_u32_e32 v4, 32, v8
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_ldexp_f32 v0, v0, v3
; GFX10-NEXT: v_ldexp_f32 v2, v2, v4
; GFX10-NEXT: v_ldexp_f32 v1, v1, v7
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX10-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_clz_i32_u32_e32 v6, v1
; GFX11TRUE16-NEXT: v_clz_i32_u32_e32 v7, v5
; GFX11TRUE16-NEXT: v_clz_i32_u32_e32 v8, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_min_u32_e32 v6, 32, v6
; GFX11TRUE16-NEXT: v_min_u32_e32 v7, 32, v7
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_min_u32_e32 v8, 32, v8
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[0:1], v6, v[0:1]
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[4:5], v7, v[4:5]
; GFX11TRUE16-NEXT: v_lshlrev_b64 v[2:3], v8, v[2:3]
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v7, 32, v7
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_or_b32_e32 v1, v5, v4
; GFX11TRUE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v3, 32, v6
; GFX11TRUE16-NEXT: v_sub_nc_u32_e32 v4, 32, v8
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11TRUE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_ldexp_f32 v0, v0, v3
; GFX11TRUE16-NEXT: v_ldexp_f32 v1, v1, v7
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_ldexp_f32 v2, v2, v4
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_uitofp_v3i64_to_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_clz_i32_u32_e32 v6, v1
; GFX11FAKE16-NEXT: v_clz_i32_u32_e32 v7, v5
; GFX11FAKE16-NEXT: v_clz_i32_u32_e32 v8, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_min_u32_e32 v6, 32, v6
; GFX11FAKE16-NEXT: v_min_u32_e32 v7, 32, v7
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_min_u32_e32 v8, 32, v8
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[0:1], v6, v[0:1]
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[4:5], v7, v[4:5]
; GFX11FAKE16-NEXT: v_lshlrev_b64 v[2:3], v8, v[2:3]
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v7, 32, v7
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_or_b32_e32 v1, v5, v4
; GFX11FAKE16-NEXT: v_or_b32_e32 v2, v3, v2
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v3, 32, v6
; GFX11FAKE16-NEXT: v_sub_nc_u32_e32 v4, 32, v8
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11FAKE16-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_ldexp_f32 v0, v0, v3
; GFX11FAKE16-NEXT: v_ldexp_f32 v1, v1, v7
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_ldexp_f32 v2, v2, v4
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v2, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v0
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v2
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v2, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v2, v0, 0x7060302
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <3 x i64> %x to <3 x bfloat>
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_uitofp_v4i64_to_v4bf16(<4 x i64> %x) {
; GCN-LABEL: v_uitofp_v4i64_to_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_ffbh_u32_e32 v8, v7
; GCN-NEXT: v_ffbh_u32_e32 v9, v5
; GCN-NEXT: v_ffbh_u32_e32 v10, v3
; GCN-NEXT: v_ffbh_u32_e32 v11, v1
; GCN-NEXT: v_min_u32_e32 v8, 32, v8
; GCN-NEXT: v_min_u32_e32 v9, 32, v9
; GCN-NEXT: v_min_u32_e32 v10, 32, v10
; GCN-NEXT: v_min_u32_e32 v11, 32, v11
; GCN-NEXT: v_lshl_b64 v[6:7], v[6:7], v8
; GCN-NEXT: v_sub_i32_e32 v8, vcc, 32, v8
; GCN-NEXT: v_lshl_b64 v[4:5], v[4:5], v9
; GCN-NEXT: v_sub_i32_e32 v9, vcc, 32, v9
; GCN-NEXT: v_lshl_b64 v[2:3], v[2:3], v10
; GCN-NEXT: v_sub_i32_e32 v10, vcc, 32, v10
; GCN-NEXT: v_lshl_b64 v[0:1], v[0:1], v11
; GCN-NEXT: v_sub_i32_e32 v11, vcc, 32, v11
; GCN-NEXT: v_min_u32_e32 v6, 1, v6
; GCN-NEXT: v_min_u32_e32 v4, 1, v4
; GCN-NEXT: v_min_u32_e32 v2, 1, v2
; GCN-NEXT: v_min_u32_e32 v0, 1, v0
; GCN-NEXT: v_or_b32_e32 v6, v7, v6
; GCN-NEXT: v_or_b32_e32 v4, v5, v4
; GCN-NEXT: v_or_b32_e32 v2, v3, v2
; GCN-NEXT: v_or_b32_e32 v0, v1, v0
; GCN-NEXT: v_cvt_f32_u32_e32 v1, v6
; GCN-NEXT: v_cvt_f32_u32_e32 v3, v4
; GCN-NEXT: v_cvt_f32_u32_e32 v2, v2
; GCN-NEXT: v_cvt_f32_u32_e32 v0, v0
; GCN-NEXT: v_ldexp_f32_e32 v4, v1, v8
; GCN-NEXT: v_ldexp_f32_e32 v3, v3, v9
; GCN-NEXT: v_ldexp_f32_e32 v1, v2, v10
; GCN-NEXT: v_ldexp_f32_e32 v0, v0, v11
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_uitofp_v4i64_to_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_ffbh_u32_e32 v8, v7
; GFX7-NEXT: v_min_u32_e32 v8, 32, v8
; GFX7-NEXT: v_lshl_b64 v[6:7], v[6:7], v8
; GFX7-NEXT: v_min_u32_e32 v6, 1, v6
; GFX7-NEXT: v_or_b32_e32 v6, v7, v6
; GFX7-NEXT: v_cvt_f32_u32_e32 v6, v6
; GFX7-NEXT: v_sub_i32_e32 v7, vcc, 32, v8
; GFX7-NEXT: v_ffbh_u32_e32 v8, v5
; GFX7-NEXT: v_ldexp_f32_e32 v6, v6, v7
; GFX7-NEXT: v_ffbh_u32_e32 v7, v3
; GFX7-NEXT: v_min_u32_e32 v7, 32, v7
; GFX7-NEXT: v_lshl_b64 v[2:3], v[2:3], v7
; GFX7-NEXT: v_min_u32_e32 v8, 32, v8
; GFX7-NEXT: v_min_u32_e32 v2, 1, v2
; GFX7-NEXT: v_lshl_b64 v[4:5], v[4:5], v8
; GFX7-NEXT: v_or_b32_e32 v2, v3, v2
; GFX7-NEXT: v_ffbh_u32_e32 v3, v1
; GFX7-NEXT: v_min_u32_e32 v3, 32, v3
; GFX7-NEXT: v_min_u32_e32 v4, 1, v4
; GFX7-NEXT: v_lshl_b64 v[0:1], v[0:1], v3
; GFX7-NEXT: v_or_b32_e32 v4, v5, v4
; GFX7-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX7-NEXT: v_min_u32_e32 v0, 1, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v2, v2
; GFX7-NEXT: v_or_b32_e32 v0, v1, v0
; GFX7-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v8
; GFX7-NEXT: v_ldexp_f32_e32 v4, v4, v5
; GFX7-NEXT: v_sub_i32_e32 v5, vcc, 32, v7
; GFX7-NEXT: v_ldexp_f32_e32 v1, v2, v5
; GFX7-NEXT: v_sub_i32_e32 v2, vcc, 32, v3
; GFX7-NEXT: v_ldexp_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uitofp_v4i64_to_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_ffbh_u32_e32 v8, v5
; GFX8-NEXT: v_min_u32_e32 v8, 32, v8
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX8-NEXT: v_sub_u32_e32 v5, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v8, v4, v5
; GFX8-NEXT: v_bfe_u32 v4, v8, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v8
; GFX8-NEXT: v_add_u32_e32 v9, vcc, s4, v4
; GFX8-NEXT: v_ffbh_u32_e32 v4, v7
; GFX8-NEXT: v_min_u32_e32 v10, 32, v4
; GFX8-NEXT: v_lshlrev_b64 v[4:5], v10, v[6:7]
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v8
; GFX8-NEXT: v_min_u32_e32 v4, 1, v4
; GFX8-NEXT: v_or_b32_e32 v4, v5, v4
; GFX8-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX8-NEXT: v_ffbh_u32_e32 v8, v1
; GFX8-NEXT: v_min_u32_e32 v8, 32, v8
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v8, v[0:1]
; GFX8-NEXT: v_cndmask_b32_e32 v5, v9, v6, vcc
; GFX8-NEXT: v_sub_u32_e32 v6, vcc, 32, v10
; GFX8-NEXT: v_ldexp_f32 v4, v4, v6
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_add_u32_e32 v6, vcc, s4, v6
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc
; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v6, v0, v1
; GFX8-NEXT: v_bfe_u32 v0, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v0, vcc, v0, v6
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v0
; GFX8-NEXT: v_ffbh_u32_e32 v0, v3
; GFX8-NEXT: v_min_u32_e32 v8, 32, v0
; GFX8-NEXT: v_lshlrev_b64 v[0:1], v8, v[2:3]
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v6
; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v1, v7, v2, vcc
; GFX8-NEXT: v_sub_u32_e32 v2, vcc, 32, v8
; GFX8-NEXT: v_ldexp_f32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GFX8-NEXT: v_alignbit_b32 v1, v4, v5, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_uitofp_v4i64_to_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_ffbh_u32_e32 v8, v5
; GFX9-NEXT: v_min_u32_e32 v8, 32, v8
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX9-NEXT: v_sub_u32_e32 v5, 32, v8
; GFX9-NEXT: v_ldexp_f32 v8, v4, v5
; GFX9-NEXT: v_bfe_u32 v4, v8, 16, 1
; GFX9-NEXT: v_add3_u32 v9, v4, v8, s4
; GFX9-NEXT: v_ffbh_u32_e32 v4, v7
; GFX9-NEXT: v_min_u32_e32 v10, 32, v4
; GFX9-NEXT: v_lshlrev_b64 v[4:5], v10, v[6:7]
; GFX9-NEXT: v_ffbh_u32_e32 v7, v1
; GFX9-NEXT: v_min_u32_e32 v4, 1, v4
; GFX9-NEXT: v_min_u32_e32 v7, 32, v7
; GFX9-NEXT: v_or_b32_e32 v4, v5, v4
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v7, v[0:1]
; GFX9-NEXT: v_cvt_f32_u32_e32 v4, v4
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v8
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v5, v9, v6, vcc
; GFX9-NEXT: v_sub_u32_e32 v6, 32, v10
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_ldexp_f32 v4, v4, v6
; GFX9-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX9-NEXT: v_add3_u32 v6, v6, v4, s4
; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v4
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT: v_sub_u32_e32 v1, 32, v7
; GFX9-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc
; GFX9-NEXT: v_ldexp_f32 v6, v0, v1
; GFX9-NEXT: v_bfe_u32 v0, v6, 16, 1
; GFX9-NEXT: v_add3_u32 v7, v0, v6, s4
; GFX9-NEXT: v_ffbh_u32_e32 v0, v3
; GFX9-NEXT: v_min_u32_e32 v8, 32, v0
; GFX9-NEXT: v_lshlrev_b64 v[0:1], v8, v[2:3]
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v6
; GFX9-NEXT: v_min_u32_e32 v0, 1, v0
; GFX9-NEXT: v_or_b32_e32 v0, v1, v0
; GFX9-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v1, v7, v2, vcc
; GFX9-NEXT: v_sub_u32_e32 v2, 32, v8
; GFX9-NEXT: v_ldexp_f32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT: v_perm_b32 v1, v4, v5, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_uitofp_v4i64_to_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_ffbh_u32_e32 v8, v5
; GFX10-NEXT: v_ffbh_u32_e32 v10, v1
; GFX10-NEXT: v_ffbh_u32_e32 v11, v3
; GFX10-NEXT: v_ffbh_u32_e32 v9, v7
; GFX10-NEXT: v_min_u32_e32 v8, 32, v8
; GFX10-NEXT: v_min_u32_e32 v10, 32, v10
; GFX10-NEXT: v_min_u32_e32 v11, 32, v11
; GFX10-NEXT: v_min_u32_e32 v9, 32, v9
; GFX10-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX10-NEXT: v_lshlrev_b64 v[0:1], v10, v[0:1]
; GFX10-NEXT: v_lshlrev_b64 v[2:3], v11, v[2:3]
; GFX10-NEXT: v_lshlrev_b64 v[6:7], v9, v[6:7]
; GFX10-NEXT: v_sub_nc_u32_e32 v8, 32, v8
; GFX10-NEXT: v_sub_nc_u32_e32 v9, 32, v9
; GFX10-NEXT: v_min_u32_e32 v4, 1, v4
; GFX10-NEXT: v_min_u32_e32 v0, 1, v0
; GFX10-NEXT: v_min_u32_e32 v2, 1, v2
; GFX10-NEXT: v_min_u32_e32 v6, 1, v6
; GFX10-NEXT: v_or_b32_e32 v4, v5, v4
; GFX10-NEXT: v_or_b32_e32 v0, v1, v0
; GFX10-NEXT: v_or_b32_e32 v1, v3, v2
; GFX10-NEXT: v_sub_nc_u32_e32 v5, 32, v10
; GFX10-NEXT: v_sub_nc_u32_e32 v3, 32, v11
; GFX10-NEXT: v_cvt_f32_u32_e32 v2, v4
; GFX10-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX10-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX10-NEXT: v_or_b32_e32 v6, v7, v6
; GFX10-NEXT: v_ldexp_f32 v2, v2, v8
; GFX10-NEXT: v_ldexp_f32 v0, v0, v5
; GFX10-NEXT: v_ldexp_f32 v1, v1, v3
; GFX10-NEXT: v_cvt_f32_u32_e32 v4, v6
; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_ldexp_f32 v4, v4, v9
; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v9, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v8, v3, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uitofp_v4i64_to_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_clz_i32_u32_e32 v8, v5
; GFX11-NEXT: v_clz_i32_u32_e32 v10, v1
; GFX11-NEXT: v_clz_i32_u32_e32 v11, v3
; GFX11-NEXT: v_clz_i32_u32_e32 v9, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_u32_e32 v8, 32, v8
; GFX11-NEXT: v_min_u32_e32 v10, 32, v10
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_min_u32_e32 v11, 32, v11
; GFX11-NEXT: v_min_u32_e32 v9, 32, v9
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_lshlrev_b64 v[4:5], v8, v[4:5]
; GFX11-NEXT: v_lshlrev_b64 v[0:1], v10, v[0:1]
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_lshlrev_b64 v[2:3], v11, v[2:3]
; GFX11-NEXT: v_lshlrev_b64 v[6:7], v9, v[6:7]
; GFX11-NEXT: v_sub_nc_u32_e32 v8, 32, v8
; GFX11-NEXT: v_sub_nc_u32_e32 v9, 32, v9
; GFX11-NEXT: v_min_u32_e32 v4, 1, v4
; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
; GFX11-NEXT: v_min_u32_e32 v2, 1, v2
; GFX11-NEXT: v_min_u32_e32 v6, 1, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_or_b32_e32 v4, v5, v4
; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_or_b32_e32 v1, v3, v2
; GFX11-NEXT: v_sub_nc_u32_e32 v5, 32, v10
; GFX11-NEXT: v_sub_nc_u32_e32 v3, 32, v11
; GFX11-NEXT: v_cvt_f32_u32_e32 v2, v4
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: v_or_b32_e32 v6, v7, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_ldexp_f32 v2, v2, v8
; GFX11-NEXT: v_ldexp_f32 v0, v0, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_ldexp_f32 v1, v1, v3
; GFX11-NEXT: v_cvt_f32_u32_e32 v4, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_bfe_u32 v7, v0, 16, 1
; GFX11-NEXT: v_bfe_u32 v8, v1, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-NEXT: v_ldexp_f32 v4, v4, v9
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v7, v7, v0, 0x7fff
; GFX11-NEXT: v_add3_u32 v8, v8, v1, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v2, v3, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_bfe_u32 v6, v4, 16, 1
; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v4
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v9, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v1, v8, v3, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v1, v3, v2, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = uitofp <4 x i64> %x to <4 x bfloat>
ret <4 x bfloat> %op
}
define bfloat @v_select_bf16(i1 %cond, bfloat %a, bfloat %b) {
; GCN-LABEL: v_select_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, bfloat %a, bfloat %b
ret bfloat %op
}
define bfloat @v_select_fneg_lhs_bf16(i1 %cond, bfloat %a, bfloat %b) {
; GCN-LABEL: v_select_fneg_lhs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v1, -1.0, v1
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_fneg_lhs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, -1.0, v1
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_fneg_lhs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_xor_b32_e32 v1, 0x8000, v1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_fneg_lhs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_xor_b32_e32 v1, 0x8000, v1
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_fneg_lhs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_xor_b32_e32 v1, 0x8000, v1
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_select_fneg_lhs_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_xor_b16 v1.l, 0x8000, v1.l
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_select_fneg_lhs_bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_xor_b32_e32 v1, 0x8000, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%neg.a = fneg bfloat %a
%op = select i1 %cond, bfloat %neg.a, bfloat %b
ret bfloat %op
}
define bfloat @v_select_fneg_rhs_bf16(i1 %cond, bfloat %a, bfloat %b) {
; GCN-LABEL: v_select_fneg_rhs_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v2, -1.0, v2
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_fneg_rhs_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v2, -1.0, v2
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_fneg_rhs_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_xor_b32_e32 v2, 0x8000, v2
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_fneg_rhs_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_xor_b32_e32 v2, 0x8000, v2
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_fneg_rhs_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_xor_b32_e32 v2, 0x8000, v2
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_select_fneg_rhs_bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_xor_b16 v2.l, 0x8000, v2.l
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_select_fneg_rhs_bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_xor_b32_e32 v2, 0x8000, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%neg.b = fneg bfloat %b
%op = select i1 %cond, bfloat %a, bfloat %neg.b
ret bfloat %op
}
define <2 x bfloat> @v_select_v2bf16(i1 %cond, <2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_select_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GCN-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v2, v4, v2, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v1
; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_select_v2bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v3, 16, v2
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v4, 16, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_dual_cndmask_b32 v0, v2, v1 :: v_dual_cndmask_b32 v1, v3, v4
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_select_v2bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v1
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v0, v2, v1 :: v_dual_cndmask_b32 v1, v4, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <2 x bfloat> %a, <2 x bfloat> %b
ret <2 x bfloat> %op
}
define <2 x bfloat> @v_vselect_v2bf16(<2 x i1> %cond, <2 x bfloat> %a, <2 x bfloat> %b) {
; GCN-LABEL: v_vselect_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v1, 1, v1
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v5, v3, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v4, v2, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_vselect_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v1, 1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_cndmask_b32_e32 v1, v5, v3, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v4, v2, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_vselect_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_and_b32_e32 v1, 1, v1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_vselect_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_and_b32_e32 v1, 1, v1
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v2, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_vselect_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_and_b32_e32 v1, 1, v1
; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v2
; GFX10-NEXT: v_lshrrev_b32_e32 v5, 16, v3
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_vselect_v2bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v4, 16, v3
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v5, 16, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_dual_cndmask_b32 v0, v3, v2 :: v_dual_and_b32 v1, 1, v1
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_vselect_v2bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v2
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v5, 16, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v0, v3, v2 :: v_dual_and_b32 v1, 1, v1
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select <2 x i1> %cond, <2 x bfloat> %a, <2 x bfloat> %b
ret <2 x bfloat> %op
}
define amdgpu_ps i32 @s_select_bf16(bfloat inreg %a, bfloat inreg %b, i32 %c) {
; GCN-LABEL: s_select_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v1, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s1
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_select_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s0
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s1
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_select_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mov_b32_e32 v2, s0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_select_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_mov_b32_e32 v2, s0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_select_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v1, s0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, s1, v1, vcc_lo
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_select_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_mov_b32_e32 v1, s0
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, s1, v1, vcc_lo
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: ; return to shader part epilog
%cond = icmp eq i32 %c, 0
%op = select i1 %cond, bfloat %a, bfloat %b
%cast = bitcast bfloat %op to i16
%zext = zext i16 %cast to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %zext)
ret i32 %readlane
}
define amdgpu_ps i32 @s_select_v2bf16(<2 x bfloat> inreg %a, <2 x bfloat> inreg %b, i32 %c) {
; GCN-LABEL: s_select_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s3
; GCN-NEXT: v_mul_f32_e64 v3, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v4, 1.0, s2
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_select_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s3
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s0
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s2
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_select_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_lshr_b32 s2, s0, 16
; GFX8-NEXT: s_lshr_b32 s3, s1, 16
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: v_mov_b32_e32 v2, s2
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s1
; GFX8-NEXT: v_mov_b32_e32 v2, s0
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX8-NEXT: v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_select_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_lshr_b32 s2, s0, 16
; GFX9-NEXT: s_lshr_b32 s3, s1, 16
; GFX9-NEXT: v_mov_b32_e32 v1, s3
; GFX9-NEXT: v_mov_b32_e32 v2, s2
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_mov_b32_e32 v1, s1
; GFX9-NEXT: v_mov_b32_e32 v2, s0
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX9-NEXT: s_mov_b32 s0, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v0, v1, s0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_select_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_lshr_b32 s2, s0, 16
; GFX10-NEXT: v_mov_b32_e32 v2, s0
; GFX10-NEXT: v_mov_b32_e32 v1, s2
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: s_lshr_b32 s3, s1, 16
; GFX10-NEXT: v_cndmask_b32_e32 v0, s3, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, s1, v2, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11TRUE16-LABEL: s_select_v2bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_lshr_b32 s2, s1, 16
; GFX11TRUE16-NEXT: s_lshr_b32 s3, s0, 16
; GFX11TRUE16-NEXT: v_mov_b16_e32 v1.l, s2
; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, s3
; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.l, s1
; GFX11TRUE16-NEXT: v_mov_b16_e32 v4.l, s0
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_cndmask_b32 v0, v1, v2 :: v_dual_cndmask_b32 v1, v3, v4
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11TRUE16-NEXT: ; return to shader part epilog
;
; GFX11FAKE16-LABEL: s_select_v2bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_lshr_b32 s2, s0, 16
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11FAKE16-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s0
; GFX11FAKE16-NEXT: s_lshr_b32 s3, s1, 16
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, s3, v1, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, s1, v2, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v1, 0x5040100
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11FAKE16-NEXT: ; return to shader part epilog
%cond = icmp eq i32 %c, 0
%op = select i1 %cond, <2 x bfloat> %a, <2 x bfloat> %b
%cast = bitcast <2 x bfloat> %op to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %cast)
ret i32 %readlane
}
define amdgpu_ps i32 @s_vselect_v2bf16(<2 x bfloat> inreg %a, <2 x bfloat> inreg %b, <2 x i32> %c) {
; GCN-LABEL: s_vselect_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v3, 1.0, s2
; GCN-NEXT: v_mul_f32_e64 v4, 1.0, s1
; GCN-NEXT: v_mul_f32_e64 v5, 1.0, s3
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_vselect_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s1
; GFX7-NEXT: v_mul_f32_e64 v5, 1.0, s3
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s2
; GFX7-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_vselect_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_lshr_b32 s2, s0, 16
; GFX8-NEXT: s_lshr_b32 s3, s1, 16
; GFX8-NEXT: v_mov_b32_e32 v2, s3
; GFX8-NEXT: v_mov_b32_e32 v3, s2
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX8-NEXT: v_mov_b32_e32 v2, s1
; GFX8-NEXT: v_mov_b32_e32 v3, s0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_vselect_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_lshr_b32 s2, s0, 16
; GFX9-NEXT: s_lshr_b32 s3, s1, 16
; GFX9-NEXT: v_mov_b32_e32 v2, s3
; GFX9-NEXT: v_mov_b32_e32 v3, s2
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT: v_mov_b32_e32 v2, s1
; GFX9-NEXT: v_mov_b32_e32 v3, s0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v3, vcc
; GFX9-NEXT: s_mov_b32 s0, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s0
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_vselect_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_lshr_b32 s2, s0, 16
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX10-NEXT: v_mov_b32_e32 v2, s2
; GFX10-NEXT: v_mov_b32_e32 v3, s0
; GFX10-NEXT: s_lshr_b32 s0, s1, 16
; GFX10-NEXT: v_cndmask_b32_e32 v1, s0, v2, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, s1, v3, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11TRUE16-LABEL: s_vselect_v2bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_lshr_b32 s2, s1, 16
; GFX11TRUE16-NEXT: s_lshr_b32 s3, s0, 16
; GFX11TRUE16-NEXT: v_mov_b16_e32 v2.l, s2
; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.l, s3
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX11TRUE16-NEXT: v_mov_b16_e32 v4.l, s1
; GFX11TRUE16-NEXT: v_mov_b16_e32 v5.l, s0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v3, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v4, v5, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11TRUE16-NEXT: ; return to shader part epilog
;
; GFX11FAKE16-LABEL: s_vselect_v2bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_lshr_b32 s2, s0, 16
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX11FAKE16-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s0
; GFX11FAKE16-NEXT: s_lshr_b32 s0, s1, 16
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, s0, v2, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, s1, v3, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11FAKE16-NEXT: ; return to shader part epilog
%cond = icmp eq <2 x i32> %c, zeroinitializer
%op = select <2 x i1> %cond, <2 x bfloat> %a, <2 x bfloat> %b
%cast = bitcast <2 x bfloat> %op to i32
%readlane = call i32 @llvm.amdgcn.readfirstlane(i32 %cast)
ret i32 %readlane
}
define <3 x bfloat> @v_select_v3bf16(i1 %cond, <3 x bfloat> %a, <3 x bfloat> %b) {
; GCN-LABEL: v_select_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GCN-NEXT: v_alignbit_b32 v2, v5, v4, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_cndmask_b32_e32 v0, v6, v3, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_alignbit_b32 v2, v2, v4, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v0, v6, v3, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v0
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v3bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v3, v1 :: v_dual_cndmask_b32 v1, v4, v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <3 x bfloat> %a, <3 x bfloat> %b
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_select_v4bf16(i1 %cond, <4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_select_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GCN-NEXT: v_alignbit_b32 v2, v6, v5, 16
; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GCN-NEXT: v_alignbit_b32 v4, v8, v7, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v8
; GFX7-NEXT: v_alignbit_b32 v2, v2, v5, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v2, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v3, v1 :: v_dual_cndmask_b32 v1, v4, v2
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <4 x bfloat> %a, <4 x bfloat> %b
ret <4 x bfloat> %op
}
define <6 x bfloat> @v_select_v6bf16(i1 %cond, <6 x bfloat> %a, <6 x bfloat> %b) {
; GCN-LABEL: v_select_v6bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GCN-NEXT: v_alignbit_b32 v2, v8, v7, 16
; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GCN-NEXT: v_alignbit_b32 v4, v10, v9, 16
; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
; GCN-NEXT: v_alignbit_b32 v6, v12, v11, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GCN-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v6bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v10
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_alignbit_b32 v2, v2, v7, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v9
; GFX7-NEXT: v_alignbit_b32 v5, v6, v5, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v12
; GFX7-NEXT: v_alignbit_b32 v4, v4, v7, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v11
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_alignbit_b32 v6, v6, v7, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v6bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v4, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v2, v6, v3, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v6bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v4, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v2, v6, v3, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v6bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v5, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v6, v3, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v6bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v4, v1 :: v_dual_cndmask_b32 v1, v5, v2
; GFX11-NEXT: v_cndmask_b32_e32 v2, v6, v3, vcc_lo
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <6 x bfloat> %a, <6 x bfloat> %b
ret <6 x bfloat> %op
}
define <8 x bfloat> @v_select_v8bf16(i1 %cond, <8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_select_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GCN-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GCN-NEXT: v_alignbit_b32 v2, v10, v9, 16
; GCN-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GCN-NEXT: v_alignbit_b32 v4, v12, v11, 16
; GCN-NEXT: v_alignbit_b32 v5, v6, v5, 16
; GCN-NEXT: v_alignbit_b32 v6, v14, v13, 16
; GCN-NEXT: v_alignbit_b32 v7, v8, v7, 16
; GCN-NEXT: v_alignbit_b32 v8, v16, v15, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v7, v8, v7, vcc
; GCN-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GCN-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v10
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v12
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_alignbit_b32 v2, v2, v9, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v11
; GFX7-NEXT: v_alignbit_b32 v5, v6, v5, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_alignbit_b32 v4, v4, v9, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v13
; GFX7-NEXT: v_alignbit_b32 v7, v8, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v16
; GFX7-NEXT: v_alignbit_b32 v6, v6, v9, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v15
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_alignbit_b32 v8, v8, v9, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v7, v8, v7, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v5, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v6, v2, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v2, v7, v3, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v3, v8, v4, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v5, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v6, v2, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v2, v7, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v8, v4, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v6, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v3, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v3, v8, v4, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v8bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v5, v1 :: v_dual_cndmask_b32 v1, v6, v2
; GFX11-NEXT: v_dual_cndmask_b32 v2, v7, v3 :: v_dual_cndmask_b32 v3, v8, v4
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <8 x bfloat> %a, <8 x bfloat> %b
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_select_v16bf16(i1 %cond, <16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_select_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v17
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v19
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v22
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v21
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_alignbit_b32 v5, v5, v6, 16
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v24
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v23
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v26
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v25
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v28
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v27
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v30
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v29
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
; GCN-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:4
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: v_alignbit_b32 v11, v12, v11, 16
; GCN-NEXT: buffer_load_dword v12, off, s[0:3], s32
; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GCN-NEXT: v_alignbit_b32 v17, v17, v18, 16
; GCN-NEXT: v_alignbit_b32 v13, v14, v13, 16
; GCN-NEXT: v_alignbit_b32 v14, v19, v20, 16
; GCN-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GCN-NEXT: v_cndmask_b32_e32 v13, v14, v13, vcc
; GCN-NEXT: v_cndmask_b32_e32 v11, v17, v11, vcc
; GCN-NEXT: v_cndmask_b32_e32 v9, v9, v8, vcc
; GCN-NEXT: v_cndmask_b32_e32 v7, v7, v6, vcc
; GCN-NEXT: v_cndmask_b32_e32 v5, v5, v4, vcc
; GCN-NEXT: v_cndmask_b32_e32 v3, v3, v2, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v10
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v12
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: v_alignbit_b32 v14, v14, v16, 16
; GCN-NEXT: v_cndmask_b32_e32 v15, v14, v15, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v18
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_alignbit_b32 v3, v4, v3, 16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v20
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_alignbit_b32 v2, v2, v17, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v19
; GFX7-NEXT: v_alignbit_b32 v5, v6, v5, 16
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v22
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_alignbit_b32 v4, v4, v17, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v21
; GFX7-NEXT: v_alignbit_b32 v7, v8, v7, 16
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v24
; GFX7-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_alignbit_b32 v6, v6, v17, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v23
; GFX7-NEXT: v_alignbit_b32 v9, v10, v9, 16
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v26
; GFX7-NEXT: v_alignbit_b32 v8, v8, v17, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v25
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_alignbit_b32 v10, v10, v17, 16
; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v28
; GFX7-NEXT: v_alignbit_b32 v11, v12, v11, 16
; GFX7-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v27
; GFX7-NEXT: v_alignbit_b32 v17, v17, v18, 16
; GFX7-NEXT: buffer_load_dword v18, off, s[0:3], s32
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_alignbit_b32 v13, v14, v13, 16
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v30
; GFX7-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v29
; GFX7-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_alignbit_b32 v14, v14, v19, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v13, v14, v13, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v11, v17, v11, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v9, v10, v9, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v7, v8, v7, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v5, v6, v5, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v3, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v18
; GFX7-NEXT: v_alignbit_b32 v12, v12, v16, 16
; GFX7-NEXT: v_cndmask_b32_e32 v15, v12, v15, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v9, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v10, v2, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v2, v11, v3, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v3, v12, v4, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v4, v13, v5, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v5, v14, v6, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v6, v15, v7, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v7, v16, v8, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v9, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v10, v2, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v2, v11, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v12, v4, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v4, v13, v5, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v5, v14, v6, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v6, v15, v7, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v7, v16, v8, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v9, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v10, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v11, v3, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v3, v12, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v4, v13, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v5, v14, v6, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v6, v15, v7, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v7, v16, v8, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v16bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v9, v1 :: v_dual_cndmask_b32 v1, v10, v2
; GFX11-NEXT: v_dual_cndmask_b32 v2, v11, v3 :: v_dual_cndmask_b32 v3, v12, v4
; GFX11-NEXT: v_dual_cndmask_b32 v4, v13, v5 :: v_dual_cndmask_b32 v5, v14, v6
; GFX11-NEXT: v_dual_cndmask_b32 v6, v15, v7 :: v_dual_cndmask_b32 v7, v16, v8
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <16 x bfloat> %a, <16 x bfloat> %b
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_select_v32bf16(i1 %cond, <32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_select_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_alignbit_b32 v0, v0, v1, 16
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v3
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v5
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_alignbit_b32 v2, v2, v3, 16
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v7
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_alignbit_b32 v3, v3, v4, 16
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v9
; GCN-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GCN-NEXT: v_alignbit_b32 v4, v4, v5, 16
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v11
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_alignbit_b32 v5, v5, v6, 16
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v13
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_alignbit_b32 v6, v6, v7, 16
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v16
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v15
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_alignbit_b32 v7, v7, v8, 16
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v18
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v17
; GCN-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GCN-NEXT: v_alignbit_b32 v8, v8, v9, 16
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v19
; GCN-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GCN-NEXT: v_alignbit_b32 v9, v9, v10, 16
; GCN-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:12
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v22
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v21
; GCN-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GCN-NEXT: v_alignbit_b32 v10, v10, v11, 16
; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:8
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v24
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v23
; GCN-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GCN-NEXT: v_alignbit_b32 v11, v11, v12, 16
; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:20
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v26
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v25
; GCN-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GCN-NEXT: v_alignbit_b32 v12, v12, v13, 16
; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:16
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v28
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v27
; GCN-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GCN-NEXT: v_alignbit_b32 v13, v13, v14, 16
; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:28
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v30
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v29
; GCN-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GCN-NEXT: v_alignbit_b32 v14, v14, v20, 16
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:24
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: s_waitcnt vmcnt(4)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GCN-NEXT: v_alignbit_b32 v15, v15, v16, 16
; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:36
; GCN-NEXT: s_waitcnt vmcnt(4)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v17
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v18
; GCN-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GCN-NEXT: v_alignbit_b32 v16, v16, v17, 16
; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:32
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v19
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v20
; GCN-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GCN-NEXT: v_alignbit_b32 v17, v17, v19, 16
; GCN-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:44
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v21
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:40
; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GCN-NEXT: v_alignbit_b32 v18, v20, v18, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v21
; GCN-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:52
; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:48
; GCN-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GCN-NEXT: v_alignbit_b32 v19, v19, v20, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v21
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
; GCN-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:60
; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:56
; GCN-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GCN-NEXT: v_alignbit_b32 v20, v20, v21, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v22
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v23
; GCN-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:68
; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:64
; GCN-NEXT: v_lshrrev_b32_e32 v21, 16, v21
; GCN-NEXT: v_alignbit_b32 v21, v21, v22, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v23
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v24
; GCN-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:76
; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:72
; GCN-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GCN-NEXT: v_alignbit_b32 v22, v22, v23, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v24
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v25
; GCN-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:84
; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:80
; GCN-NEXT: v_lshrrev_b32_e32 v23, 16, v23
; GCN-NEXT: v_alignbit_b32 v23, v23, v24, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v24, 1.0, v25
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v26
; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:92
; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:88
; GCN-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GCN-NEXT: v_alignbit_b32 v24, v24, v25, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v25, 1.0, v26
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v27
; GCN-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:100
; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:96
; GCN-NEXT: v_lshrrev_b32_e32 v25, 16, v25
; GCN-NEXT: v_alignbit_b32 v25, v25, v26, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v26, 1.0, v27
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v28
; GCN-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:108
; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:104
; GCN-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GCN-NEXT: v_alignbit_b32 v26, v26, v27, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v27, 1.0, v28
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v29
; GCN-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:116
; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:112
; GCN-NEXT: v_lshrrev_b32_e32 v27, 16, v27
; GCN-NEXT: v_alignbit_b32 v27, v27, v28, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v28, 1.0, v29
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v30
; GCN-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:124
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:120
; GCN-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GCN-NEXT: v_alignbit_b32 v28, v28, v29, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v29, 1.0, v30
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v31
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GCN-NEXT: v_lshrrev_b32_e32 v29, 16, v29
; GCN-NEXT: v_alignbit_b32 v29, v29, v30, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v30, 1.0, v31
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v32
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:128
; GCN-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GCN-NEXT: v_alignbit_b32 v30, v30, v31, 16
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v32
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v33
; GCN-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GCN-NEXT: v_alignbit_b32 v31, v31, v32, 16
; GCN-NEXT: v_cndmask_b32_e32 v31, v31, v30, vcc
; GCN-NEXT: v_cndmask_b32_e32 v29, v29, v14, vcc
; GCN-NEXT: v_cndmask_b32_e32 v28, v28, v13, vcc
; GCN-NEXT: v_cndmask_b32_e32 v27, v27, v12, vcc
; GCN-NEXT: v_cndmask_b32_e32 v26, v26, v11, vcc
; GCN-NEXT: v_cndmask_b32_e32 v25, v25, v10, vcc
; GCN-NEXT: v_cndmask_b32_e32 v24, v24, v9, vcc
; GCN-NEXT: v_cndmask_b32_e32 v23, v23, v8, vcc
; GCN-NEXT: v_cndmask_b32_e32 v22, v22, v7, vcc
; GCN-NEXT: v_cndmask_b32_e32 v13, v21, v6, vcc
; GCN-NEXT: v_cndmask_b32_e32 v11, v20, v5, vcc
; GCN-NEXT: v_cndmask_b32_e32 v9, v19, v4, vcc
; GCN-NEXT: v_cndmask_b32_e32 v7, v18, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v5, v17, v2, vcc
; GCN-NEXT: v_cndmask_b32_e32 v3, v16, v1, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v15, v0, vcc
; GCN-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_lshlrev_b32_e32 v6, 16, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_lshlrev_b32_e32 v8, 16, v9
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_lshlrev_b32_e32 v12, 16, v13
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_lshlrev_b32_e32 v14, 16, v22
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v22
; GCN-NEXT: v_lshlrev_b32_e32 v16, 16, v23
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v23
; GCN-NEXT: v_lshlrev_b32_e32 v18, 16, v24
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v24
; GCN-NEXT: v_lshlrev_b32_e32 v20, 16, v25
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v25
; GCN-NEXT: v_lshlrev_b32_e32 v22, 16, v26
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v26
; GCN-NEXT: v_lshlrev_b32_e32 v24, 16, v27
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v27
; GCN-NEXT: v_lshlrev_b32_e32 v26, 16, v28
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v28
; GCN-NEXT: v_lshlrev_b32_e32 v28, 16, v29
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_select_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_alignbit_b32 v1, v2, v1, 16
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v4
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_alignbit_b32 v2, v2, v3, 16
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v5
; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v7
; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v10
; GFX7-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_alignbit_b32 v5, v5, v6, 16
; GFX7-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:12
; GFX7-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:16
; GFX7-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:24
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_alignbit_b32 v17, v18, v17, 16
; GFX7-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:76
; GFX7-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GFX7-NEXT: v_mul_f32_e32 v27, 1.0, v27
; GFX7-NEXT: v_alignbit_b32 v13, v14, v13, 16
; GFX7-NEXT: v_alignbit_b32 v27, v28, v27, 16
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_alignbit_b32 v11, v12, v11, 16
; GFX7-NEXT: v_alignbit_b32 v23, v24, v23, 16
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_alignbit_b32 v15, v16, v15, 16
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_alignbit_b32 v19, v20, v19, 16
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_alignbit_b32 v21, v22, v21, 16
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_alignbit_b32 v25, v26, v25, 16
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GFX7-NEXT: v_mul_f32_e32 v29, 1.0, v29
; GFX7-NEXT: v_alignbit_b32 v29, v30, v29, 16
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:32
; GFX7-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
; GFX7-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:116
; GFX7-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:52
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100
; GFX7-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:68
; GFX7-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:84
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92
; GFX7-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:108
; GFX7-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:124
; GFX7-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:128
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32
; GFX7-NEXT: s_waitcnt vmcnt(14)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(13)
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: s_waitcnt vmcnt(12)
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_alignbit_b32 v6, v6, v7, 16
; GFX7-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX7-NEXT: v_cndmask_b32_e32 v1, v6, v1, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(12)
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: s_waitcnt vmcnt(11)
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX7-NEXT: s_waitcnt vmcnt(9)
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX7-NEXT: s_waitcnt vmcnt(7)
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX7-NEXT: s_waitcnt vmcnt(6)
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GFX7-NEXT: s_waitcnt vmcnt(5)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v26, 1.0, v26
; GFX7-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GFX7-NEXT: v_mul_f32_e32 v28, 1.0, v28
; GFX7-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_mul_f32_e32 v30, 1.0, v30
; GFX7-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_mul_f32_e32 v33, 1.0, v33
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX7-NEXT: v_alignbit_b32 v7, v7, v8, 16
; GFX7-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:28
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX7-NEXT: v_alignbit_b32 v8, v8, v9, 16
; GFX7-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:36
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX7-NEXT: v_alignbit_b32 v9, v9, v10, 16
; GFX7-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_cndmask_b32_e32 v9, v9, v4, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v6, 16, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX7-NEXT: v_alignbit_b32 v10, v10, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_cndmask_b32_e32 v10, v10, v5, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v5, v8, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v3, v7, v2, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v9
; GFX7-NEXT: v_lshlrev_b32_e32 v8, 16, v10
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v12, v12, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_cndmask_b32_e32 v11, v12, v11, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v10, 16, v11
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v14, v14, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_cndmask_b32_e32 v13, v14, v13, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v12, 16, v13
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v16, v16, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72
; GFX7-NEXT: v_cndmask_b32_e32 v15, v16, v15, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v14, 16, v15
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v18, v18, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:80
; GFX7-NEXT: v_cndmask_b32_e32 v17, v18, v17, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v16, 16, v17
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v20, v20, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
; GFX7-NEXT: v_cndmask_b32_e32 v19, v20, v19, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v18, 16, v19
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v22, v22, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
; GFX7-NEXT: v_cndmask_b32_e32 v21, v22, v21, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v20, 16, v21
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v24, v24, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:104
; GFX7-NEXT: v_cndmask_b32_e32 v23, v24, v23, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v22, 16, v23
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v26, v26, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:112
; GFX7-NEXT: v_cndmask_b32_e32 v25, v26, v25, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v24, 16, v25
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v28, v28, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:120
; GFX7-NEXT: v_cndmask_b32_e32 v27, v28, v27, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v26, 16, v27
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_alignbit_b32 v30, v30, v31, 16
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX7-NEXT: v_cndmask_b32_e32 v29, v30, v29, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v28, 16, v29
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GFX7-NEXT: v_alignbit_b32 v31, v31, v32, 16
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_lshrrev_b32_e32 v32, 16, v32
; GFX7-NEXT: v_alignbit_b32 v32, v32, v33, 16
; GFX7-NEXT: v_cndmask_b32_e32 v31, v32, v31, vcc
; GFX7-NEXT: v_lshlrev_b32_e32 v30, 16, v31
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_select_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v17, v1, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v1, v18, v2, vcc
; GFX8-NEXT: buffer_load_dword v17, off, s[0:3], s32
; GFX8-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:4
; GFX8-NEXT: v_cndmask_b32_e32 v2, v19, v3, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v3, v20, v4, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v4, v21, v5, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v5, v22, v6, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v6, v23, v7, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v7, v24, v8, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v8, v25, v9, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v9, v26, v10, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v10, v27, v11, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v11, v28, v12, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v12, v29, v13, vcc
; GFX8-NEXT: v_cndmask_b32_e32 v13, v30, v14, vcc
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_cndmask_b32_e32 v14, v17, v15, vcc
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_cndmask_b32_e32 v15, v18, v16, vcc
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_select_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v17, v1, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v1, v18, v2, vcc
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:4
; GFX9-NEXT: v_cndmask_b32_e32 v2, v19, v3, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v3, v20, v4, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v4, v21, v5, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v5, v22, v6, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v6, v23, v7, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v7, v24, v8, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v8, v25, v9, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v9, v26, v10, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v10, v27, v11, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v11, v28, v12, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v12, v29, v13, vcc
; GFX9-NEXT: v_cndmask_b32_e32 v13, v30, v14, vcc
; GFX9-NEXT: s_waitcnt vmcnt(1)
; GFX9-NEXT: v_cndmask_b32_e32 v14, v17, v15, vcc
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_cndmask_b32_e32 v15, v18, v16, vcc
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_select_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v17, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, v18, v2, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v2, v19, v3, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v3, v20, v4, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v4, v21, v5, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v5, v22, v6, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v6, v23, v7, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v7, v24, v8, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v8, v25, v9, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v9, v26, v10, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v10, v27, v11, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v11, v28, v12, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v12, v29, v13, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v13, v30, v14, vcc_lo
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: v_cndmask_b32_e32 v14, v31, v15, vcc_lo
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_cndmask_b32_e32 v15, v32, v16, vcc_lo
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_select_v32bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: s_clause 0x1
; GFX11-NEXT: scratch_load_b32 v31, off, s32
; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11-NEXT: v_dual_cndmask_b32 v0, v17, v1 :: v_dual_cndmask_b32 v1, v18, v2
; GFX11-NEXT: v_dual_cndmask_b32 v2, v19, v3 :: v_dual_cndmask_b32 v3, v20, v4
; GFX11-NEXT: v_dual_cndmask_b32 v4, v21, v5 :: v_dual_cndmask_b32 v5, v22, v6
; GFX11-NEXT: v_dual_cndmask_b32 v6, v23, v7 :: v_dual_cndmask_b32 v7, v24, v8
; GFX11-NEXT: v_dual_cndmask_b32 v8, v25, v9 :: v_dual_cndmask_b32 v9, v26, v10
; GFX11-NEXT: v_dual_cndmask_b32 v10, v27, v11 :: v_dual_cndmask_b32 v11, v28, v12
; GFX11-NEXT: v_dual_cndmask_b32 v12, v29, v13 :: v_dual_cndmask_b32 v13, v30, v14
; GFX11-NEXT: s_waitcnt vmcnt(0)
; GFX11-NEXT: v_dual_cndmask_b32 v14, v31, v15 :: v_dual_cndmask_b32 v15, v32, v16
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = select i1 %cond, <32 x bfloat> %a, <32 x bfloat> %b
ret <32 x bfloat> %op
}
define amdgpu_ps <2 x i32> @s_select_v3bf16(<3 x bfloat> inreg %a, <3 x bfloat> inreg %b, i32 %c) {
; GCN-LABEL: s_select_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v3, 1.0, s4
; GCN-NEXT: v_mul_f32_e64 v4, 1.0, s3
; GCN-NEXT: v_mul_f32_e64 v5, 1.0, s2
; GCN-NEXT: v_mul_f32_e64 v6, 1.0, s5
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GCN-NEXT: v_alignbit_b32 v2, v3, v4, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v6, v5, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_readfirstlane_b32 s0, v1
; GCN-NEXT: v_readfirstlane_b32 s1, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_select_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GFX7-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s4
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s3
; GFX7-NEXT: v_alignbit_b32 v2, v2, v3, 16
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s2
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s5
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_readfirstlane_b32 s0, v1
; GFX7-NEXT: v_readfirstlane_b32 s1, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_select_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_mov_b32_e32 v1, s2
; GFX8-NEXT: v_mov_b32_e32 v2, s0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: v_mov_b32_e32 v2, s1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: v_readfirstlane_b32 s1, v1
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_select_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_mov_b32_e32 v1, s2
; GFX9-NEXT: v_mov_b32_e32 v2, s0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_mov_b32_e32 v1, s3
; GFX9-NEXT: v_mov_b32_e32 v2, s1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: v_readfirstlane_b32 s1, v1
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_select_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v1, s0
; GFX10-NEXT: v_mov_b32_e32 v2, s1
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, s2, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, s3, v2, vcc_lo
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX10-NEXT: v_readfirstlane_b32 s1, v1
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_select_v3bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_dual_mov_b32 v1, s0 :: v_dual_mov_b32 v2, s1
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, s2, v1, vcc_lo
; GFX11-NEXT: v_cndmask_b32_e32 v1, s3, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_readfirstlane_b32 s0, v0
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_readfirstlane_b32 s1, v1
; GFX11-NEXT: ; return to shader part epilog
%cond = icmp eq i32 %c, 0
%op = select i1 %cond, <3 x bfloat> %a, <3 x bfloat> %b
%cast = bitcast <3 x bfloat> %op to i48
%elt0 = trunc i48 %cast to i32
%elt1.hi = lshr i48 %cast, 32
%elt1 = trunc i48 %elt1.hi to i32
%readlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt0)
%readlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt1)
%bv.0 = insertelement <2 x i32> poison, i32 %readlane0, i32 0
%bv.1 = insertelement <2 x i32> %bv.0, i32 %readlane1, i32 1
ret <2 x i32> %bv.1
}
define amdgpu_ps <2 x i32> @s_select_v4bf16(<4 x bfloat> inreg %a, <4 x bfloat> inreg %b, i32 %c) {
; GCN-LABEL: s_select_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GCN-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v3, 1.0, s5
; GCN-NEXT: v_mul_f32_e64 v4, 1.0, s4
; GCN-NEXT: v_mul_f32_e64 v5, 1.0, s3
; GCN-NEXT: v_mul_f32_e64 v6, 1.0, s2
; GCN-NEXT: v_mul_f32_e64 v7, 1.0, s7
; GCN-NEXT: v_mul_f32_e64 v8, 1.0, s6
; GCN-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GCN-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GCN-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GCN-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GCN-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GCN-NEXT: v_alignbit_b32 v2, v3, v4, 16
; GCN-NEXT: v_alignbit_b32 v3, v5, v6, 16
; GCN-NEXT: v_alignbit_b32 v4, v7, v8, 16
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GCN-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GCN-NEXT: v_readfirstlane_b32 s0, v1
; GCN-NEXT: v_readfirstlane_b32 s1, v0
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_select_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v1, 1.0, s1
; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s0
; GFX7-NEXT: v_alignbit_b32 v1, v1, v2, 16
; GFX7-NEXT: v_mul_f32_e64 v2, 1.0, s5
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s4
; GFX7-NEXT: v_alignbit_b32 v2, v2, v3, 16
; GFX7-NEXT: v_mul_f32_e64 v3, 1.0, s3
; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s2
; GFX7-NEXT: v_alignbit_b32 v3, v3, v4, 16
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s7
; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT: v_mul_f32_e64 v5, 1.0, s6
; GFX7-NEXT: v_alignbit_b32 v4, v4, v5, 16
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v4, v3, vcc
; GFX7-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
; GFX7-NEXT: v_readfirstlane_b32 s0, v1
; GFX7-NEXT: v_readfirstlane_b32 s1, v0
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_select_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: v_mov_b32_e32 v1, s3
; GFX8-NEXT: v_mov_b32_e32 v2, s1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_mov_b32_e32 v1, s2
; GFX8-NEXT: v_mov_b32_e32 v2, s0
; GFX8-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX8-NEXT: v_readfirstlane_b32 s0, v1
; GFX8-NEXT: v_readfirstlane_b32 s1, v0
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_select_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: v_mov_b32_e32 v1, s3
; GFX9-NEXT: v_mov_b32_e32 v2, s1
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_mov_b32_e32 v1, s2
; GFX9-NEXT: v_mov_b32_e32 v2, s0
; GFX9-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc
; GFX9-NEXT: v_readfirstlane_b32 s0, v1
; GFX9-NEXT: v_readfirstlane_b32 s1, v0
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_select_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: v_mov_b32_e32 v1, s1
; GFX10-NEXT: v_mov_b32_e32 v2, s0
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, s3, v1, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e32 v1, s2, v2, vcc_lo
; GFX10-NEXT: v_readfirstlane_b32 s1, v0
; GFX10-NEXT: v_readfirstlane_b32 s0, v1
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_select_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: v_dual_mov_b32 v1, s1 :: v_dual_mov_b32 v2, s0
; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v0, s3, v1, vcc_lo
; GFX11-NEXT: v_cndmask_b32_e32 v1, s2, v2, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_readfirstlane_b32 s1, v0
; GFX11-NEXT: v_readfirstlane_b32 s0, v1
; GFX11-NEXT: ; return to shader part epilog
%cond = icmp eq i32 %c, 0
%op = select i1 %cond, <4 x bfloat> %a, <4 x bfloat> %b
%cast = bitcast <4 x bfloat> %op to <2 x i32>
%elt0 = extractelement <2 x i32> %cast, i32 0
%elt1 = extractelement <2 x i32> %cast, i32 1
%readlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt0)
%readlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt1)
%bv.0 = insertelement <2 x i32> poison, i32 %readlane0, i32 0
%bv.1 = insertelement <2 x i32> %bv.0, i32 %readlane1, i32 1
ret <2 x i32> %bv.1
}
define amdgpu_ps <2 x i32> @s_vselect_v4bf16(<4 x bfloat> inreg %a, <4 x bfloat> inreg %b, <4 x i32> %c) {
; GCN-LABEL: s_vselect_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: v_mul_f32_e64 v4, 1.0, s0
; GCN-NEXT: v_mul_f32_e64 v5, 1.0, s4
; GCN-NEXT: v_mul_f32_e64 v6, 1.0, s1
; GCN-NEXT: v_mul_f32_e64 v7, 1.0, s5
; GCN-NEXT: v_mul_f32_e64 v8, 1.0, s2
; GCN-NEXT: v_mul_f32_e64 v9, 1.0, s6
; GCN-NEXT: v_mul_f32_e64 v10, 1.0, s3
; GCN-NEXT: v_mul_f32_e64 v11, 1.0, s7
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
; GCN-NEXT: v_cndmask_b32_e32 v3, v11, v10, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GCN-NEXT: v_cndmask_b32_e32 v2, v9, v8, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v7, v6, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v5, v4, vcc
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_or_b32_e32 v2, v2, v3
; GCN-NEXT: v_or_b32_e32 v0, v0, v1
; GCN-NEXT: v_readfirstlane_b32 s0, v0
; GCN-NEXT: v_readfirstlane_b32 s1, v2
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_vselect_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: v_mul_f32_e64 v10, 1.0, s3
; GFX7-NEXT: v_mul_f32_e64 v11, 1.0, s7
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
; GFX7-NEXT: v_mul_f32_e64 v8, 1.0, s2
; GFX7-NEXT: v_mul_f32_e64 v9, 1.0, s6
; GFX7-NEXT: v_cndmask_b32_e32 v3, v11, v10, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GFX7-NEXT: v_mul_f32_e64 v6, 1.0, s1
; GFX7-NEXT: v_mul_f32_e64 v7, 1.0, s5
; GFX7-NEXT: v_cndmask_b32_e32 v2, v9, v8, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX7-NEXT: v_mul_f32_e64 v4, 1.0, s0
; GFX7-NEXT: v_mul_f32_e64 v5, 1.0, s4
; GFX7-NEXT: v_cndmask_b32_e32 v1, v7, v6, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v5, v4, vcc
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_or_b32_e32 v2, v2, v3
; GFX7-NEXT: v_or_b32_e32 v0, v0, v1
; GFX7-NEXT: v_readfirstlane_b32 s0, v0
; GFX7-NEXT: v_readfirstlane_b32 s1, v2
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_vselect_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_lshr_b32 s4, s1, 16
; GFX8-NEXT: s_lshr_b32 s5, s3, 16
; GFX8-NEXT: v_mov_b32_e32 v4, s5
; GFX8-NEXT: v_mov_b32_e32 v5, s4
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_mov_b32_e32 v4, s3
; GFX8-NEXT: v_mov_b32_e32 v5, s1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
; GFX8-NEXT: s_lshr_b32 s1, s0, 16
; GFX8-NEXT: s_lshr_b32 s3, s2, 16
; GFX8-NEXT: v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_mov_b32_e32 v3, s3
; GFX8-NEXT: v_mov_b32_e32 v4, s1
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX8-NEXT: v_mov_b32_e32 v3, s2
; GFX8-NEXT: v_mov_b32_e32 v4, s0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_readfirstlane_b32 s0, v0
; GFX8-NEXT: v_readfirstlane_b32 s1, v2
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_vselect_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_lshr_b32 s4, s1, 16
; GFX9-NEXT: s_lshr_b32 s5, s3, 16
; GFX9-NEXT: v_mov_b32_e32 v4, s5
; GFX9-NEXT: v_mov_b32_e32 v5, s4
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_mov_b32_e32 v4, s3
; GFX9-NEXT: v_mov_b32_e32 v5, s1
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2
; GFX9-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc
; GFX9-NEXT: s_mov_b32 s1, 0x5040100
; GFX9-NEXT: s_lshr_b32 s3, s0, 16
; GFX9-NEXT: s_lshr_b32 s4, s2, 16
; GFX9-NEXT: v_perm_b32 v2, v3, v2, s1
; GFX9-NEXT: v_mov_b32_e32 v3, s4
; GFX9-NEXT: v_mov_b32_e32 v4, s3
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT: v_mov_b32_e32 v3, s2
; GFX9-NEXT: v_mov_b32_e32 v4, s0
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v3, v4, vcc
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s1
; GFX9-NEXT: v_readfirstlane_b32 s0, v0
; GFX9-NEXT: v_readfirstlane_b32 s1, v2
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_vselect_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_lshr_b32 s4, s1, 16
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX10-NEXT: v_mov_b32_e32 v4, s4
; GFX10-NEXT: s_lshr_b32 s4, s3, 16
; GFX10-NEXT: s_lshr_b32 s5, s0, 16
; GFX10-NEXT: v_mov_b32_e32 v6, s0
; GFX10-NEXT: s_lshr_b32 s0, s2, 16
; GFX10-NEXT: v_cndmask_b32_e32 v3, s4, v4, vcc_lo
; GFX10-NEXT: v_mov_b32_e32 v4, s5
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX10-NEXT: v_mov_b32_e32 v5, s1
; GFX10-NEXT: v_cndmask_b32_e32 v1, s0, v4, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, s2, v6, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v2, s3, v5, vcc_lo
; GFX10-NEXT: v_readfirstlane_b32 s0, v0
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX10-NEXT: v_readfirstlane_b32 s1, v1
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11TRUE16-LABEL: s_vselect_v4bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_lshr_b32 s4, s3, 16
; GFX11TRUE16-NEXT: s_lshr_b32 s5, s1, 16
; GFX11TRUE16-NEXT: v_mov_b16_e32 v4.l, s4
; GFX11TRUE16-NEXT: v_mov_b16_e32 v5.l, s5
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX11TRUE16-NEXT: s_lshr_b32 s4, s0, 16
; GFX11TRUE16-NEXT: v_mov_b16_e32 v7.l, s2
; GFX11TRUE16-NEXT: v_mov_b16_e32 v8.l, s0
; GFX11TRUE16-NEXT: v_mov_b16_e32 v6.l, s4
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc_lo
; GFX11TRUE16-NEXT: v_mov_b16_e32 v4.l, s3
; GFX11TRUE16-NEXT: s_lshr_b32 s3, s2, 16
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX11TRUE16-NEXT: v_mov_b16_e32 v5.l, s3
; GFX11TRUE16-NEXT: v_mov_b16_e32 v9.l, s1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v5, v6, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v4, v9, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11TRUE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_readfirstlane_b32 s1, v1
; GFX11TRUE16-NEXT: ; return to shader part epilog
;
; GFX11FAKE16-LABEL: s_vselect_v4bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_lshr_b32 s4, s1, 16
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v3
; GFX11FAKE16-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s1
; GFX11FAKE16-NEXT: s_lshr_b32 s4, s3, 16
; GFX11FAKE16-NEXT: s_lshr_b32 s5, s0, 16
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, s4, v4, vcc_lo
; GFX11FAKE16-NEXT: v_mov_b32_e32 v4, s5
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v1
; GFX11FAKE16-NEXT: v_mov_b32_e32 v6, s0
; GFX11FAKE16-NEXT: s_lshr_b32 s0, s2, 16
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instid1(SALU_CYCLE_1)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, s0, v4, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, s2, v6, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v2
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, s3, v5, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_readfirstlane_b32 s0, v0
; GFX11FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_readfirstlane_b32 s1, v1
; GFX11FAKE16-NEXT: ; return to shader part epilog
%cond = icmp eq <4 x i32> %c, zeroinitializer
%op = select <4 x i1> %cond, <4 x bfloat> %a, <4 x bfloat> %b
%cast = bitcast <4 x bfloat> %op to <2 x i32>
%elt0 = extractelement <2 x i32> %cast, i32 0
%elt1 = extractelement <2 x i32> %cast, i32 1
%readlane0 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt0)
%readlane1 = call i32 @llvm.amdgcn.readfirstlane(i32 %elt1)
%bv.0 = insertelement <2 x i32> poison, i32 %readlane0, i32 0
%bv.1 = insertelement <2 x i32> %bv.0, i32 %readlane1, i32 1
ret <2 x i32> %bv.1
}
define <4 x bfloat> @v_vselect_v4bf16(<4 x i1> %cond, <4 x bfloat> %a, <4 x bfloat> %b) {
; GCN-LABEL: v_vselect_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_and_b32_e32 v1, 1, v1
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_and_b32_e32 v2, 1, v2
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v3, 1, v3
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GCN-NEXT: v_cndmask_b32_e32 v3, v11, v7, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GCN-NEXT: v_cndmask_b32_e32 v2, v10, v6, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v9, v5, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v8, v4, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_vselect_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v3, 1, v3
; GFX7-NEXT: v_and_b32_e32 v2, 1, v2
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX7-NEXT: v_and_b32_e32 v1, 1, v1
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_cndmask_b32_e32 v3, v11, v7, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_cndmask_b32_e32 v2, v10, v6, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_cndmask_b32_e32 v1, v9, v5, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v8, v4, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_vselect_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v3, 1, v3
; GFX8-NEXT: v_and_b32_e32 v2, 1, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v8, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v7
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX8-NEXT: v_and_b32_e32 v1, 1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v9, v8, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v2, v7, v5, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v6
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v7, v5, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v6, v4, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v3
; GFX8-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_vselect_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v2, 1, v2
; GFX9-NEXT: v_and_b32_e32 v3, 1, v3
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v2, v7, v5, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX9-NEXT: v_and_b32_e32 v1, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v7, v5, vcc
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v6, v4, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v6
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_perm_b32 v1, v3, v2, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_vselect_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v2, 1, v2
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_and_b32_e32 v1, 1, v1
; GFX10-NEXT: v_and_b32_e32 v3, 1, v3
; GFX10-NEXT: v_lshrrev_b32_e32 v8, 16, v4
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX10-NEXT: v_lshrrev_b32_e32 v9, 16, v6
; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v5, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX10-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX10-NEXT: v_cndmask_b32_e32 v0, v6, v4, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v9, v8, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v3, v7, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_vselect_v4bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: v_mov_b16_e32 v8.l, v7.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v9.l, v5.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v8, v9, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v8, 16, v6
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v9, 16, v4
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v6, v4, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v8, v9, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v5, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_vselect_v4bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v8, 16, v4
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v9, 16, v6
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v2, v7, v5 :: v_dual_and_b32 v3, 1, v3
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v0, v6, v4 :: v_dual_and_b32 v1, 1, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v9, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, v7, v5, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select <4 x i1> %cond, <4 x bfloat> %a, <4 x bfloat> %b
ret <4 x bfloat> %op
}
define <8 x bfloat> @v_vselect_v8bf16(<8 x i1> %cond, <8 x bfloat> %a, <8 x bfloat> %b) {
; GCN-LABEL: v_vselect_v8bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v1, 1, v1
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v2, 1, v2
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v3, 1, v3
; GCN-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_and_b32_e32 v4, 1, v4
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GCN-NEXT: v_and_b32_e32 v5, 1, v5
; GCN-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GCN-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GCN-NEXT: v_and_b32_e32 v6, 1, v6
; GCN-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GCN-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GCN-NEXT: v_and_b32_e32 v7, 1, v7
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GCN-NEXT: v_cndmask_b32_e32 v7, v23, v15, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GCN-NEXT: v_cndmask_b32_e32 v6, v22, v14, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GCN-NEXT: v_cndmask_b32_e32 v5, v21, v13, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GCN-NEXT: v_cndmask_b32_e32 v4, v20, v12, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GCN-NEXT: v_cndmask_b32_e32 v3, v19, v11, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GCN-NEXT: v_cndmask_b32_e32 v2, v18, v10, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v17, v9, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v16, v8, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_vselect_v8bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v7, 1, v7
; GFX7-NEXT: v_and_b32_e32 v6, 1, v6
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GFX7-NEXT: v_and_b32_e32 v5, 1, v5
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: v_cndmask_b32_e32 v7, v23, v15, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GFX7-NEXT: v_and_b32_e32 v4, 1, v4
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: v_cndmask_b32_e32 v6, v22, v14, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GFX7-NEXT: v_and_b32_e32 v3, 1, v3
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: v_cndmask_b32_e32 v5, v21, v13, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GFX7-NEXT: v_and_b32_e32 v2, 1, v2
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_cndmask_b32_e32 v4, v20, v12, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX7-NEXT: v_and_b32_e32 v1, 1, v1
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_cndmask_b32_e32 v3, v19, v11, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_cndmask_b32_e32 v2, v18, v10, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_cndmask_b32_e32 v1, v17, v9, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_cndmask_b32_e32 v0, v16, v8, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_vselect_v8bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v7, 1, v7
; GFX8-NEXT: v_and_b32_e32 v6, 1, v6
; GFX8-NEXT: v_lshrrev_b32_e32 v16, 16, v11
; GFX8-NEXT: v_lshrrev_b32_e32 v17, 16, v15
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GFX8-NEXT: v_and_b32_e32 v5, 1, v5
; GFX8-NEXT: v_cndmask_b32_e32 v7, v17, v16, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GFX8-NEXT: v_and_b32_e32 v4, 1, v4
; GFX8-NEXT: v_cndmask_b32_e32 v6, v15, v11, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v10
; GFX8-NEXT: v_lshrrev_b32_e32 v15, 16, v14
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GFX8-NEXT: v_and_b32_e32 v3, 1, v3
; GFX8-NEXT: v_cndmask_b32_e32 v5, v15, v11, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GFX8-NEXT: v_and_b32_e32 v2, 1, v2
; GFX8-NEXT: v_cndmask_b32_e32 v4, v14, v10, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v13
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX8-NEXT: v_and_b32_e32 v1, 1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v11, v10, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v2, v13, v9, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v8
; GFX8-NEXT: v_lshrrev_b32_e32 v10, 16, v12
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v10, v9, vcc
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v12, v8, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v3
; GFX8-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v7
; GFX8-NEXT: v_or_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v3, v6, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_vselect_v8bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v6, 1, v6
; GFX9-NEXT: v_and_b32_e32 v7, 1, v7
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GFX9-NEXT: v_and_b32_e32 v4, 1, v4
; GFX9-NEXT: v_cndmask_b32_e32 v6, v15, v11, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GFX9-NEXT: v_and_b32_e32 v5, 1, v5
; GFX9-NEXT: v_cndmask_b32_e32 v7, v15, v11, vcc
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GFX9-NEXT: v_and_b32_e32 v2, 1, v2
; GFX9-NEXT: v_cndmask_b32_e32 v4, v14, v10, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v14
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GFX9-NEXT: v_and_b32_e32 v3, 1, v3
; GFX9-NEXT: v_cndmask_b32_e32 v5, v11, v10, vcc
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v2, v13, v9, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v13
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX9-NEXT: v_and_b32_e32 v1, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v10, v9, vcc
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v12, v8, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v12
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v9, v8, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_perm_b32 v1, v3, v2, s4
; GFX9-NEXT: v_perm_b32 v2, v5, v4, s4
; GFX9-NEXT: v_perm_b32 v3, v7, v6, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_vselect_v8bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_and_b32_e32 v6, 1, v6
; GFX10-NEXT: v_and_b32_e32 v4, 1, v4
; GFX10-NEXT: v_and_b32_e32 v5, 1, v5
; GFX10-NEXT: v_and_b32_e32 v2, 1, v2
; GFX10-NEXT: v_lshrrev_b32_e32 v16, 16, v10
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX10-NEXT: v_lshrrev_b32_e32 v17, 16, v14
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_and_b32_e32 v1, 1, v1
; GFX10-NEXT: v_and_b32_e32 v3, 1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v6, v15, v11, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX10-NEXT: v_and_b32_e32 v7, 1, v7
; GFX10-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX10-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX10-NEXT: v_cndmask_b32_e32 v4, v14, v10, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX10-NEXT: v_lshrrev_b32_e32 v10, 16, v8
; GFX10-NEXT: v_lshrrev_b32_e32 v14, 16, v12
; GFX10-NEXT: v_cndmask_b32_e32 v5, v17, v16, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX10-NEXT: v_cndmask_b32_e32 v2, v13, v9, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX10-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX10-NEXT: v_cndmask_b32_e32 v0, v12, v8, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v14, v10, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v3, v13, v9, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v7, v15, v11, vcc_lo
; GFX10-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX10-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_vselect_v8bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v16.l, v15.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v17.l, v11.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX11TRUE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11TRUE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11TRUE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v6, v16, v17, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v16, 16, v14
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v17, 16, v10
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v4, v14, v10, vcc_lo
; GFX11TRUE16-NEXT: v_mov_b16_e32 v10.l, v13.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v14.l, v9.l
; GFX11TRUE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v5, v16, v17, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v10, v14, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v7, 1, v7
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v10, 16, v12
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v14, 16, v8
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v12, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v10, v14, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v13, v9, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v7, v15, v11, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_vselect_v8bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v16, 16, v10
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v17, 16, v14
; GFX11FAKE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11FAKE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v6, v15, v11 :: v_dual_and_b32 v1, 1, v1
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX11FAKE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v4, v14, v10 :: v_dual_and_b32 v3, 1, v3
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v10, 16, v8
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v14, 16, v12
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v5, v17, v16, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v2, v13, v9 :: v_dual_and_b32 v7, 1, v7
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v12, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v14, v10, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, v13, v9, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v7, v15, v11, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select <8 x i1> %cond, <8 x bfloat> %a, <8 x bfloat> %b
ret <8 x bfloat> %op
}
define <16 x bfloat> @v_vselect_v16bf16(<16 x i1> %cond, <16 x bfloat> %a, <16 x bfloat> %b) {
; GCN-LABEL: v_vselect_v16bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: v_writelane_b32 v31, s30, 0
; GCN-NEXT: v_writelane_b32 v31, s31, 1
; GCN-NEXT: v_writelane_b32 v31, s34, 2
; GCN-NEXT: v_writelane_b32 v31, s35, 3
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v1
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v2
; GCN-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v3
; GCN-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v4
; GCN-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v5
; GCN-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v6
; GCN-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v7
; GCN-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v8
; GCN-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v0
; GCN-NEXT: v_and_b32_e32 v0, 1, v9
; GCN-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v0
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v16
; GCN-NEXT: v_and_b32_e32 v1, 1, v10
; GCN-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v1
; GCN-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v17
; GCN-NEXT: v_and_b32_e32 v3, 1, v11
; GCN-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v3
; GCN-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:8
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v18
; GCN-NEXT: v_and_b32_e32 v5, 1, v12
; GCN-NEXT: v_cmp_eq_u32_e64 s[26:27], 1, v5
; GCN-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:12
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v19
; GCN-NEXT: v_and_b32_e32 v7, 1, v13
; GCN-NEXT: v_and_b32_e32 v8, 1, v14
; GCN-NEXT: v_cmp_eq_u32_e64 s[28:29], 1, v7
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32
; GCN-NEXT: v_cmp_eq_u32_e64 s[30:31], 1, v8
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:64
; GCN-NEXT: v_and_b32_e32 v9, 1, v15
; GCN-NEXT: v_cmp_eq_u32_e64 s[34:35], 1, v9
; GCN-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:60
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_cndmask_b32_e64 v15, v8, v7, s[34:35]
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:56
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v30
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_cndmask_b32_e64 v14, v9, v8, s[30:31]
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:52
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v29
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_cndmask_b32_e64 v13, v7, v9, s[28:29]
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:48
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v28
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_cndmask_b32_e64 v12, v8, v9, s[26:27]
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:44
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v27
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_cndmask_b32_e64 v11, v7, v9, s[24:25]
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:40
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v26
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_cndmask_b32_e64 v10, v8, v9, s[22:23]
; GCN-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v25
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_cndmask_b32_e64 v9, v7, v9, s[20:21]
; GCN-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v24
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_cndmask_b32_e64 v8, v8, v16, s[18:19]
; GCN-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:28
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v23
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_cndmask_b32_e64 v7, v7, v17, s[16:17]
; GCN-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:24
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v22
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GCN-NEXT: v_cndmask_b32_e64 v16, v16, v18, s[14:15]
; GCN-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:16
; GCN-NEXT: v_mul_f32_e32 v19, 1.0, v20
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v21
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GCN-NEXT: v_cndmask_b32_e64 v17, v17, v20, s[12:13]
; GCN-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:20
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GCN-NEXT: v_cndmask_b32_e64 v19, v20, v19, s[10:11]
; GCN-NEXT: v_cndmask_b32_e64 v5, v18, v5, s[8:9]
; GCN-NEXT: v_cndmask_b32_e64 v3, v6, v3, s[6:7]
; GCN-NEXT: v_cndmask_b32_e64 v1, v4, v1, s[4:5]
; GCN-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_readlane_b32 s35, v31, 3
; GCN-NEXT: v_readlane_b32 s34, v31, 2
; GCN-NEXT: v_readlane_b32 s31, v31, 1
; GCN-NEXT: v_readlane_b32 s30, v31, 0
; GCN-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GCN-NEXT: s_mov_b64 exec, s[4:5]
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_vselect_v16bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v1
; GFX7-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v2
; GFX7-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v3
; GFX7-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v4
; GFX7-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v5
; GFX7-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v6
; GFX7-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v7
; GFX7-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v8
; GFX7-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v9
; GFX7-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v10
; GFX7-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v0
; GFX7-NEXT: v_and_b32_e32 v0, 1, v11
; GFX7-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v0
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:64
; GFX7-NEXT: v_and_b32_e32 v2, 1, v12
; GFX7-NEXT: v_writelane_b32 v31, s30, 0
; GFX7-NEXT: v_cmp_eq_u32_e64 s[26:27], 1, v2
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:60
; GFX7-NEXT: v_and_b32_e32 v3, 1, v13
; GFX7-NEXT: v_writelane_b32 v31, s31, 1
; GFX7-NEXT: v_cmp_eq_u32_e64 s[28:29], 1, v3
; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:56
; GFX7-NEXT: v_and_b32_e32 v4, 1, v14
; GFX7-NEXT: v_writelane_b32 v31, s34, 2
; GFX7-NEXT: v_cmp_eq_u32_e64 s[30:31], 1, v4
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:52
; GFX7-NEXT: v_and_b32_e32 v5, 1, v15
; GFX7-NEXT: v_writelane_b32 v31, s35, 3
; GFX7-NEXT: v_cmp_eq_u32_e64 s[34:35], 1, v5
; GFX7-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:48
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: s_waitcnt vmcnt(5)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v15, v1, v0, s[34:35]
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:44
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v30
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_cndmask_b32_e64 v14, v2, v1, s[30:31]
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:40
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v29
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_cndmask_b32_e64 v13, v3, v2, s[28:29]
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:36
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v28
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_cndmask_b32_e64 v12, v4, v3, s[26:27]
; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:32
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v27
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_cndmask_b32_e64 v11, v5, v4, s[24:25]
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:28
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v26
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: v_readlane_b32 s35, v31, 3
; GFX7-NEXT: v_readlane_b32 s34, v31, 2
; GFX7-NEXT: v_readlane_b32 s31, v31, 1
; GFX7-NEXT: v_readlane_b32 s30, v31, 0
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_cndmask_b32_e64 v10, v0, v5, s[22:23]
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:24
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v25
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_cndmask_b32_e64 v9, v1, v5, s[20:21]
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v24
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_cndmask_b32_e64 v8, v2, v5, s[18:19]
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:8
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v23
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_cndmask_b32_e64 v7, v3, v5, s[16:17]
; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:12
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v22
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_cndmask_b32_e64 v6, v4, v5, s[14:15]
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:16
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v21
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_cndmask_b32_e64 v5, v0, v5, s[12:13]
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:20
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(4)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_cndmask_b32_e64 v2, v2, v17, s[4:5]
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_cndmask_b32_e64 v3, v3, v18, s[6:7]
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_cndmask_b32_e64 v4, v4, v19, s[8:9]
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_cndmask_b32_e64 v20, v0, v20, s[10:11]
; GFX7-NEXT: v_cndmask_b32_e32 v0, v1, v16, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v20
; GFX7-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX7-NEXT: s_mov_b64 exec, s[4:5]
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_vselect_v16bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v1
; GFX8-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v2
; GFX8-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v3
; GFX8-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v4
; GFX8-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v5
; GFX8-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v6
; GFX8-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v7
; GFX8-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v8
; GFX8-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v9
; GFX8-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v10
; GFX8-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v0
; GFX8-NEXT: buffer_load_dword v0, off, s[0:3], s32
; GFX8-NEXT: v_writelane_b32 v31, s30, 0
; GFX8-NEXT: v_and_b32_e32 v2, 1, v12
; GFX8-NEXT: v_and_b32_e32 v3, 1, v13
; GFX8-NEXT: v_writelane_b32 v31, s31, 1
; GFX8-NEXT: v_cmp_eq_u32_e64 s[26:27], 1, v2
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v22
; GFX8-NEXT: v_cmp_eq_u32_e64 s[28:29], 1, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v30
; GFX8-NEXT: v_writelane_b32 v31, s34, 2
; GFX8-NEXT: v_and_b32_e32 v1, 1, v11
; GFX8-NEXT: v_and_b32_e32 v4, 1, v14
; GFX8-NEXT: v_and_b32_e32 v5, 1, v15
; GFX8-NEXT: v_cndmask_b32_e64 v6, v3, v2, s[28:29]
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v20
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v28
; GFX8-NEXT: v_writelane_b32 v31, s35, 3
; GFX8-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v23
; GFX8-NEXT: v_cmp_eq_u32_e64 s[30:31], 1, v4
; GFX8-NEXT: v_cmp_eq_u32_e64 s[34:35], 1, v5
; GFX8-NEXT: v_cndmask_b32_e64 v10, v3, v2, s[20:21]
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v21
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v29
; GFX8-NEXT: v_cndmask_b32_e64 v5, v5, v4, s[24:25]
; GFX8-NEXT: v_lshrrev_b32_e32 v4, 16, v19
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v27
; GFX8-NEXT: v_cndmask_b32_e64 v3, v9, v4, s[16:17]
; GFX8-NEXT: v_lshrrev_b32_e32 v15, 16, v24
; GFX8-NEXT: v_cndmask_b32_e64 v7, v30, v22, s[26:27]
; GFX8-NEXT: v_cndmask_b32_e64 v4, v27, v19, s[14:15]
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v6
; GFX8-NEXT: v_cndmask_b32_e64 v8, v29, v21, s[22:23]
; GFX8-NEXT: v_cndmask_b32_e64 v11, v28, v20, s[18:19]
; GFX8-NEXT: v_cndmask_b32_e64 v9, v26, v18, s[10:11]
; GFX8-NEXT: v_cndmask_b32_e64 v14, v25, v17, s[6:7]
; GFX8-NEXT: v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v10
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v4, v11, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v5, v8, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshrrev_b32_e32 v2, 16, v0
; GFX8-NEXT: v_cndmask_b32_e64 v12, v0, v23, s[30:31]
; GFX8-NEXT: v_cndmask_b32_e64 v13, v2, v1, s[34:35]
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v18
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v26
; GFX8-NEXT: v_cndmask_b32_e64 v2, v1, v0, s[12:13]
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v17
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v25
; GFX8-NEXT: v_cndmask_b32_e64 v1, v1, v0, s[8:9]
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v16
; GFX8-NEXT: v_cndmask_b32_e64 v0, v15, v0, s[4:5]
; GFX8-NEXT: v_cndmask_b32_e32 v15, v24, v16, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v13
; GFX8-NEXT: v_or_b32_sdwa v0, v15, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v1, v14, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v2, v9, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v7, v12, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_readlane_b32 s35, v31, 3
; GFX8-NEXT: v_readlane_b32 s34, v31, 2
; GFX8-NEXT: v_readlane_b32 s31, v31, 1
; GFX8-NEXT: v_readlane_b32 s30, v31, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_vselect_v16bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v4, 1, v4
; GFX9-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v4
; GFX9-NEXT: v_and_b32_e32 v4, 1, v14
; GFX9-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v4
; GFX9-NEXT: v_and_b32_e32 v4, 1, v15
; GFX9-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v4
; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32
; GFX9-NEXT: v_and_b32_e32 v12, 1, v12
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v12
; GFX9-NEXT: v_and_b32_e32 v12, 1, v13
; GFX9-NEXT: v_and_b32_e32 v10, 1, v10
; GFX9-NEXT: v_and_b32_e32 v6, 1, v6
; GFX9-NEXT: v_and_b32_e32 v5, 1, v5
; GFX9-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v12
; GFX9-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v10
; GFX9-NEXT: v_and_b32_e32 v10, 1, v11
; GFX9-NEXT: v_and_b32_e32 v8, 1, v8
; GFX9-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v6
; GFX9-NEXT: v_and_b32_e32 v6, 1, v7
; GFX9-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v5
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v22
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v30
; GFX9-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v10
; GFX9-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v8
; GFX9-NEXT: v_and_b32_e32 v8, 1, v9
; GFX9-NEXT: v_cndmask_b32_e64 v7, v7, v5, s[4:5]
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v21
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v29
; GFX9-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v8
; GFX9-NEXT: v_cndmask_b32_e64 v5, v9, v5, s[8:9]
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v20
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v28
; GFX9-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v6
; GFX9-NEXT: v_cndmask_b32_e64 v9, v11, v9, s[12:13]
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v19
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v27
; GFX9-NEXT: v_and_b32_e32 v2, 1, v2
; GFX9-NEXT: v_cndmask_b32_e64 v11, v13, v11, s[16:17]
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v23
; GFX9-NEXT: v_cndmask_b32_e32 v6, v30, v22, vcc
; GFX9-NEXT: v_and_b32_e32 v3, 1, v3
; GFX9-NEXT: v_cndmask_b32_e64 v15, v26, v18, s[18:19]
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v2, v25, v17, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX9-NEXT: v_and_b32_e32 v1, 1, v1
; GFX9-NEXT: v_cndmask_b32_e64 v8, v29, v21, s[6:7]
; GFX9-NEXT: v_cndmask_b32_e64 v10, v28, v20, s[10:11]
; GFX9-NEXT: v_cndmask_b32_e64 v12, v27, v19, s[14:15]
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v5, v5, v8, s4
; GFX9-NEXT: v_perm_b32 v6, v7, v6, s4
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_cndmask_b32_e64 v14, v4, v23, s[20:21]
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX9-NEXT: v_cndmask_b32_e64 v13, v4, v13, s[22:23]
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v18
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v26
; GFX9-NEXT: v_cndmask_b32_e64 v4, v18, v4, s[24:25]
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v25
; GFX9-NEXT: v_cndmask_b32_e32 v3, v18, v17, vcc
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v24, v16, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v24
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v17, v16, vcc
; GFX9-NEXT: v_perm_b32 v0, v1, v0, s4
; GFX9-NEXT: v_perm_b32 v1, v3, v2, s4
; GFX9-NEXT: v_perm_b32 v2, v4, v15, s4
; GFX9-NEXT: v_perm_b32 v3, v11, v12, s4
; GFX9-NEXT: v_perm_b32 v4, v9, v10, s4
; GFX9-NEXT: v_perm_b32 v7, v13, v14, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_vselect_v16bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32
; GFX10-NEXT: v_and_b32_e32 v12, 1, v12
; GFX10-NEXT: v_and_b32_e32 v13, 1, v13
; GFX10-NEXT: v_and_b32_e32 v10, 1, v10
; GFX10-NEXT: v_lshrrev_b32_e32 v33, 16, v22
; GFX10-NEXT: v_lshrrev_b32_e32 v34, 16, v30
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v12
; GFX10-NEXT: v_and_b32_e32 v11, 1, v11
; GFX10-NEXT: v_and_b32_e32 v8, 1, v8
; GFX10-NEXT: v_lshrrev_b32_e32 v35, 16, v21
; GFX10-NEXT: v_lshrrev_b32_e32 v36, 16, v29
; GFX10-NEXT: v_cndmask_b32_e32 v12, v30, v22, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v13
; GFX10-NEXT: v_and_b32_e32 v9, 1, v9
; GFX10-NEXT: v_and_b32_e32 v6, 1, v6
; GFX10-NEXT: v_lshrrev_b32_e32 v37, 16, v20
; GFX10-NEXT: v_lshrrev_b32_e32 v38, 16, v28
; GFX10-NEXT: v_cndmask_b32_e32 v13, v34, v33, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v10
; GFX10-NEXT: v_and_b32_e32 v4, 1, v4
; GFX10-NEXT: v_and_b32_e32 v2, 1, v2
; GFX10-NEXT: v_and_b32_e32 v3, 1, v3
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v10, v29, v21, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v11
; GFX10-NEXT: v_lshrrev_b32_e32 v51, 16, v17
; GFX10-NEXT: v_lshrrev_b32_e32 v52, 16, v25
; GFX10-NEXT: v_and_b32_e32 v1, 1, v1
; GFX10-NEXT: v_and_b32_e32 v5, 1, v5
; GFX10-NEXT: v_cndmask_b32_e32 v11, v36, v35, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v8
; GFX10-NEXT: v_lshrrev_b32_e32 v53, 16, v16
; GFX10-NEXT: v_lshrrev_b32_e32 v54, 16, v24
; GFX10-NEXT: v_and_b32_e32 v7, 1, v7
; GFX10-NEXT: v_lshrrev_b32_e32 v49, 16, v18
; GFX10-NEXT: v_cndmask_b32_e32 v8, v28, v20, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v9
; GFX10-NEXT: v_lshrrev_b32_e32 v50, 16, v26
; GFX10-NEXT: v_and_b32_e32 v14, 1, v14
; GFX10-NEXT: v_lshrrev_b32_e32 v39, 16, v19
; GFX10-NEXT: v_lshrrev_b32_e32 v48, 16, v27
; GFX10-NEXT: v_cndmask_b32_e32 v9, v38, v37, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX10-NEXT: v_and_b32_e32 v15, 1, v15
; GFX10-NEXT: v_lshrrev_b32_e32 v32, 16, v23
; GFX10-NEXT: v_cndmask_b32_e32 v6, v27, v19, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX10-NEXT: v_cndmask_b32_e32 v4, v26, v18, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX10-NEXT: v_cndmask_b32_e32 v2, v25, v17, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX10-NEXT: v_cndmask_b32_e32 v3, v52, v51, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v24, v16, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX10-NEXT: v_cndmask_b32_e32 v1, v54, v53, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v5, v50, v49, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX10-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX10-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX10-NEXT: v_cndmask_b32_e32 v7, v48, v39, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v14
; GFX10-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
; GFX10-NEXT: v_perm_b32 v5, v11, v10, 0x5040100
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v31
; GFX10-NEXT: v_cndmask_b32_e32 v14, v31, v23, vcc_lo
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v15
; GFX10-NEXT: v_cndmask_b32_e32 v15, v3, v32, vcc_lo
; GFX10-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX10-NEXT: v_perm_b32 v6, v13, v12, 0x5040100
; GFX10-NEXT: v_perm_b32 v7, v15, v14, 0x5040100
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_vselect_v16bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: scratch_load_b32 v31, off, s32
; GFX11TRUE16-NEXT: v_mov_b16_e32 v49.l, v26.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v50.l, v18.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GFX11TRUE16-NEXT: v_and_b32_e32 v12, 1, v12
; GFX11TRUE16-NEXT: v_mov_b16_e32 v33.l, v30.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v34.l, v22.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v53.l, v24.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v54.l, v16.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GFX11TRUE16-NEXT: v_and_b32_e32 v10, 1, v10
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v12
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11TRUE16-NEXT: v_mov_b16_e32 v35.l, v29.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v36.l, v21.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v51.l, v25.l
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v12, v33, v34, vcc_lo
; GFX11TRUE16-NEXT: v_mov_b16_e32 v52.l, v17.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v25, 16, v25
; GFX11TRUE16-NEXT: v_and_b32_e32 v8, 1, v8
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v10
; GFX11TRUE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11TRUE16-NEXT: v_mov_b16_e32 v37.l, v28.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v38.l, v20.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v10, v35, v36, vcc_lo
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GFX11TRUE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v8
; GFX11TRUE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v48.l, v19.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v8, v37, v38, vcc_lo
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v27, 16, v27
; GFX11TRUE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11TRUE16-NEXT: v_and_b32_e32 v7, 1, v7
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v21, 16, v21
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v29, 16, v29
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v6, v39, v48, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11TRUE16-NEXT: v_and_b32_e32 v9, 1, v9
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v4, v49, v50, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v11, 1, v11
; GFX11TRUE16-NEXT: v_mov_b16_e32 v32.l, v23.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v23, 16, v23
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v51, v52, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v13, 1, v13
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v53, v54, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v13
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v13, v30, v22, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v11
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v11, v29, v21, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v9
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v9, v28, v20, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v7, v27, v19, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v24, v16, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v25, v17, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v5, v26, v18, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v5, v11, v10, 0x5040100
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v16, 16, v31
; GFX11TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.l
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v3.l, v16.l
; GFX11TRUE16-NEXT: v_and_b32_e32 v14, 1, v14
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v14
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_cndmask_b32 v14, v17, v32 :: v_dual_and_b32 v15, 1, v15
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v15
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v15, v3, v23, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v6, v13, v12, 0x5040100
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_perm_b32 v7, v15, v14, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_vselect_v16bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: scratch_load_b32 v31, off, s32
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v19
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v27
; GFX11FAKE16-NEXT: v_and_b32_e32 v12, 1, v12
; GFX11FAKE16-NEXT: v_and_b32_e32 v13, 1, v13
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v22
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v30
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v18
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v12
; GFX11FAKE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v26
; GFX11FAKE16-NEXT: v_and_b32_e32 v10, 1, v10
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v12, v30, v22 :: v_dual_and_b32 v11, 1, v11
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v13
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v21
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v29
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v16
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v13, v34, v33, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v10
; GFX11FAKE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v24
; GFX11FAKE16-NEXT: v_and_b32_e32 v8, 1, v8
; GFX11FAKE16-NEXT: v_and_b32_e32 v9, 1, v9
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v10, v29, v21, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v11
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v20
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v28
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v17
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v11, v36, v35, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v8
; GFX11FAKE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v25
; GFX11FAKE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v23
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v8, v28, v20, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v9
; GFX11FAKE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11FAKE16-NEXT: v_and_b32_e32 v15, 1, v15
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v9, v38, v37, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v6, v27, v19, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v4, v26, v18 :: v_dual_and_b32 v7, 1, v7
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v25, v17, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, v52, v51, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v24, v16, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v54, v53, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v5, v50, v49, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v7, v48, v39, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v5, v11, v10, 0x5040100
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v31
; GFX11FAKE16-NEXT: v_and_b32_e32 v14, 1, v14
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v14
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v14, v31, v23, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v15
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v15, v3, v32, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v6, v13, v12, 0x5040100
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_perm_b32 v7, v15, v14, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select <16 x i1> %cond, <16 x bfloat> %a, <16 x bfloat> %b
ret <16 x bfloat> %op
}
define <32 x bfloat> @v_vselect_v32bf16(<32 x i1> %cond, <32 x bfloat> %a, <32 x bfloat> %b) {
; GCN-LABEL: v_vselect_v32bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GCN-NEXT: buffer_store_dword v61, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GCN-NEXT: v_and_b32_e32 v0, 1, v0
; GCN-NEXT: v_and_b32_e32 v1, 1, v1
; GCN-NEXT: v_and_b32_e32 v2, 1, v2
; GCN-NEXT: v_and_b32_e32 v36, 1, v13
; GCN-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:52
; GCN-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:180
; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:56
; GCN-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:184
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:188
; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:64
; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:192
; GCN-NEXT: v_and_b32_e32 v53, 1, v26
; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:84
; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:88
; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:92
; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:96
; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:100
; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:108
; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:112
; GCN-NEXT: v_and_b32_e32 v27, 1, v27
; GCN-NEXT: v_and_b32_e32 v28, 1, v28
; GCN-NEXT: v_and_b32_e32 v29, 1, v29
; GCN-NEXT: v_and_b32_e32 v30, 1, v30
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:116
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:120
; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:124
; GCN-NEXT: buffer_load_dword v26, off, s[0:3], s32
; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:252
; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:248
; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:244
; GCN-NEXT: s_waitcnt expcnt(6)
; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:240
; GCN-NEXT: s_waitcnt vmcnt(14)
; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v37
; GCN-NEXT: v_mul_f32_e32 v38, 1.0, v38
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v36
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v43
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v44
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v30
; GCN-NEXT: v_cndmask_b32_e64 v30, v37, v36, s[4:5]
; GCN-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:236
; GCN-NEXT: s_waitcnt expcnt(5)
; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:232
; GCN-NEXT: s_waitcnt expcnt(4)
; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:228
; GCN-NEXT: s_waitcnt expcnt(3)
; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:224
; GCN-NEXT: s_waitcnt expcnt(2)
; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:220
; GCN-NEXT: s_waitcnt expcnt(1)
; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:216
; GCN-NEXT: s_waitcnt expcnt(0)
; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:212
; GCN-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:128
; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v42
; GCN-NEXT: s_waitcnt vmcnt(10)
; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v45
; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
; GCN-NEXT: s_waitcnt vmcnt(9)
; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v46
; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
; GCN-NEXT: s_waitcnt vmcnt(8)
; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v47
; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
; GCN-NEXT: s_waitcnt vmcnt(7)
; GCN-NEXT: v_mul_f32_e32 v36, 1.0, v36
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v29
; GCN-NEXT: v_cndmask_b32_e64 v29, v43, v42, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v28
; GCN-NEXT: v_cndmask_b32_e64 v28, v44, v41, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v27
; GCN-NEXT: v_cndmask_b32_e64 v27, v45, v55, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v53
; GCN-NEXT: v_cndmask_b32_e64 v36, v36, v54, s[4:5]
; GCN-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:4
; GCN-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:132
; GCN-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:8
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:136
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:12
; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:140
; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:16
; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:144
; GCN-NEXT: v_and_b32_e32 v3, 1, v3
; GCN-NEXT: v_and_b32_e32 v4, 1, v4
; GCN-NEXT: v_and_b32_e32 v5, 1, v5
; GCN-NEXT: v_and_b32_e32 v6, 1, v6
; GCN-NEXT: v_and_b32_e32 v18, 1, v18
; GCN-NEXT: v_and_b32_e32 v22, 1, v22
; GCN-NEXT: v_and_b32_e32 v23, 1, v23
; GCN-NEXT: v_and_b32_e32 v24, 1, v24
; GCN-NEXT: v_and_b32_e32 v25, 1, v25
; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
; GCN-NEXT: s_waitcnt vmcnt(14)
; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v56
; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
; GCN-NEXT: s_waitcnt vmcnt(13)
; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v57
; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
; GCN-NEXT: s_waitcnt vmcnt(12)
; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v58
; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
; GCN-NEXT: s_waitcnt vmcnt(11)
; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v59
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v25
; GCN-NEXT: v_cndmask_b32_e64 v25, v46, v52, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v24
; GCN-NEXT: v_cndmask_b32_e64 v24, v47, v51, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v23
; GCN-NEXT: v_cndmask_b32_e64 v23, v56, v50, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v22
; GCN-NEXT: v_cndmask_b32_e64 v22, v57, v49, s[4:5]
; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:68
; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:196
; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:72
; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:200
; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:76
; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:204
; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:80
; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:208
; GCN-NEXT: v_and_b32_e32 v19, 1, v19
; GCN-NEXT: v_and_b32_e32 v20, 1, v20
; GCN-NEXT: v_and_b32_e32 v21, 1, v21
; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
; GCN-NEXT: s_waitcnt vmcnt(14)
; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v60
; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v61
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v46
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v47
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v56
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v57
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v21
; GCN-NEXT: v_cndmask_b32_e64 v21, v58, v48, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v20
; GCN-NEXT: v_cndmask_b32_e64 v20, v59, v39, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v19
; GCN-NEXT: v_cndmask_b32_e64 v19, v57, v56, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v18
; GCN-NEXT: v_cndmask_b32_e64 v18, v47, v46, s[4:5]
; GCN-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:20
; GCN-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:148
; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:24
; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:152
; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:28
; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:156
; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:32
; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:160
; GCN-NEXT: v_and_b32_e32 v7, 1, v7
; GCN-NEXT: v_and_b32_e32 v8, 1, v8
; GCN-NEXT: v_and_b32_e32 v9, 1, v9
; GCN-NEXT: v_and_b32_e32 v10, 1, v10
; GCN-NEXT: v_and_b32_e32 v14, 1, v14
; GCN-NEXT: v_and_b32_e32 v15, 1, v15
; GCN-NEXT: v_and_b32_e32 v16, 1, v16
; GCN-NEXT: v_and_b32_e32 v17, 1, v17
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v17
; GCN-NEXT: v_cndmask_b32_e64 v17, v52, v51, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v16
; GCN-NEXT: v_cndmask_b32_e64 v16, v50, v49, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v15
; GCN-NEXT: v_cndmask_b32_e64 v15, v35, v34, s[4:5]
; GCN-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v14
; GCN-NEXT: v_cndmask_b32_e64 v14, v33, v32, s[4:5]
; GCN-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
; GCN-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:164
; GCN-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:40
; GCN-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:168
; GCN-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:44
; GCN-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:172
; GCN-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:48
; GCN-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:176
; GCN-NEXT: v_and_b32_e32 v11, 1, v11
; GCN-NEXT: v_and_b32_e32 v12, 1, v12
; GCN-NEXT: v_cndmask_b32_e32 v38, v38, v40, vcc
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:256
; GCN-NEXT: v_and_b32_e32 v26, 1, v26
; GCN-NEXT: v_mul_f32_e32 v53, 1.0, v53
; GCN-NEXT: v_mul_f32_e32 v54, 1.0, v54
; GCN-NEXT: v_mul_f32_e32 v55, 1.0, v55
; GCN-NEXT: v_mul_f32_e32 v41, 1.0, v41
; GCN-NEXT: v_mul_f32_e32 v42, 1.0, v42
; GCN-NEXT: v_mul_f32_e32 v43, 1.0, v43
; GCN-NEXT: v_mul_f32_e32 v44, 1.0, v44
; GCN-NEXT: v_mul_f32_e32 v45, 1.0, v45
; GCN-NEXT: s_waitcnt vmcnt(14)
; GCN-NEXT: v_mul_f32_e32 v39, 1.0, v39
; GCN-NEXT: v_mul_f32_e32 v48, 1.0, v48
; GCN-NEXT: v_mul_f32_e32 v46, 1.0, v46
; GCN-NEXT: s_waitcnt vmcnt(13)
; GCN-NEXT: v_mul_f32_e32 v47, 1.0, v47
; GCN-NEXT: s_waitcnt vmcnt(12)
; GCN-NEXT: v_mul_f32_e32 v56, 1.0, v56
; GCN-NEXT: s_waitcnt vmcnt(11)
; GCN-NEXT: v_mul_f32_e32 v57, 1.0, v57
; GCN-NEXT: s_waitcnt vmcnt(10)
; GCN-NEXT: v_mul_f32_e32 v58, 1.0, v58
; GCN-NEXT: s_waitcnt vmcnt(9)
; GCN-NEXT: v_mul_f32_e32 v59, 1.0, v59
; GCN-NEXT: s_waitcnt vmcnt(8)
; GCN-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GCN-NEXT: s_waitcnt vmcnt(7)
; GCN-NEXT: v_mul_f32_e32 v33, 1.0, v33
; GCN-NEXT: s_waitcnt vmcnt(6)
; GCN-NEXT: v_mul_f32_e32 v34, 1.0, v34
; GCN-NEXT: s_waitcnt vmcnt(5)
; GCN-NEXT: v_mul_f32_e32 v35, 1.0, v35
; GCN-NEXT: s_waitcnt vmcnt(4)
; GCN-NEXT: v_mul_f32_e32 v49, 1.0, v49
; GCN-NEXT: s_waitcnt vmcnt(3)
; GCN-NEXT: v_mul_f32_e32 v50, 1.0, v50
; GCN-NEXT: s_waitcnt vmcnt(2)
; GCN-NEXT: v_mul_f32_e32 v51, 1.0, v51
; GCN-NEXT: s_waitcnt vmcnt(1)
; GCN-NEXT: v_mul_f32_e32 v52, 1.0, v52
; GCN-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GCN-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GCN-NEXT: v_mul_f32_e32 v37, 1.0, v37
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v40, 1.0, v40
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v12
; GCN-NEXT: v_cndmask_b32_e32 v12, v31, v13, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v11
; GCN-NEXT: v_cndmask_b32_e32 v11, v52, v51, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
; GCN-NEXT: v_cndmask_b32_e32 v10, v50, v49, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v9
; GCN-NEXT: v_cndmask_b32_e32 v9, v35, v34, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v8
; GCN-NEXT: v_cndmask_b32_e32 v8, v33, v32, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GCN-NEXT: v_cndmask_b32_e32 v7, v59, v58, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GCN-NEXT: v_cndmask_b32_e32 v6, v57, v56, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GCN-NEXT: v_cndmask_b32_e32 v5, v47, v46, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GCN-NEXT: v_cndmask_b32_e32 v4, v48, v39, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GCN-NEXT: v_cndmask_b32_e32 v3, v45, v44, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GCN-NEXT: v_cndmask_b32_e32 v2, v43, v42, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GCN-NEXT: v_cndmask_b32_e32 v1, v41, v55, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GCN-NEXT: v_cndmask_b32_e32 v0, v54, v53, vcc
; GCN-NEXT: v_cmp_eq_u32_e32 vcc, 1, v26
; GCN-NEXT: v_cndmask_b32_e32 v31, v40, v37, vcc
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GCN-NEXT: v_and_b32_e32 v13, 0xffff0000, v38
; GCN-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GCN-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GCN-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GCN-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GCN-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GCN-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GCN-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GCN-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GCN-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GCN-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GCN-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GCN-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GCN-NEXT: v_and_b32_e32 v26, 0xffff0000, v36
; GCN-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GCN-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GCN-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GCN-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GCN-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GCN-NEXT: buffer_load_dword v61, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GCN-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GCN-NEXT: s_waitcnt vmcnt(0)
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_vselect_v32bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_and_b32_e32 v24, 1, v24
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v24
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:228
; GFX7-NEXT: v_and_b32_e32 v25, 1, v25
; GFX7-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v25
; GFX7-NEXT: v_and_b32_e32 v30, 1, v30
; GFX7-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v30
; GFX7-NEXT: v_and_b32_e32 v29, 1, v29
; GFX7-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v29
; GFX7-NEXT: v_and_b32_e32 v28, 1, v28
; GFX7-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v28
; GFX7-NEXT: v_and_b32_e32 v27, 1, v27
; GFX7-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v27
; GFX7-NEXT: v_and_b32_e32 v26, 1, v26
; GFX7-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v26
; GFX7-NEXT: v_and_b32_e32 v23, 1, v23
; GFX7-NEXT: v_and_b32_e32 v22, 1, v22
; GFX7-NEXT: v_and_b32_e32 v21, 1, v21
; GFX7-NEXT: v_and_b32_e32 v20, 1, v20
; GFX7-NEXT: v_and_b32_e32 v19, 1, v19
; GFX7-NEXT: v_and_b32_e32 v18, 1, v18
; GFX7-NEXT: v_and_b32_e32 v17, 1, v17
; GFX7-NEXT: v_and_b32_e32 v16, 1, v16
; GFX7-NEXT: v_and_b32_e32 v15, 1, v15
; GFX7-NEXT: v_and_b32_e32 v14, 1, v14
; GFX7-NEXT: v_and_b32_e32 v13, 1, v13
; GFX7-NEXT: v_and_b32_e32 v12, 1, v12
; GFX7-NEXT: v_and_b32_e32 v11, 1, v11
; GFX7-NEXT: v_and_b32_e32 v10, 1, v10
; GFX7-NEXT: v_and_b32_e32 v9, 1, v9
; GFX7-NEXT: v_and_b32_e32 v8, 1, v8
; GFX7-NEXT: v_and_b32_e32 v7, 1, v7
; GFX7-NEXT: v_and_b32_e32 v6, 1, v6
; GFX7-NEXT: v_and_b32_e32 v5, 1, v5
; GFX7-NEXT: v_and_b32_e32 v4, 1, v4
; GFX7-NEXT: v_and_b32_e32 v3, 1, v3
; GFX7-NEXT: v_and_b32_e32 v2, 1, v2
; GFX7-NEXT: v_and_b32_e32 v1, 1, v1
; GFX7-NEXT: v_and_b32_e32 v0, 1, v0
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:252
; GFX7-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:256
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_and_b32_e32 v24, 1, v24
; GFX7-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v24
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:124
; GFX7-NEXT: s_waitcnt vmcnt(3)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: s_waitcnt vmcnt(2)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v31, 1.0, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_cndmask_b32_e64 v30, v25, v24, s[12:13]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:120
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:248
; GFX7-NEXT: v_and_b32_e32 v30, 0xffff0000, v30
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_cndmask_b32_e64 v29, v25, v24, s[14:15]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:244
; GFX7-NEXT: v_and_b32_e32 v29, 0xffff0000, v29
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_cndmask_b32_e64 v28, v25, v24, s[16:17]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:112
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:240
; GFX7-NEXT: v_and_b32_e32 v28, 0xffff0000, v28
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_cndmask_b32_e64 v27, v25, v24, s[10:11]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:108
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:236
; GFX7-NEXT: v_and_b32_e32 v27, 0xffff0000, v27
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_cndmask_b32_e64 v26, v25, v24, s[8:9]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:104
; GFX7-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:232
; GFX7-NEXT: v_and_b32_e32 v26, 0xffff0000, v26
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v25, 1.0, v25
; GFX7-NEXT: v_cndmask_b32_e64 v25, v25, v24, s[6:7]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:128
; GFX7-NEXT: v_and_b32_e32 v25, 0xffff0000, v25
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_cndmask_b32_e64 v31, v31, v24, s[4:5]
; GFX7-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:100
; GFX7-NEXT: v_and_b32_e32 v31, 0xffff0000, v31
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v24, 1.0, v24
; GFX7-NEXT: v_cndmask_b32_e32 v24, v32, v24, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v23
; GFX7-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:96
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:224
; GFX7-NEXT: v_and_b32_e32 v24, 0xffff0000, v24
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v23, 1.0, v23
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v23, v32, v23, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v22
; GFX7-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:92
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:220
; GFX7-NEXT: v_and_b32_e32 v23, 0xffff0000, v23
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v22, 1.0, v22
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v22, v32, v22, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v21
; GFX7-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:88
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:216
; GFX7-NEXT: v_and_b32_e32 v22, 0xffff0000, v22
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v21, 1.0, v21
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v21, v32, v21, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v20
; GFX7-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:84
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:212
; GFX7-NEXT: v_and_b32_e32 v21, 0xffff0000, v21
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v20, 1.0, v20
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v20, v32, v20, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v19
; GFX7-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:208
; GFX7-NEXT: v_and_b32_e32 v20, 0xffff0000, v20
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v19, 1.0, v19
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v19, v32, v19, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v18
; GFX7-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:76
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:204
; GFX7-NEXT: v_and_b32_e32 v19, 0xffff0000, v19
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v18, 1.0, v18
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v18, v32, v18, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v17
; GFX7-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:200
; GFX7-NEXT: v_and_b32_e32 v18, 0xffff0000, v18
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v17, 1.0, v17
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v17, v32, v17, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v16
; GFX7-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:68
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:196
; GFX7-NEXT: v_and_b32_e32 v17, 0xffff0000, v17
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v16, 1.0, v16
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v16, v32, v16, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v15
; GFX7-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:64
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:192
; GFX7-NEXT: v_and_b32_e32 v16, 0xffff0000, v16
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v15, 1.0, v15
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v15, v32, v15, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v14
; GFX7-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:60
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:188
; GFX7-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v14, 1.0, v14
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v14, v32, v14, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v13
; GFX7-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:56
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:184
; GFX7-NEXT: v_and_b32_e32 v14, 0xffff0000, v14
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v13, 1.0, v13
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v13, v32, v13, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v12
; GFX7-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:52
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:180
; GFX7-NEXT: v_and_b32_e32 v13, 0xffff0000, v13
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v12, 1.0, v12
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v12, v32, v12, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v11
; GFX7-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:48
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:176
; GFX7-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v11, v32, v11, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v10
; GFX7-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:44
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:172
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v10, v32, v10, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v9
; GFX7-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:168
; GFX7-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v9, v32, v9, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v8
; GFX7-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:36
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:164
; GFX7-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v8, v32, v8, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v7
; GFX7-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:160
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v7, v32, v7, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v6
; GFX7-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:28
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:156
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v6, v32, v6, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v5
; GFX7-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:152
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v5, v32, v5, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v4
; GFX7-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:20
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:148
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v4, v32, v4, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v3
; GFX7-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:144
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v3, v32, v3, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v2
; GFX7-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:12
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:140
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v2, v32, v2, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX7-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:136
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v1, v32, v1, vcc
; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX7-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:4
; GFX7-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_waitcnt vmcnt(1)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: s_waitcnt vmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v32, 1.0, v32
; GFX7-NEXT: v_cndmask_b32_e32 v0, v32, v0, vcc
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_vselect_v32bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_store_dword v34, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: v_writelane_b32 v34, s30, 0
; GFX8-NEXT: v_writelane_b32 v34, s31, 1
; GFX8-NEXT: v_writelane_b32 v34, s34, 2
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_writelane_b32 v34, s35, 3
; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v1
; GFX8-NEXT: v_writelane_b32 v34, s36, 4
; GFX8-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v2
; GFX8-NEXT: v_writelane_b32 v34, s37, 5
; GFX8-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v3
; GFX8-NEXT: v_writelane_b32 v34, s38, 6
; GFX8-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v4
; GFX8-NEXT: v_writelane_b32 v34, s39, 7
; GFX8-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v5
; GFX8-NEXT: v_writelane_b32 v34, s40, 8
; GFX8-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v6
; GFX8-NEXT: v_writelane_b32 v34, s41, 9
; GFX8-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v7
; GFX8-NEXT: v_writelane_b32 v34, s42, 10
; GFX8-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v8
; GFX8-NEXT: v_writelane_b32 v34, s43, 11
; GFX8-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v9
; GFX8-NEXT: v_writelane_b32 v34, s44, 12
; GFX8-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v10
; GFX8-NEXT: v_writelane_b32 v34, s45, 13
; GFX8-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v11
; GFX8-NEXT: v_writelane_b32 v34, s46, 14
; GFX8-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v12
; GFX8-NEXT: v_writelane_b32 v34, s47, 15
; GFX8-NEXT: v_cmp_eq_u32_e64 s[26:27], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v13
; GFX8-NEXT: v_writelane_b32 v34, s48, 16
; GFX8-NEXT: v_cmp_eq_u32_e64 s[28:29], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v14
; GFX8-NEXT: v_writelane_b32 v34, s49, 17
; GFX8-NEXT: v_cmp_eq_u32_e64 s[30:31], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v15
; GFX8-NEXT: v_writelane_b32 v34, s50, 18
; GFX8-NEXT: v_cmp_eq_u32_e64 s[34:35], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v16
; GFX8-NEXT: v_writelane_b32 v34, s51, 19
; GFX8-NEXT: v_cmp_eq_u32_e64 s[36:37], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v17
; GFX8-NEXT: v_writelane_b32 v34, s52, 20
; GFX8-NEXT: v_cmp_eq_u32_e64 s[38:39], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v18
; GFX8-NEXT: v_writelane_b32 v34, s53, 21
; GFX8-NEXT: v_cmp_eq_u32_e64 s[40:41], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v19
; GFX8-NEXT: v_writelane_b32 v34, s54, 22
; GFX8-NEXT: v_cmp_eq_u32_e64 s[42:43], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v20
; GFX8-NEXT: v_writelane_b32 v34, s55, 23
; GFX8-NEXT: v_cmp_eq_u32_e64 s[44:45], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v21
; GFX8-NEXT: v_writelane_b32 v34, s56, 24
; GFX8-NEXT: v_cmp_eq_u32_e64 s[46:47], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v22
; GFX8-NEXT: v_writelane_b32 v34, s57, 25
; GFX8-NEXT: v_cmp_eq_u32_e64 s[48:49], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v23
; GFX8-NEXT: v_writelane_b32 v34, s58, 26
; GFX8-NEXT: v_cmp_eq_u32_e64 s[50:51], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v24
; GFX8-NEXT: v_writelane_b32 v34, s59, 27
; GFX8-NEXT: v_cmp_eq_u32_e64 s[52:53], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v25
; GFX8-NEXT: v_writelane_b32 v34, s60, 28
; GFX8-NEXT: v_cmp_eq_u32_e64 s[54:55], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v26
; GFX8-NEXT: v_writelane_b32 v34, s61, 29
; GFX8-NEXT: v_cmp_eq_u32_e64 s[56:57], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v27
; GFX8-NEXT: v_writelane_b32 v34, s62, 30
; GFX8-NEXT: v_cmp_eq_u32_e64 s[58:59], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v28
; GFX8-NEXT: v_writelane_b32 v34, s63, 31
; GFX8-NEXT: v_cmp_eq_u32_e64 s[60:61], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v29
; GFX8-NEXT: v_writelane_b32 v34, s64, 32
; GFX8-NEXT: v_cmp_eq_u32_e64 s[62:63], 1, v0
; GFX8-NEXT: v_and_b32_e32 v0, 1, v30
; GFX8-NEXT: v_writelane_b32 v34, s65, 33
; GFX8-NEXT: v_cmp_eq_u32_e64 s[64:65], 1, v0
; GFX8-NEXT: buffer_load_ushort v0, off, s[0:3], s32
; GFX8-NEXT: v_writelane_b32 v34, s66, 34
; GFX8-NEXT: v_writelane_b32 v34, s67, 35
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_and_b32_e32 v0, 1, v0
; GFX8-NEXT: v_cmp_eq_u32_e64 s[66:67], 1, v0
; GFX8-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:68
; GFX8-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4
; GFX8-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72
; GFX8-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8
; GFX8-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76
; GFX8-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
; GFX8-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
; GFX8-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:16
; GFX8-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:84
; GFX8-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:20
; GFX8-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:88
; GFX8-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:24
; GFX8-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92
; GFX8-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:28
; GFX8-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:96
; GFX8-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:32
; GFX8-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:100
; GFX8-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:36
; GFX8-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:104
; GFX8-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:40
; GFX8-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:108
; GFX8-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:44
; GFX8-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:112
; GFX8-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:48
; GFX8-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116
; GFX8-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:52
; GFX8-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
; GFX8-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:56
; GFX8-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:124
; GFX8-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
; GFX8-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:128
; GFX8-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX8-NEXT: s_waitcnt vmcnt(1)
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v29
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: v_lshrrev_b32_e32 v28, 16, v32
; GFX8-NEXT: v_cndmask_b32_e64 v28, v33, v28, s[66:67]
; GFX8-NEXT: v_cndmask_b32_e64 v29, v29, v32, s[64:65]
; GFX8-NEXT: v_lshrrev_b32_e32 v32, 16, v31
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v30
; GFX8-NEXT: v_cndmask_b32_e64 v32, v33, v32, s[62:63]
; GFX8-NEXT: v_cndmask_b32_e64 v30, v30, v31, s[60:61]
; GFX8-NEXT: v_lshrrev_b32_e32 v31, 16, v27
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v26
; GFX8-NEXT: v_cndmask_b32_e64 v31, v33, v31, s[58:59]
; GFX8-NEXT: v_cndmask_b32_e64 v26, v26, v27, s[56:57]
; GFX8-NEXT: v_lshrrev_b32_e32 v27, 16, v25
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v24
; GFX8-NEXT: v_cndmask_b32_e64 v27, v33, v27, s[54:55]
; GFX8-NEXT: v_cndmask_b32_e64 v24, v24, v25, s[52:53]
; GFX8-NEXT: v_lshrrev_b32_e32 v25, 16, v23
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v22
; GFX8-NEXT: v_cndmask_b32_e64 v25, v33, v25, s[50:51]
; GFX8-NEXT: v_cndmask_b32_e64 v22, v22, v23, s[48:49]
; GFX8-NEXT: v_lshrrev_b32_e32 v23, 16, v21
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v20
; GFX8-NEXT: v_cndmask_b32_e64 v23, v33, v23, s[46:47]
; GFX8-NEXT: v_cndmask_b32_e64 v20, v20, v21, s[44:45]
; GFX8-NEXT: v_lshrrev_b32_e32 v21, 16, v19
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v18
; GFX8-NEXT: v_cndmask_b32_e64 v21, v33, v21, s[42:43]
; GFX8-NEXT: v_cndmask_b32_e64 v18, v18, v19, s[40:41]
; GFX8-NEXT: v_lshrrev_b32_e32 v19, 16, v17
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v16
; GFX8-NEXT: v_cndmask_b32_e64 v19, v33, v19, s[38:39]
; GFX8-NEXT: v_cndmask_b32_e64 v16, v16, v17, s[36:37]
; GFX8-NEXT: v_lshrrev_b32_e32 v17, 16, v15
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v14
; GFX8-NEXT: v_cndmask_b32_e64 v17, v33, v17, s[34:35]
; GFX8-NEXT: v_cndmask_b32_e64 v14, v14, v15, s[30:31]
; GFX8-NEXT: v_lshrrev_b32_e32 v15, 16, v13
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v12
; GFX8-NEXT: v_cndmask_b32_e64 v15, v33, v15, s[28:29]
; GFX8-NEXT: v_cndmask_b32_e64 v12, v12, v13, s[26:27]
; GFX8-NEXT: v_lshrrev_b32_e32 v13, 16, v11
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v10
; GFX8-NEXT: v_cndmask_b32_e64 v13, v33, v13, s[24:25]
; GFX8-NEXT: v_cndmask_b32_e64 v10, v10, v11, s[22:23]
; GFX8-NEXT: v_lshrrev_b32_e32 v11, 16, v9
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v8
; GFX8-NEXT: v_cndmask_b32_e64 v11, v33, v11, s[20:21]
; GFX8-NEXT: v_cndmask_b32_e64 v8, v8, v9, s[18:19]
; GFX8-NEXT: v_lshrrev_b32_e32 v9, 16, v7
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v6
; GFX8-NEXT: v_cndmask_b32_e64 v9, v33, v9, s[16:17]
; GFX8-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[14:15]
; GFX8-NEXT: v_lshrrev_b32_e32 v7, 16, v5
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v4
; GFX8-NEXT: v_cndmask_b32_e64 v7, v33, v7, s[12:13]
; GFX8-NEXT: v_cndmask_b32_e64 v4, v4, v5, s[10:11]
; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v3
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v2
; GFX8-NEXT: v_cndmask_b32_e64 v5, v33, v5, s[8:9]
; GFX8-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[6:7]
; GFX8-NEXT: v_lshrrev_b32_e32 v3, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v33, 16, v0
; GFX8-NEXT: v_cndmask_b32_e64 v3, v33, v3, s[4:5]
; GFX8-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v3
; GFX8-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v5
; GFX8-NEXT: v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v7
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v9
; GFX8-NEXT: v_or_b32_sdwa v2, v4, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v3, v6, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v11
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v13
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v15
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v17
; GFX8-NEXT: v_or_b32_sdwa v4, v8, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v5, v10, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v6, v12, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v7, v14, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v19
; GFX8-NEXT: v_lshlrev_b32_e32 v9, 16, v21
; GFX8-NEXT: v_lshlrev_b32_e32 v10, 16, v23
; GFX8-NEXT: v_lshlrev_b32_e32 v11, 16, v25
; GFX8-NEXT: v_lshlrev_b32_e32 v12, 16, v27
; GFX8-NEXT: v_lshlrev_b32_e32 v13, 16, v31
; GFX8-NEXT: v_lshlrev_b32_e32 v14, 16, v32
; GFX8-NEXT: v_lshlrev_b32_e32 v15, 16, v28
; GFX8-NEXT: v_or_b32_sdwa v8, v16, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v9, v18, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v11, v22, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v12, v24, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v13, v26, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v14, v30, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_or_b32_sdwa v15, v29, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX8-NEXT: v_readlane_b32 s67, v34, 35
; GFX8-NEXT: v_readlane_b32 s66, v34, 34
; GFX8-NEXT: v_readlane_b32 s65, v34, 33
; GFX8-NEXT: v_readlane_b32 s64, v34, 32
; GFX8-NEXT: v_readlane_b32 s63, v34, 31
; GFX8-NEXT: v_readlane_b32 s62, v34, 30
; GFX8-NEXT: v_readlane_b32 s61, v34, 29
; GFX8-NEXT: v_readlane_b32 s60, v34, 28
; GFX8-NEXT: v_readlane_b32 s59, v34, 27
; GFX8-NEXT: v_readlane_b32 s58, v34, 26
; GFX8-NEXT: v_readlane_b32 s57, v34, 25
; GFX8-NEXT: v_readlane_b32 s56, v34, 24
; GFX8-NEXT: v_readlane_b32 s55, v34, 23
; GFX8-NEXT: v_readlane_b32 s54, v34, 22
; GFX8-NEXT: v_readlane_b32 s53, v34, 21
; GFX8-NEXT: v_readlane_b32 s52, v34, 20
; GFX8-NEXT: v_readlane_b32 s51, v34, 19
; GFX8-NEXT: v_readlane_b32 s50, v34, 18
; GFX8-NEXT: v_readlane_b32 s49, v34, 17
; GFX8-NEXT: v_readlane_b32 s48, v34, 16
; GFX8-NEXT: v_readlane_b32 s47, v34, 15
; GFX8-NEXT: v_readlane_b32 s46, v34, 14
; GFX8-NEXT: v_readlane_b32 s45, v34, 13
; GFX8-NEXT: v_readlane_b32 s44, v34, 12
; GFX8-NEXT: v_readlane_b32 s43, v34, 11
; GFX8-NEXT: v_readlane_b32 s42, v34, 10
; GFX8-NEXT: v_readlane_b32 s41, v34, 9
; GFX8-NEXT: v_readlane_b32 s40, v34, 8
; GFX8-NEXT: v_readlane_b32 s39, v34, 7
; GFX8-NEXT: v_readlane_b32 s38, v34, 6
; GFX8-NEXT: v_readlane_b32 s37, v34, 5
; GFX8-NEXT: v_readlane_b32 s36, v34, 4
; GFX8-NEXT: v_readlane_b32 s35, v34, 3
; GFX8-NEXT: v_readlane_b32 s34, v34, 2
; GFX8-NEXT: v_readlane_b32 s31, v34, 1
; GFX8-NEXT: v_readlane_b32 s30, v34, 0
; GFX8-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX8-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX8-NEXT: s_mov_b64 exec, s[4:5]
; GFX8-NEXT: s_waitcnt vmcnt(0)
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_vselect_v32bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: v_writelane_b32 v33, s30, 0
; GFX9-NEXT: v_writelane_b32 v33, s31, 1
; GFX9-NEXT: v_writelane_b32 v33, s34, 2
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_writelane_b32 v33, s35, 3
; GFX9-NEXT: v_cmp_eq_u32_e64 s[4:5], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v3
; GFX9-NEXT: v_writelane_b32 v33, s36, 4
; GFX9-NEXT: v_cmp_eq_u32_e64 s[6:7], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v2
; GFX9-NEXT: v_writelane_b32 v33, s37, 5
; GFX9-NEXT: v_cmp_eq_u32_e64 s[8:9], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v5
; GFX9-NEXT: v_writelane_b32 v33, s38, 6
; GFX9-NEXT: v_cmp_eq_u32_e64 s[10:11], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v4
; GFX9-NEXT: v_writelane_b32 v33, s39, 7
; GFX9-NEXT: v_cmp_eq_u32_e64 s[12:13], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v7
; GFX9-NEXT: v_writelane_b32 v33, s40, 8
; GFX9-NEXT: v_cmp_eq_u32_e64 s[14:15], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v6
; GFX9-NEXT: v_writelane_b32 v33, s41, 9
; GFX9-NEXT: v_cmp_eq_u32_e64 s[16:17], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v9
; GFX9-NEXT: v_writelane_b32 v33, s42, 10
; GFX9-NEXT: v_cmp_eq_u32_e64 s[18:19], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v8
; GFX9-NEXT: v_writelane_b32 v33, s43, 11
; GFX9-NEXT: v_cmp_eq_u32_e64 s[20:21], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v11
; GFX9-NEXT: v_writelane_b32 v33, s44, 12
; GFX9-NEXT: v_cmp_eq_u32_e64 s[22:23], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v10
; GFX9-NEXT: v_writelane_b32 v33, s45, 13
; GFX9-NEXT: v_cmp_eq_u32_e64 s[24:25], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v13
; GFX9-NEXT: v_writelane_b32 v33, s46, 14
; GFX9-NEXT: v_cmp_eq_u32_e64 s[26:27], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v12
; GFX9-NEXT: v_writelane_b32 v33, s47, 15
; GFX9-NEXT: v_cmp_eq_u32_e64 s[28:29], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v15
; GFX9-NEXT: v_writelane_b32 v33, s48, 16
; GFX9-NEXT: v_cmp_eq_u32_e64 s[30:31], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v14
; GFX9-NEXT: v_writelane_b32 v33, s49, 17
; GFX9-NEXT: v_cmp_eq_u32_e64 s[34:35], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v17
; GFX9-NEXT: v_writelane_b32 v33, s50, 18
; GFX9-NEXT: v_cmp_eq_u32_e64 s[36:37], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v16
; GFX9-NEXT: v_writelane_b32 v33, s51, 19
; GFX9-NEXT: v_cmp_eq_u32_e64 s[38:39], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v19
; GFX9-NEXT: v_writelane_b32 v33, s52, 20
; GFX9-NEXT: v_cmp_eq_u32_e64 s[40:41], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v18
; GFX9-NEXT: v_writelane_b32 v33, s53, 21
; GFX9-NEXT: v_cmp_eq_u32_e64 s[42:43], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v21
; GFX9-NEXT: v_writelane_b32 v33, s54, 22
; GFX9-NEXT: v_cmp_eq_u32_e64 s[44:45], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v20
; GFX9-NEXT: v_writelane_b32 v33, s55, 23
; GFX9-NEXT: v_cmp_eq_u32_e64 s[46:47], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v23
; GFX9-NEXT: v_writelane_b32 v33, s56, 24
; GFX9-NEXT: v_cmp_eq_u32_e64 s[48:49], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v22
; GFX9-NEXT: v_writelane_b32 v33, s57, 25
; GFX9-NEXT: v_cmp_eq_u32_e64 s[50:51], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v25
; GFX9-NEXT: v_writelane_b32 v33, s58, 26
; GFX9-NEXT: v_cmp_eq_u32_e64 s[52:53], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v24
; GFX9-NEXT: v_writelane_b32 v33, s59, 27
; GFX9-NEXT: v_cmp_eq_u32_e64 s[54:55], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v27
; GFX9-NEXT: v_writelane_b32 v33, s60, 28
; GFX9-NEXT: v_cmp_eq_u32_e64 s[56:57], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v26
; GFX9-NEXT: v_writelane_b32 v33, s61, 29
; GFX9-NEXT: v_cmp_eq_u32_e64 s[58:59], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v29
; GFX9-NEXT: v_writelane_b32 v33, s62, 30
; GFX9-NEXT: v_cmp_eq_u32_e64 s[60:61], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v28
; GFX9-NEXT: v_writelane_b32 v33, s63, 31
; GFX9-NEXT: v_cmp_eq_u32_e64 s[62:63], 1, v0
; GFX9-NEXT: buffer_load_ushort v0, off, s[0:3], s32
; GFX9-NEXT: v_writelane_b32 v33, s64, 32
; GFX9-NEXT: v_writelane_b32 v33, s65, 33
; GFX9-NEXT: v_writelane_b32 v33, s66, 34
; GFX9-NEXT: v_and_b32_e32 v1, 1, v1
; GFX9-NEXT: v_writelane_b32 v33, s67, 35
; GFX9-NEXT: v_cmp_eq_u32_e32 vcc, 1, v1
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_and_b32_e32 v0, 1, v0
; GFX9-NEXT: v_cmp_eq_u32_e64 s[64:65], 1, v0
; GFX9-NEXT: v_and_b32_e32 v0, 1, v30
; GFX9-NEXT: v_cmp_eq_u32_e64 s[66:67], 1, v0
; GFX9-NEXT: buffer_load_dword v0, off, s[0:3], s32 offset:68
; GFX9-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:4
; GFX9-NEXT: buffer_load_dword v2, off, s[0:3], s32 offset:72
; GFX9-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:8
; GFX9-NEXT: buffer_load_dword v4, off, s[0:3], s32 offset:76
; GFX9-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:12
; GFX9-NEXT: buffer_load_dword v6, off, s[0:3], s32 offset:80
; GFX9-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:16
; GFX9-NEXT: buffer_load_dword v8, off, s[0:3], s32 offset:84
; GFX9-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:20
; GFX9-NEXT: buffer_load_dword v10, off, s[0:3], s32 offset:88
; GFX9-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:24
; GFX9-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:92
; GFX9-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:28
; GFX9-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:96
; GFX9-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:32
; GFX9-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:100
; GFX9-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:36
; GFX9-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:104
; GFX9-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:40
; GFX9-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:108
; GFX9-NEXT: buffer_load_dword v21, off, s[0:3], s32 offset:44
; GFX9-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:112
; GFX9-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:48
; GFX9-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:116
; GFX9-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:52
; GFX9-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:120
; GFX9-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:56
; GFX9-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:124
; GFX9-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:60
; GFX9-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:128
; GFX9-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:64
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: v_cndmask_b32_e64 v29, v31, v32, s[66:67]
; GFX9-NEXT: v_lshrrev_b32_e32 v32, 16, v32
; GFX9-NEXT: v_lshrrev_b32_e32 v31, 16, v31
; GFX9-NEXT: v_cndmask_b32_e64 v31, v31, v32, s[64:65]
; GFX9-NEXT: v_cndmask_b32_e64 v32, v28, v30, s[62:63]
; GFX9-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GFX9-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GFX9-NEXT: v_cndmask_b32_e64 v28, v28, v30, s[60:61]
; GFX9-NEXT: v_cndmask_b32_e64 v30, v26, v27, s[58:59]
; GFX9-NEXT: v_lshrrev_b32_e32 v27, 16, v27
; GFX9-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GFX9-NEXT: v_cndmask_b32_e64 v26, v26, v27, s[56:57]
; GFX9-NEXT: v_cndmask_b32_e64 v27, v24, v25, s[54:55]
; GFX9-NEXT: v_lshrrev_b32_e32 v25, 16, v25
; GFX9-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GFX9-NEXT: v_cndmask_b32_e64 v24, v24, v25, s[52:53]
; GFX9-NEXT: v_cndmask_b32_e64 v25, v22, v23, s[50:51]
; GFX9-NEXT: v_lshrrev_b32_e32 v23, 16, v23
; GFX9-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX9-NEXT: v_cndmask_b32_e64 v22, v22, v23, s[48:49]
; GFX9-NEXT: v_cndmask_b32_e64 v23, v20, v21, s[46:47]
; GFX9-NEXT: v_lshrrev_b32_e32 v21, 16, v21
; GFX9-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GFX9-NEXT: v_cndmask_b32_e64 v20, v20, v21, s[44:45]
; GFX9-NEXT: v_cndmask_b32_e64 v21, v18, v19, s[42:43]
; GFX9-NEXT: v_lshrrev_b32_e32 v19, 16, v19
; GFX9-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT: v_cndmask_b32_e64 v18, v18, v19, s[40:41]
; GFX9-NEXT: v_cndmask_b32_e64 v19, v16, v17, s[38:39]
; GFX9-NEXT: v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX9-NEXT: v_cndmask_b32_e64 v16, v16, v17, s[36:37]
; GFX9-NEXT: v_cndmask_b32_e64 v17, v14, v15, s[34:35]
; GFX9-NEXT: v_lshrrev_b32_e32 v15, 16, v15
; GFX9-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX9-NEXT: v_cndmask_b32_e64 v14, v14, v15, s[30:31]
; GFX9-NEXT: v_cndmask_b32_e64 v15, v12, v13, s[28:29]
; GFX9-NEXT: v_lshrrev_b32_e32 v13, 16, v13
; GFX9-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX9-NEXT: v_cndmask_b32_e64 v12, v12, v13, s[26:27]
; GFX9-NEXT: v_cndmask_b32_e64 v13, v10, v11, s[24:25]
; GFX9-NEXT: v_lshrrev_b32_e32 v11, 16, v11
; GFX9-NEXT: v_lshrrev_b32_e32 v10, 16, v10
; GFX9-NEXT: v_cndmask_b32_e64 v10, v10, v11, s[22:23]
; GFX9-NEXT: v_cndmask_b32_e64 v11, v8, v9, s[20:21]
; GFX9-NEXT: v_lshrrev_b32_e32 v9, 16, v9
; GFX9-NEXT: v_lshrrev_b32_e32 v8, 16, v8
; GFX9-NEXT: v_cndmask_b32_e64 v8, v8, v9, s[18:19]
; GFX9-NEXT: v_cndmask_b32_e64 v9, v6, v7, s[16:17]
; GFX9-NEXT: v_lshrrev_b32_e32 v7, 16, v7
; GFX9-NEXT: v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT: v_cndmask_b32_e64 v6, v6, v7, s[14:15]
; GFX9-NEXT: v_cndmask_b32_e64 v7, v4, v5, s[12:13]
; GFX9-NEXT: v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT: v_lshrrev_b32_e32 v4, 16, v4
; GFX9-NEXT: v_cndmask_b32_e64 v4, v4, v5, s[10:11]
; GFX9-NEXT: v_cndmask_b32_e64 v5, v2, v3, s[8:9]
; GFX9-NEXT: v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[6:7]
; GFX9-NEXT: v_cndmask_b32_e64 v3, v0, v1, s[4:5]
; GFX9-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x5040100
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v2, v5, s4
; GFX9-NEXT: v_perm_b32 v2, v4, v7, s4
; GFX9-NEXT: v_perm_b32 v3, v6, v9, s4
; GFX9-NEXT: v_perm_b32 v4, v8, v11, s4
; GFX9-NEXT: v_perm_b32 v5, v10, v13, s4
; GFX9-NEXT: v_perm_b32 v6, v12, v15, s4
; GFX9-NEXT: v_perm_b32 v7, v14, v17, s4
; GFX9-NEXT: v_perm_b32 v8, v16, v19, s4
; GFX9-NEXT: v_perm_b32 v9, v18, v21, s4
; GFX9-NEXT: v_perm_b32 v10, v20, v23, s4
; GFX9-NEXT: v_perm_b32 v11, v22, v25, s4
; GFX9-NEXT: v_perm_b32 v12, v24, v27, s4
; GFX9-NEXT: v_perm_b32 v13, v26, v30, s4
; GFX9-NEXT: v_perm_b32 v14, v28, v32, s4
; GFX9-NEXT: v_perm_b32 v15, v31, v29, s4
; GFX9-NEXT: v_readlane_b32 s67, v33, 35
; GFX9-NEXT: v_readlane_b32 s66, v33, 34
; GFX9-NEXT: v_readlane_b32 s65, v33, 33
; GFX9-NEXT: v_readlane_b32 s64, v33, 32
; GFX9-NEXT: v_readlane_b32 s63, v33, 31
; GFX9-NEXT: v_readlane_b32 s62, v33, 30
; GFX9-NEXT: v_readlane_b32 s61, v33, 29
; GFX9-NEXT: v_readlane_b32 s60, v33, 28
; GFX9-NEXT: v_readlane_b32 s59, v33, 27
; GFX9-NEXT: v_readlane_b32 s58, v33, 26
; GFX9-NEXT: v_readlane_b32 s57, v33, 25
; GFX9-NEXT: v_readlane_b32 s56, v33, 24
; GFX9-NEXT: v_readlane_b32 s55, v33, 23
; GFX9-NEXT: v_readlane_b32 s54, v33, 22
; GFX9-NEXT: v_readlane_b32 s53, v33, 21
; GFX9-NEXT: v_readlane_b32 s52, v33, 20
; GFX9-NEXT: v_readlane_b32 s51, v33, 19
; GFX9-NEXT: v_readlane_b32 s50, v33, 18
; GFX9-NEXT: v_readlane_b32 s49, v33, 17
; GFX9-NEXT: v_readlane_b32 s48, v33, 16
; GFX9-NEXT: v_readlane_b32 s47, v33, 15
; GFX9-NEXT: v_readlane_b32 s46, v33, 14
; GFX9-NEXT: v_readlane_b32 s45, v33, 13
; GFX9-NEXT: v_readlane_b32 s44, v33, 12
; GFX9-NEXT: v_readlane_b32 s43, v33, 11
; GFX9-NEXT: v_readlane_b32 s42, v33, 10
; GFX9-NEXT: v_readlane_b32 s41, v33, 9
; GFX9-NEXT: v_readlane_b32 s40, v33, 8
; GFX9-NEXT: v_readlane_b32 s39, v33, 7
; GFX9-NEXT: v_readlane_b32 s38, v33, 6
; GFX9-NEXT: v_readlane_b32 s37, v33, 5
; GFX9-NEXT: v_readlane_b32 s36, v33, 4
; GFX9-NEXT: v_readlane_b32 s35, v33, 3
; GFX9-NEXT: v_readlane_b32 s34, v33, 2
; GFX9-NEXT: v_readlane_b32 s31, v33, 1
; GFX9-NEXT: v_readlane_b32 s30, v33, 0
; GFX9-NEXT: s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT: s_mov_b64 exec, s[4:5]
; GFX9-NEXT: s_waitcnt vmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_vselect_v32bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_or_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: v_and_b32_e32 v29, 1, v29
; GFX10-NEXT: v_and_b32_e32 v30, 1, v30
; GFX10-NEXT: v_and_b32_e32 v28, 1, v28
; GFX10-NEXT: v_and_b32_e32 v26, 1, v26
; GFX10-NEXT: v_and_b32_e32 v24, 1, v24
; GFX10-NEXT: v_and_b32_e32 v22, 1, v22
; GFX10-NEXT: v_and_b32_e32 v20, 1, v20
; GFX10-NEXT: v_and_b32_e32 v18, 1, v18
; GFX10-NEXT: v_and_b32_e32 v16, 1, v16
; GFX10-NEXT: v_and_b32_e32 v14, 1, v14
; GFX10-NEXT: v_and_b32_e32 v12, 1, v12
; GFX10-NEXT: s_clause 0x14
; GFX10-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
; GFX10-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:124
; GFX10-NEXT: buffer_load_ushort v33, off, s[0:3], s32
; GFX10-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:128
; GFX10-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:64
; GFX10-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:48
; GFX10-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:116
; GFX10-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:52
; GFX10-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:120
; GFX10-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:56
; GFX10-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:32
; GFX10-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:100
; GFX10-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:36
; GFX10-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:104
; GFX10-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:40
; GFX10-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:108
; GFX10-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:44
; GFX10-NEXT: buffer_load_dword v64, off, s[0:3], s32 offset:112
; GFX10-NEXT: buffer_load_dword v65, off, s[0:3], s32 offset:72
; GFX10-NEXT: buffer_load_dword v66, off, s[0:3], s32 offset:76
; GFX10-NEXT: buffer_load_dword v67, off, s[0:3], s32 offset:80
; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v29
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:92
; GFX10-NEXT: buffer_load_dword v68, off, s[0:3], s32 offset:28
; GFX10-NEXT: v_cmp_eq_u32_e64 s4, 1, v30
; GFX10-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:96
; GFX10-NEXT: v_cmp_eq_u32_e64 s5, 1, v28
; GFX10-NEXT: buffer_load_dword v28, off, s[0:3], s32 offset:88
; GFX10-NEXT: v_cmp_eq_u32_e64 s6, 1, v26
; GFX10-NEXT: v_cmp_eq_u32_e64 s7, 1, v24
; GFX10-NEXT: buffer_load_dword v24, off, s[0:3], s32 offset:84
; GFX10-NEXT: v_cmp_eq_u32_e64 s8, 1, v22
; GFX10-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:20
; GFX10-NEXT: v_cmp_eq_u32_e64 s9, 1, v20
; GFX10-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:16
; GFX10-NEXT: v_cmp_eq_u32_e64 s10, 1, v18
; GFX10-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:12
; GFX10-NEXT: v_cmp_eq_u32_e64 s11, 1, v16
; GFX10-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:8
; GFX10-NEXT: v_cmp_eq_u32_e64 s12, 1, v14
; GFX10-NEXT: s_clause 0x1
; GFX10-NEXT: buffer_load_dword v14, off, s[0:3], s32 offset:68
; GFX10-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:24
; GFX10-NEXT: v_cmp_eq_u32_e64 s13, 1, v12
; GFX10-NEXT: buffer_load_dword v12, off, s[0:3], s32 offset:4
; GFX10-NEXT: v_writelane_b32 v40, s30, 0
; GFX10-NEXT: v_and_b32_e32 v0, 1, v0
; GFX10-NEXT: v_and_b32_e32 v2, 1, v2
; GFX10-NEXT: v_and_b32_e32 v4, 1, v4
; GFX10-NEXT: v_and_b32_e32 v6, 1, v6
; GFX10-NEXT: v_and_b32_e32 v8, 1, v8
; GFX10-NEXT: v_and_b32_e32 v10, 1, v10
; GFX10-NEXT: v_writelane_b32 v40, s31, 1
; GFX10-NEXT: v_and_b32_e32 v1, 1, v1
; GFX10-NEXT: v_and_b32_e32 v3, 1, v3
; GFX10-NEXT: v_and_b32_e32 v5, 1, v5
; GFX10-NEXT: v_and_b32_e32 v7, 1, v7
; GFX10-NEXT: v_and_b32_e32 v9, 1, v9
; GFX10-NEXT: v_and_b32_e32 v11, 1, v11
; GFX10-NEXT: v_and_b32_e32 v13, 1, v13
; GFX10-NEXT: v_and_b32_e32 v15, 1, v15
; GFX10-NEXT: v_and_b32_e32 v17, 1, v17
; GFX10-NEXT: v_and_b32_e32 v19, 1, v19
; GFX10-NEXT: v_and_b32_e32 v21, 1, v21
; GFX10-NEXT: v_and_b32_e32 v23, 1, v23
; GFX10-NEXT: v_and_b32_e32 v25, 1, v25
; GFX10-NEXT: v_and_b32_e32 v27, 1, v27
; GFX10-NEXT: v_cmp_eq_u32_e64 s14, 1, v10
; GFX10-NEXT: v_cmp_eq_u32_e64 s15, 1, v8
; GFX10-NEXT: v_cmp_eq_u32_e64 s16, 1, v6
; GFX10-NEXT: v_cmp_eq_u32_e64 s17, 1, v4
; GFX10-NEXT: v_cmp_eq_u32_e64 s18, 1, v2
; GFX10-NEXT: v_cmp_eq_u32_e64 s19, 1, v0
; GFX10-NEXT: v_writelane_b32 v40, s34, 2
; GFX10-NEXT: v_cmp_eq_u32_e64 s20, 1, v27
; GFX10-NEXT: v_cmp_eq_u32_e64 s21, 1, v25
; GFX10-NEXT: v_cmp_eq_u32_e64 s22, 1, v23
; GFX10-NEXT: v_cmp_eq_u32_e64 s23, 1, v21
; GFX10-NEXT: v_cmp_eq_u32_e64 s24, 1, v19
; GFX10-NEXT: v_cmp_eq_u32_e64 s25, 1, v17
; GFX10-NEXT: v_cmp_eq_u32_e64 s26, 1, v15
; GFX10-NEXT: v_cmp_eq_u32_e64 s27, 1, v13
; GFX10-NEXT: v_cmp_eq_u32_e64 s28, 1, v11
; GFX10-NEXT: v_cmp_eq_u32_e64 s29, 1, v7
; GFX10-NEXT: v_cmp_eq_u32_e64 vcc_hi, 1, v3
; GFX10-NEXT: v_cmp_eq_u32_e64 s30, 1, v1
; GFX10-NEXT: v_cmp_eq_u32_e64 s31, 1, v5
; GFX10-NEXT: v_cmp_eq_u32_e64 s34, 1, v9
; GFX10-NEXT: s_waitcnt vmcnt(32)
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v31
; GFX10-NEXT: s_waitcnt vmcnt(31)
; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v32
; GFX10-NEXT: s_waitcnt vmcnt(30)
; GFX10-NEXT: v_and_b32_e32 v2, 1, v33
; GFX10-NEXT: s_waitcnt vmcnt(29)
; GFX10-NEXT: v_lshrrev_b32_e32 v4, 16, v34
; GFX10-NEXT: s_waitcnt vmcnt(28)
; GFX10-NEXT: v_cndmask_b32_e64 v15, v34, v35, s4
; GFX10-NEXT: v_lshrrev_b32_e32 v3, 16, v35
; GFX10-NEXT: v_cndmask_b32_e64 v17, v32, v31, s5
; GFX10-NEXT: s_waitcnt vmcnt(25)
; GFX10-NEXT: v_cndmask_b32_e64 v19, v37, v38, s7
; GFX10-NEXT: s_waitcnt vmcnt(24)
; GFX10-NEXT: v_lshrrev_b32_e32 v6, 16, v39
; GFX10-NEXT: s_waitcnt vmcnt(23)
; GFX10-NEXT: v_cndmask_b32_e64 v13, v39, v48, s6
; GFX10-NEXT: v_lshrrev_b32_e32 v5, 16, v48
; GFX10-NEXT: v_lshrrev_b32_e32 v7, 16, v38
; GFX10-NEXT: v_lshrrev_b32_e32 v8, 16, v37
; GFX10-NEXT: v_lshrrev_b32_e32 v9, 16, v36
; GFX10-NEXT: s_waitcnt vmcnt(18)
; GFX10-NEXT: v_cndmask_b32_e64 v27, v52, v53, s10
; GFX10-NEXT: s_waitcnt vmcnt(17)
; GFX10-NEXT: v_lshrrev_b32_e32 v25, 16, v54
; GFX10-NEXT: s_waitcnt vmcnt(16)
; GFX10-NEXT: v_cndmask_b32_e64 v21, v54, v55, s9
; GFX10-NEXT: s_waitcnt vmcnt(15)
; GFX10-NEXT: v_cndmask_b32_e64 v11, v64, v36, s8
; GFX10-NEXT: v_lshrrev_b32_e32 v10, 16, v64
; GFX10-NEXT: v_lshrrev_b32_e32 v23, 16, v55
; GFX10-NEXT: v_lshrrev_b32_e32 v31, 16, v53
; GFX10-NEXT: v_lshrrev_b32_e32 v32, 16, v52
; GFX10-NEXT: v_cndmask_b32_e64 v33, v50, v51, s11
; GFX10-NEXT: v_lshrrev_b32_e32 v34, 16, v51
; GFX10-NEXT: v_lshrrev_b32_e32 v35, 16, v50
; GFX10-NEXT: s_waitcnt vmcnt(9)
; GFX10-NEXT: v_cndmask_b32_e64 v36, v30, v49, s12
; GFX10-NEXT: v_lshrrev_b32_e32 v37, 16, v49
; GFX10-NEXT: v_lshrrev_b32_e32 v30, 16, v30
; GFX10-NEXT: v_cndmask_b32_e64 v38, v29, v68, s13
; GFX10-NEXT: v_lshrrev_b32_e32 v39, 16, v68
; GFX10-NEXT: v_lshrrev_b32_e32 v29, 16, v29
; GFX10-NEXT: s_waitcnt vmcnt(6)
; GFX10-NEXT: v_cndmask_b32_e64 v49, v24, v22, s15
; GFX10-NEXT: v_lshrrev_b32_e32 v22, 16, v22
; GFX10-NEXT: v_lshrrev_b32_e32 v24, 16, v24
; GFX10-NEXT: s_waitcnt vmcnt(5)
; GFX10-NEXT: v_cndmask_b32_e64 v50, v67, v20, s16
; GFX10-NEXT: v_lshrrev_b32_e32 v20, 16, v20
; GFX10-NEXT: v_lshrrev_b32_e32 v51, 16, v67
; GFX10-NEXT: s_waitcnt vmcnt(4)
; GFX10-NEXT: v_cndmask_b32_e64 v52, v66, v18, s17
; GFX10-NEXT: v_lshrrev_b32_e32 v18, 16, v18
; GFX10-NEXT: s_waitcnt vmcnt(1)
; GFX10-NEXT: v_cndmask_b32_e64 v48, v28, v26, s14
; GFX10-NEXT: v_lshrrev_b32_e32 v26, 16, v26
; GFX10-NEXT: v_lshrrev_b32_e32 v28, 16, v28
; GFX10-NEXT: v_lshrrev_b32_e32 v53, 16, v66
; GFX10-NEXT: v_cndmask_b32_e64 v54, v65, v16, s18
; GFX10-NEXT: v_lshrrev_b32_e32 v16, 16, v16
; GFX10-NEXT: v_lshrrev_b32_e32 v55, 16, v65
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: v_cndmask_b32_e64 v64, v14, v12, s19
; GFX10-NEXT: v_lshrrev_b32_e32 v12, 16, v12
; GFX10-NEXT: v_lshrrev_b32_e32 v14, 16, v14
; GFX10-NEXT: v_cmp_eq_u32_e64 s4, 1, v2
; GFX10-NEXT: v_cndmask_b32_e32 v65, v1, v0, vcc_lo
; GFX10-NEXT: v_cndmask_b32_e64 v66, v6, v5, s20
; GFX10-NEXT: v_cndmask_b32_e64 v67, v8, v7, s21
; GFX10-NEXT: v_cndmask_b32_e64 v68, v10, v9, s22
; GFX10-NEXT: v_cndmask_b32_e64 v10, v25, v23, s23
; GFX10-NEXT: v_cndmask_b32_e64 v9, v32, v31, s24
; GFX10-NEXT: v_cndmask_b32_e64 v8, v35, v34, s25
; GFX10-NEXT: v_cndmask_b32_e64 v7, v30, v37, s26
; GFX10-NEXT: v_cndmask_b32_e64 v6, v29, v39, s27
; GFX10-NEXT: v_cndmask_b32_e64 v5, v28, v26, s28
; GFX10-NEXT: v_cndmask_b32_e64 v20, v51, v20, s29
; GFX10-NEXT: v_cndmask_b32_e64 v0, v14, v12, s30
; GFX10-NEXT: v_cndmask_b32_e64 v1, v55, v16, vcc_hi
; GFX10-NEXT: v_cndmask_b32_e64 v2, v53, v18, s31
; GFX10-NEXT: v_cndmask_b32_e64 v12, v24, v22, s34
; GFX10-NEXT: v_cndmask_b32_e64 v16, v4, v3, s4
; GFX10-NEXT: v_perm_b32 v0, v0, v64, 0x5040100
; GFX10-NEXT: v_perm_b32 v1, v1, v54, 0x5040100
; GFX10-NEXT: v_perm_b32 v2, v2, v52, 0x5040100
; GFX10-NEXT: v_perm_b32 v3, v20, v50, 0x5040100
; GFX10-NEXT: v_perm_b32 v4, v12, v49, 0x5040100
; GFX10-NEXT: v_perm_b32 v5, v5, v48, 0x5040100
; GFX10-NEXT: v_perm_b32 v6, v6, v38, 0x5040100
; GFX10-NEXT: v_perm_b32 v7, v7, v36, 0x5040100
; GFX10-NEXT: v_perm_b32 v8, v8, v33, 0x5040100
; GFX10-NEXT: v_perm_b32 v9, v9, v27, 0x5040100
; GFX10-NEXT: v_perm_b32 v10, v10, v21, 0x5040100
; GFX10-NEXT: v_perm_b32 v11, v68, v11, 0x5040100
; GFX10-NEXT: v_perm_b32 v12, v67, v19, 0x5040100
; GFX10-NEXT: v_perm_b32 v13, v66, v13, 0x5040100
; GFX10-NEXT: v_perm_b32 v14, v65, v17, 0x5040100
; GFX10-NEXT: v_perm_b32 v15, v16, v15, 0x5040100
; GFX10-NEXT: v_readlane_b32 s34, v40, 2
; GFX10-NEXT: v_readlane_b32 s31, v40, 1
; GFX10-NEXT: v_readlane_b32 s30, v40, 0
; GFX10-NEXT: s_or_saveexec_b32 s4, -1
; GFX10-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX10-NEXT: s_waitcnt_depctr 0xffe3
; GFX10-NEXT: s_mov_b32 exec_lo, s4
; GFX10-NEXT: s_waitcnt vmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_vselect_v32bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: s_clause 0x1f
; GFX11TRUE16-NEXT: scratch_load_u16 v31, off, s32
; GFX11TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:128
; GFX11TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:64
; GFX11TRUE16-NEXT: scratch_load_b32 v34, off, s32 offset:124
; GFX11TRUE16-NEXT: scratch_load_b32 v35, off, s32 offset:60
; GFX11TRUE16-NEXT: scratch_load_b32 v36, off, s32 offset:120
; GFX11TRUE16-NEXT: scratch_load_b32 v37, off, s32 offset:56
; GFX11TRUE16-NEXT: scratch_load_b32 v38, off, s32 offset:116
; GFX11TRUE16-NEXT: scratch_load_b32 v39, off, s32 offset:52
; GFX11TRUE16-NEXT: scratch_load_b32 v48, off, s32 offset:112
; GFX11TRUE16-NEXT: scratch_load_b32 v49, off, s32 offset:48
; GFX11TRUE16-NEXT: scratch_load_b32 v50, off, s32 offset:108
; GFX11TRUE16-NEXT: scratch_load_b32 v51, off, s32 offset:44
; GFX11TRUE16-NEXT: scratch_load_b32 v52, off, s32 offset:104
; GFX11TRUE16-NEXT: scratch_load_b32 v53, off, s32 offset:40
; GFX11TRUE16-NEXT: scratch_load_b32 v54, off, s32 offset:100
; GFX11TRUE16-NEXT: scratch_load_b32 v55, off, s32 offset:36
; GFX11TRUE16-NEXT: scratch_load_b32 v64, off, s32 offset:96
; GFX11TRUE16-NEXT: scratch_load_b32 v65, off, s32 offset:32
; GFX11TRUE16-NEXT: scratch_load_b32 v66, off, s32 offset:92
; GFX11TRUE16-NEXT: scratch_load_b32 v67, off, s32 offset:28
; GFX11TRUE16-NEXT: scratch_load_b32 v68, off, s32 offset:88
; GFX11TRUE16-NEXT: scratch_load_b32 v69, off, s32 offset:24
; GFX11TRUE16-NEXT: scratch_load_b32 v70, off, s32 offset:84
; GFX11TRUE16-NEXT: scratch_load_b32 v71, off, s32 offset:20
; GFX11TRUE16-NEXT: scratch_load_b32 v80, off, s32 offset:80
; GFX11TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:16
; GFX11TRUE16-NEXT: scratch_load_b32 v82, off, s32 offset:76
; GFX11TRUE16-NEXT: scratch_load_b32 v83, off, s32 offset:12
; GFX11TRUE16-NEXT: scratch_load_b32 v84, off, s32 offset:72
; GFX11TRUE16-NEXT: scratch_load_b32 v85, off, s32 offset:8
; GFX11TRUE16-NEXT: scratch_load_b32 v86, off, s32 offset:68
; GFX11TRUE16-NEXT: scratch_load_b32 v87, off, s32 offset:4
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(31)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v96.l, v32.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(30)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v97.l, v33.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(29)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v98.l, v34.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(28)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v99.l, v35.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(27)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v100.l, v36.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(26)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v101.l, v37.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(25)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v102.l, v38.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(24)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v103.l, v39.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(23)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v112.l, v48.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(22)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v113.l, v49.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(21)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v114.l, v50.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(20)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v115.l, v51.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(19)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v116.l, v52.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(18)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v117.l, v53.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(17)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v118.l, v54.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(16)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v119.l, v55.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(15)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v128.l, v64.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(14)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v129.l, v65.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(13)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v130.l, v66.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(12)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v131.l, v67.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(11)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v132.l, v68.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(10)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v133.l, v69.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(9)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v134.l, v70.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(8)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v135.l, v71.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v71, 16, v71
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v70, 16, v70
; GFX11TRUE16-NEXT: v_and_b32_e32 v30, 1, v30
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(5)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v146.l, v82.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(4)
; GFX11TRUE16-NEXT: v_mov_b16_e64 v147.l, v83.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v83, 16, v83
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v82, 16, v82
; GFX11TRUE16-NEXT: v_and_b32_e32 v28, 1, v28
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v30
; GFX11TRUE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(3)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v30.l, v84.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v84, 16, v84
; GFX11TRUE16-NEXT: v_and_b32_e32 v26, 1, v26
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v96, v96, v97, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v28
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(2)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v97.l, v85.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v85, 16, v85
; GFX11TRUE16-NEXT: v_and_b32_e32 v24, 1, v24
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v98, v98, v99, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v26
; GFX11TRUE16-NEXT: v_and_b32_e32 v7, 1, v7
; GFX11TRUE16-NEXT: v_mov_b16_e64 v144.l, v80.l
; GFX11TRUE16-NEXT: v_mov_b16_e64 v145.l, v81.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v81, 16, v81
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v26, v100, v101, vcc_lo
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v80, 16, v80
; GFX11TRUE16-NEXT: v_and_b32_e32 v22, 1, v22
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v24
; GFX11TRUE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v69, 16, v69
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v68, 16, v68
; GFX11TRUE16-NEXT: v_and_b32_e32 v20, 1, v20
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v24, v102, v103, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v22
; GFX11TRUE16-NEXT: v_and_b32_e32 v11, 1, v11
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v67, 16, v67
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v66, 16, v66
; GFX11TRUE16-NEXT: v_and_b32_e32 v18, 1, v18
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v22, v112, v113, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v20
; GFX11TRUE16-NEXT: v_and_b32_e32 v9, 1, v9
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v65, 16, v65
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v64, 16, v64
; GFX11TRUE16-NEXT: v_and_b32_e32 v16, 1, v16
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v20, v114, v115, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v18
; GFX11TRUE16-NEXT: v_and_b32_e32 v15, 1, v15
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v55, 16, v55
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v54, 16, v54
; GFX11TRUE16-NEXT: v_and_b32_e32 v14, 1, v14
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v18, v116, v117, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v16
; GFX11TRUE16-NEXT: v_and_b32_e32 v13, 1, v13
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v53, 16, v53
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v52, 16, v52
; GFX11TRUE16-NEXT: v_and_b32_e32 v12, 1, v12
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v16, v118, v119, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v14
; GFX11TRUE16-NEXT: v_and_b32_e32 v19, 1, v19
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v51, 16, v51
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v50, 16, v50
; GFX11TRUE16-NEXT: v_and_b32_e32 v10, 1, v10
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v14, v128, v129, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v12
; GFX11TRUE16-NEXT: v_and_b32_e32 v17, 1, v17
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v49, 16, v49
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v48, 16, v48
; GFX11TRUE16-NEXT: v_and_b32_e32 v8, 1, v8
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v12, v130, v131, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v10
; GFX11TRUE16-NEXT: v_and_b32_e32 v23, 1, v23
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v39, 16, v39
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v38, 16, v38
; GFX11TRUE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v10, v132, v133, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v8
; GFX11TRUE16-NEXT: v_and_b32_e32 v21, 1, v21
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v37, 16, v37
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v36, 16, v36
; GFX11TRUE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v8, v134, v135, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11TRUE16-NEXT: v_and_b32_e32 v27, 1, v27
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v35, 16, v35
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v34, 16, v34
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v6, v144, v145, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11TRUE16-NEXT: v_and_b32_e32 v25, 1, v25
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v33, 16, v33
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v32, 16, v32
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v4, v146, v147, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v31, 1, v31
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(1)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v28.l, v86.l
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0)
; GFX11TRUE16-NEXT: v_mov_b16_e32 v99.l, v87.l
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v87, 16, v87
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v30, v97, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v29, 1, v29
; GFX11TRUE16-NEXT: v_lshrrev_b32_e32 v86, 16, v86
; GFX11TRUE16-NEXT: v_mov_b16_e32 v30.l, v84.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v84.l, v85.l
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v28, v99, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v31
; GFX11TRUE16-NEXT: v_mov_b16_e32 v28.l, v86.l
; GFX11TRUE16-NEXT: v_mov_b16_e32 v85.l, v87.l
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v29
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v29, v34, v35, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v27
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v27, v36, v37, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v25
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v25, v38, v39, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v23
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v21
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v21, v50, v51, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v19
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v19, v52, v53, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v17
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v17, v54, v55, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v15
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v15, v64, v65, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v13
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v13, v66, v67, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v11
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v11, v68, v69, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v7, v80, v81, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v30, v84, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v28, v85, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v5, v82, v83, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v9
; GFX11TRUE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v6, v13, v12, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v9, v70, v71, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v5, v11, v10, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v7, v15, v14, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v10, v21, v20, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v11, v23, v22, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v8, v17, v16, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v9, v19, v18, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v12, v25, v24, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v13, v27, v26, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v14, v29, v98, 0x5040100
; GFX11TRUE16-NEXT: v_perm_b32 v15, v31, v96, 0x5040100
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_vselect_v32bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: s_clause 0x1f
; GFX11FAKE16-NEXT: scratch_load_u16 v31, off, s32
; GFX11FAKE16-NEXT: scratch_load_b32 v32, off, s32 offset:128
; GFX11FAKE16-NEXT: scratch_load_b32 v33, off, s32 offset:64
; GFX11FAKE16-NEXT: scratch_load_b32 v34, off, s32 offset:124
; GFX11FAKE16-NEXT: scratch_load_b32 v35, off, s32 offset:60
; GFX11FAKE16-NEXT: scratch_load_b32 v36, off, s32 offset:120
; GFX11FAKE16-NEXT: scratch_load_b32 v37, off, s32 offset:56
; GFX11FAKE16-NEXT: scratch_load_b32 v38, off, s32 offset:116
; GFX11FAKE16-NEXT: scratch_load_b32 v39, off, s32 offset:52
; GFX11FAKE16-NEXT: scratch_load_b32 v48, off, s32 offset:112
; GFX11FAKE16-NEXT: scratch_load_b32 v49, off, s32 offset:48
; GFX11FAKE16-NEXT: scratch_load_b32 v50, off, s32 offset:108
; GFX11FAKE16-NEXT: scratch_load_b32 v51, off, s32 offset:44
; GFX11FAKE16-NEXT: scratch_load_b32 v52, off, s32 offset:104
; GFX11FAKE16-NEXT: scratch_load_b32 v53, off, s32 offset:40
; GFX11FAKE16-NEXT: scratch_load_b32 v54, off, s32 offset:100
; GFX11FAKE16-NEXT: scratch_load_b32 v55, off, s32 offset:36
; GFX11FAKE16-NEXT: scratch_load_b32 v64, off, s32 offset:96
; GFX11FAKE16-NEXT: scratch_load_b32 v65, off, s32 offset:32
; GFX11FAKE16-NEXT: scratch_load_b32 v66, off, s32 offset:92
; GFX11FAKE16-NEXT: scratch_load_b32 v67, off, s32 offset:28
; GFX11FAKE16-NEXT: scratch_load_b32 v68, off, s32 offset:88
; GFX11FAKE16-NEXT: scratch_load_b32 v69, off, s32 offset:24
; GFX11FAKE16-NEXT: scratch_load_b32 v70, off, s32 offset:84
; GFX11FAKE16-NEXT: scratch_load_b32 v71, off, s32 offset:20
; GFX11FAKE16-NEXT: scratch_load_b32 v80, off, s32 offset:80
; GFX11FAKE16-NEXT: scratch_load_b32 v81, off, s32 offset:16
; GFX11FAKE16-NEXT: scratch_load_b32 v82, off, s32 offset:76
; GFX11FAKE16-NEXT: scratch_load_b32 v83, off, s32 offset:12
; GFX11FAKE16-NEXT: scratch_load_b32 v84, off, s32 offset:72
; GFX11FAKE16-NEXT: scratch_load_b32 v85, off, s32 offset:8
; GFX11FAKE16-NEXT: scratch_load_b32 v86, off, s32 offset:68
; GFX11FAKE16-NEXT: scratch_load_b32 v87, off, s32 offset:4
; GFX11FAKE16-NEXT: v_and_b32_e32 v30, 1, v30
; GFX11FAKE16-NEXT: v_and_b32_e32 v28, 1, v28
; GFX11FAKE16-NEXT: v_and_b32_e32 v26, 1, v26
; GFX11FAKE16-NEXT: v_and_b32_e32 v24, 1, v24
; GFX11FAKE16-NEXT: v_and_b32_e32 v22, 1, v22
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v30
; GFX11FAKE16-NEXT: v_and_b32_e32 v3, 1, v3
; GFX11FAKE16-NEXT: v_and_b32_e32 v20, 1, v20
; GFX11FAKE16-NEXT: v_and_b32_e32 v18, 1, v18
; GFX11FAKE16-NEXT: v_and_b32_e32 v16, 1, v16
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(30)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v28
; GFX11FAKE16-NEXT: v_and_b32_e32 v1, 1, v1
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v33, 16, v33
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v32, 16, v32
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 1, v0
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(28)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v28, v34, v35, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v26
; GFX11FAKE16-NEXT: v_and_b32_e32 v7, 1, v7
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v35, 16, v35
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v34, 16, v34
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 1, v2
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(26)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v26, v36, v37, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v24
; GFX11FAKE16-NEXT: v_and_b32_e32 v5, 1, v5
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v37, 16, v37
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v36, 16, v36
; GFX11FAKE16-NEXT: v_and_b32_e32 v4, 1, v4
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(24)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v24, v38, v39, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v22
; GFX11FAKE16-NEXT: v_and_b32_e32 v11, 1, v11
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v39, 16, v39
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v38, 16, v38
; GFX11FAKE16-NEXT: v_and_b32_e32 v6, 1, v6
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(22)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v22, v48, v49, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v20
; GFX11FAKE16-NEXT: v_and_b32_e32 v9, 1, v9
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v49, 16, v49
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v48, 16, v48
; GFX11FAKE16-NEXT: v_and_b32_e32 v8, 1, v8
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(20)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v20, v50, v51, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v18
; GFX11FAKE16-NEXT: v_and_b32_e32 v15, 1, v15
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v51, 16, v51
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v50, 16, v50
; GFX11FAKE16-NEXT: v_and_b32_e32 v10, 1, v10
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(18)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v18, v52, v53, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v16
; GFX11FAKE16-NEXT: v_and_b32_e32 v13, 1, v13
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v53, 16, v53
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v52, 16, v52
; GFX11FAKE16-NEXT: v_and_b32_e32 v12, 1, v12
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(16)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v16, v54, v55, vcc_lo
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v55, 16, v55
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v54, 16, v54
; GFX11FAKE16-NEXT: v_and_b32_e32 v14, 1, v14
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v14
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(14)
; GFX11FAKE16-NEXT: v_dual_cndmask_b32 v14, v64, v65 :: v_dual_and_b32 v19, 1, v19
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v12
; GFX11FAKE16-NEXT: v_and_b32_e32 v17, 1, v17
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v65, 16, v65
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v64, 16, v64
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(12)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v12, v66, v67, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v10
; GFX11FAKE16-NEXT: v_and_b32_e32 v23, 1, v23
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v67, 16, v67
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v66, 16, v66
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(10)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v10, v68, v69, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v8
; GFX11FAKE16-NEXT: v_and_b32_e32 v21, 1, v21
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v69, 16, v69
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v68, 16, v68
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(8)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v8, v70, v71, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v6
; GFX11FAKE16-NEXT: v_and_b32_e32 v27, 1, v27
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v71, 16, v71
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v70, 16, v70
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(6)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v6, v80, v81, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v4
; GFX11FAKE16-NEXT: v_and_b32_e32 v25, 1, v25
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v81, 16, v81
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v80, 16, v80
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(4)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v4, v82, v83, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v31, 1, v31
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v83, 16, v83
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v82, 16, v82
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v84, v85, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v29, 1, v29
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v85, 16, v85
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v84, 16, v84
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v86, v87, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v31
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v87, 16, v87
; GFX11FAKE16-NEXT: v_lshrrev_b32_e32 v86, 16, v86
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v31, v32, v33, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v29
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v29, v34, v35, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v27
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v27, v36, v37, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v25
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v25, v38, v39, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v23
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v23, v48, v49, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v21
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v21, v50, v51, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v19
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v19, v52, v53, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v17
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v17, v54, v55, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v15
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v15, v64, v65, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v13
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v13, v66, v67, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v11
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v11, v68, v69, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v7
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v7, v80, v81, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v3
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, v84, v85, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v1
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v86, v87, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v5
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_perm_b32 v0, v1, v0, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v5, v82, v83, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, 1, v9
; GFX11FAKE16-NEXT: v_perm_b32 v1, v3, v2, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v3, v7, v6, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v6, v13, v12, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v2, v5, v4, 0x5040100
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v9, v70, v71, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v5, v11, v10, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v7, v15, v14, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v10, v21, v20, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v11, v23, v22, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v4, v9, v8, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v8, v17, v16, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v9, v19, v18, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v12, v25, v24, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v13, v27, v26, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v14, v29, v28, 0x5040100
; GFX11FAKE16-NEXT: v_perm_b32 v15, v31, v30, 0x5040100
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = select <32 x i1> %cond, <32 x bfloat> %a, <32 x bfloat> %b
ret <32 x bfloat> %op
}
declare bfloat @llvm.fma.bf16(bfloat, bfloat, bfloat)
declare <2 x bfloat> @llvm.fma.v2bf16(<2 x bfloat>, <2 x bfloat>, <2 x bfloat>)
declare <3 x bfloat> @llvm.fma.v3bf16(<3 x bfloat>, <3 x bfloat>, <3 x bfloat>)
declare <4 x bfloat> @llvm.fma.v4bf16(<4 x bfloat>, <4 x bfloat>, <4 x bfloat>)
define bfloat @v_fma_bf16(bfloat %a, bfloat %b, bfloat %c) {
; GCN-LABEL: v_fma_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_fma_f32 v0, v0, v1, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fma_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_fma_f32 v0, v0, v1, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fma_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_fma_f32 v0, v0, v1, v2
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fma_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_fma_f32 v0, v0, v1, v2
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fma_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_fmac_f32_e32 v2, v0, v1
; GFX10-NEXT: v_bfe_u32 v0, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_add3_u32 v0, v0, v2, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fma_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v2, 16, v2
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_fmac_f32_e32 v2, v0, v1
; GFX11-NEXT: v_bfe_u32 v0, v2, 16, 1
; GFX11-NEXT: v_or_b32_e32 v1, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v0, v0, v2, 0x7fff
; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.fma.bf16(bfloat %a, bfloat %b, bfloat %c)
ret bfloat %op
}
define <2 x bfloat> @v_fma_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c) {
; GCN-LABEL: v_fma_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_fma_f32 v1, v1, v3, v5
; GCN-NEXT: v_fma_f32 v0, v0, v2, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fma_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_fma_f32 v1, v1, v3, v5
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_fma_f32 v0, v0, v2, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fma_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_fma_f32 v3, v5, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_fma_f32 v0, v0, v1, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fma_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_fma_f32 v3, v5, v4, v3
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_fma_f32 v0, v0, v1, v2
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fma_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_fmac_f32_e32 v3, v5, v4
; GFX10-NEXT: v_fmac_f32_e32 v2, v0, v1
; GFX10-NEXT: v_bfe_u32 v0, v3, 16, 1
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX10-NEXT: v_add3_u32 v0, v0, v3, 0x7fff
; GFX10-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fma_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_fmac_f32 v2, v0, v1 :: v_dual_fmac_f32 v3, v5, v4
; GFX11-NEXT: v_bfe_u32 v1, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT: v_bfe_u32 v0, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v2
; GFX11-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-NEXT: v_add3_u32 v0, v0, v3, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v1, v0, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <2 x bfloat> @llvm.fma.v2bf16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c)
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_fma_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b, <3 x bfloat> %c) {
; GCN-LABEL: v_fma_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_fma_f32 v2, v2, v5, v8
; GCN-NEXT: v_fma_f32 v1, v1, v4, v7
; GCN-NEXT: v_fma_f32 v0, v0, v3, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fma_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_fma_f32 v2, v2, v5, v8
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_fma_f32 v1, v1, v4, v5
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_fma_f32 v0, v0, v3, v4
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fma_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_fma_f32 v1, v1, v3, v5
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v0
; GFX8-NEXT: v_fma_f32 v3, v6, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_fma_f32 v0, v0, v2, v4
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fma_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_fma_f32 v1, v1, v3, v5
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v0
; GFX9-NEXT: v_fma_f32 v3, v6, v5, v3
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_fma_f32 v0, v0, v2, v4
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fma_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_fmac_f32_e32 v6, v8, v7
; GFX10-NEXT: v_fmac_f32_e32 v5, v1, v3
; GFX10-NEXT: v_fmac_f32_e32 v4, v0, v2
; GFX10-NEXT: v_bfe_u32 v1, v6, 16, 1
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v6
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v0, v5, 16, 1
; GFX10-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX10-NEXT: v_add3_u32 v1, v1, v6, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v5
; GFX10-NEXT: v_add3_u32 v0, v0, v5, 0x7fff
; GFX10-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_cndmask_b32_e32 v3, v0, v8, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v2, v1, 0x7060302
; GFX10-NEXT: v_alignbit_b32 v1, s4, v3, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fma_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11TRUE16-NEXT: v_dual_fmac_f32 v6, v8, v7 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_fmac_f32_e32 v4, v0, v2
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_fmac_f32_e32 v5, v1, v3
; GFX11TRUE16-NEXT: v_bfe_u32 v1, v6, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6
; GFX11TRUE16-NEXT: v_bfe_u32 v0, v5, 16, 1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_add3_u32 v1, v1, v6, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5
; GFX11TRUE16-NEXT: v_add3_u32 v0, v0, v5, 0x7fff
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v3, v0, v8, vcc_lo
; GFX11TRUE16-NEXT: v_perm_b32 v0, v2, v1, 0x7060302
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v3, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fma_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v4
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v7, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v8, 16, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX11FAKE16-NEXT: v_dual_fmac_f32 v6, v8, v7 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_fmac_f32_e32 v4, v0, v2
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v4, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v4
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_fmac_f32_e32 v5, v1, v3
; GFX11FAKE16-NEXT: v_bfe_u32 v1, v6, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v3, 0x400000, v6
; GFX11FAKE16-NEXT: v_bfe_u32 v0, v5, 16, 1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_add3_u32 v1, v1, v6, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v5
; GFX11FAKE16-NEXT: v_add3_u32 v0, v0, v5, 0x7fff
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v3, v0, v8, vcc_lo
; GFX11FAKE16-NEXT: v_perm_b32 v0, v2, v1, 0x7060302
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v3, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = call <3 x bfloat> @llvm.fma.v3bf16(<3 x bfloat> %a, <3 x bfloat> %b, <3 x bfloat> %c)
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_fma_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c) {
; GCN-LABEL: v_fma_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_fma_f32 v3, v3, v7, v11
; GCN-NEXT: v_fma_f32 v2, v2, v6, v10
; GCN-NEXT: v_fma_f32 v1, v1, v5, v9
; GCN-NEXT: v_fma_f32 v0, v0, v4, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fma_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_fma_f32 v3, v3, v7, v11
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_fma_f32 v2, v2, v6, v7
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_fma_f32 v1, v1, v5, v6
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_fma_f32 v0, v0, v4, v5
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fma_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v8, 16, v1
; GFX8-NEXT: v_fma_f32 v6, v8, v7, v6
; GFX8-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
; GFX8-NEXT: v_fma_f32 v1, v1, v3, v5
; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v4
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX8-NEXT: v_fma_f32 v3, v7, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_fma_f32 v0, v0, v2, v4
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fma_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v8, 16, v1
; GFX9-NEXT: v_fma_f32 v6, v8, v7, v6
; GFX9-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_fma_f32 v1, v1, v3, v5
; GFX9-NEXT: v_add3_u32 v7, v7, v6, s4
; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v4
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX9-NEXT: v_fma_f32 v3, v7, v5, v3
; GFX9-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_fma_f32 v0, v0, v2, v4
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fma_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v1
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX10-NEXT: v_fmac_f32_e32 v6, v8, v7
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v4
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v10, v6, 16, 1
; GFX10-NEXT: v_fmac_f32_e32 v5, v1, v3
; GFX10-NEXT: v_fmac_f32_e32 v7, v9, v8
; GFX10-NEXT: v_or_b32_e32 v1, 0x400000, v6
; GFX10-NEXT: v_fmac_f32_e32 v4, v0, v2
; GFX10-NEXT: v_add3_u32 v0, v10, v6, 0x7fff
; GFX10-NEXT: v_bfe_u32 v2, v5, 16, 1
; GFX10-NEXT: v_bfe_u32 v3, v7, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_bfe_u32 v8, v4, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5
; GFX10-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc_lo
; GFX10-NEXT: v_add3_u32 v0, v2, v5, 0x7fff
; GFX10-NEXT: v_add3_u32 v2, v3, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v7
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_add3_u32 v6, v8, v4, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4
; GFX10-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT: v_cndmask_b32_e32 v4, v0, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v3, v2, 0x7060302
; GFX10-NEXT: v_perm_b32 v1, v4, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fma_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v5
; GFX11-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_fmac_f32_e32 v5, v1, v3
; GFX11-NEXT: v_dual_fmac_f32 v6, v8, v7 :: v_dual_lshlrev_b32 v7, 16, v4
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_bfe_u32 v10, v6, 16, 1
; GFX11-NEXT: v_or_b32_e32 v1, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v2
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_fmac_f32_e32 v4, v0, v2
; GFX11-NEXT: v_add3_u32 v0, v10, v6, 0x7fff
; GFX11-NEXT: v_bfe_u32 v2, v5, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc_lo
; GFX11-NEXT: v_fmac_f32_e32 v7, v9, v8
; GFX11-NEXT: v_bfe_u32 v8, v4, 16, 1
; GFX11-NEXT: v_add3_u32 v0, v2, v5, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v5
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_bfe_u32 v3, v7, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_add3_u32 v6, v8, v4, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v4
; GFX11-NEXT: v_add3_u32 v2, v3, v7, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v7
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT: v_cndmask_b32_e32 v4, v0, v9, vcc_lo
; GFX11-NEXT: v_perm_b32 v0, v3, v2, 0x7060302
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT: v_perm_b32 v1, v4, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <4 x bfloat> @llvm.fma.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c)
ret <4 x bfloat> %op
}
declare bfloat @llvm.fmuladd.bf16(bfloat, bfloat, bfloat)
declare <2 x bfloat> @llvm.fmuladd.v2bf16(<2 x bfloat>, <2 x bfloat>, <2 x bfloat>)
declare <3 x bfloat> @llvm.fmuladd.v3bf16(<3 x bfloat>, <3 x bfloat>, <3 x bfloat>)
declare <4 x bfloat> @llvm.fmuladd.v4bf16(<4 x bfloat>, <4 x bfloat>, <4 x bfloat>)
define bfloat @v_fmuladd_bf16(bfloat %a, bfloat %b, bfloat %c) {
; GCN-LABEL: v_fmuladd_bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_mul_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GCN-NEXT: v_add_f32_e32 v0, v0, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmuladd_bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX7-NEXT: v_add_f32_e32 v0, v0, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmuladd_bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmuladd_bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmuladd_bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v3, vcc_lo
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_add_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX10-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmuladd_bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: v_dual_cndmask_b32 v0, v1, v3 :: v_dual_lshlrev_b32 v1, 16, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_add_f32_e32 v0, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v1, v1, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc_lo
; GFX11-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call bfloat @llvm.fmuladd.bf16(bfloat %a, bfloat %b, bfloat %c)
ret bfloat %op
}
define <2 x bfloat> @v_fmuladd_v2bf16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c) {
; GCN-LABEL: v_fmuladd_v2bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_mul_f32_e32 v1, v1, v3
; GCN-NEXT: v_mul_f32_e32 v0, v0, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v1, v1, v5
; GCN-NEXT: v_add_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmuladd_v2bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX7-NEXT: v_add_f32_e32 v1, v1, v3
; GFX7-NEXT: v_add_f32_e32 v0, v0, v2
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmuladd_v2bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v3, v4, v3
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX8-NEXT: v_add_f32_e32 v3, v3, v4
; GFX8-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v4, vcc, s4, v4
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, s4, v1
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX8-NEXT: v_add_f32_e32 v0, v0, v1
; GFX8-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v1, vcc, v1, v0
; GFX8-NEXT: v_add_u32_e32 v1, vcc, 0x7fff, v1
; GFX8-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmuladd_v2bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v3, v4, v3
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v4, 16, v2
; GFX9-NEXT: v_add_f32_e32 v3, v3, v4
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX9-NEXT: v_add3_u32 v4, v4, v3, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v4, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v2
; GFX9-NEXT: v_add_f32_e32 v0, v0, v1
; GFX9-NEXT: v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v1, v1, v0, s4
; GFX9-NEXT: v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmuladd_v2bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v3, v4, v3
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX10-NEXT: v_bfe_u32 v1, v3, 16, 1
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v1, v1, v3, 0x7fff
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX10-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v0, v4, v6, vcc_lo
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX10-NEXT: v_add_f32_e32 v0, v0, v2
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX10-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmuladd_v2bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v4, 16, v0
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_dual_mul_f32 v3, v4, v3 :: v_dual_and_b32 v0, 0xffff0000, v0
; GFX11-NEXT: v_mul_f32_e32 v0, v0, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v1, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v1, v1, v3, 0x7fff
; GFX11-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX11-NEXT: v_bfe_u32 v4, v0, 16, 1
; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v1, v5, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_add3_u32 v4, v4, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-NEXT: v_dual_cndmask_b32 v0, v4, v6 :: v_dual_add_f32 v1, v1, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_or_b32_e32 v4, 0x400000, v1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add_f32_e32 v0, v0, v2
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: v_bfe_u32 v3, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX11-NEXT: v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v1, v2, v4, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v5, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v1, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <2 x bfloat> @llvm.fmuladd.v2bf16(<2 x bfloat> %a, <2 x bfloat> %b, <2 x bfloat> %c)
ret <2 x bfloat> %op
}
define <3 x bfloat> @v_fmuladd_v3bf16(<3 x bfloat> %a, <3 x bfloat> %b, <3 x bfloat> %c) {
; GCN-LABEL: v_fmuladd_v3bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_mul_f32_e32 v2, v2, v5
; GCN-NEXT: v_mul_f32_e32 v1, v1, v4
; GCN-NEXT: v_mul_f32_e32 v0, v0, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v2, v2, v8
; GCN-NEXT: v_add_f32_e32 v1, v1, v7
; GCN-NEXT: v_add_f32_e32 v0, v0, v6
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmuladd_v3bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v5
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v4
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v8
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v6
; GFX7-NEXT: v_add_f32_e32 v2, v2, v5
; GFX7-NEXT: v_add_f32_e32 v1, v1, v4
; GFX7-NEXT: v_add_f32_e32 v0, v0, v3
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmuladd_v3bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, 0x7fff, v3
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v5
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_add_f32_e32 v3, v3, v5
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, s4, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmuladd_v3bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v6, vcc
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v5
; GFX9-NEXT: v_add_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_add_f32_e32 v3, v3, v5
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmuladd_v3bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX10-NEXT: v_mul_f32_e32 v3, v7, v6
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX10-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v6, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v5
; GFX10-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v2, v7, v9, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_add_f32_e32 v1, v1, v3
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cndmask_b32_e32 v0, v8, v10, vcc_lo
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_add_f32_e32 v2, v2, v5
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT: v_add_f32_e32 v0, v0, v4
; GFX10-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX10-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX10-NEXT: v_alignbit_b32 v1, s4, v1, 16
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11TRUE16-LABEL: v_fmuladd_v3bf16:
; GFX11TRUE16: ; %bb.0:
; GFX11TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11TRUE16-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11TRUE16-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX11TRUE16-NEXT: v_mul_f32_e32 v3, v7, v6
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX11TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v6, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v5
; GFX11TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX11TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v7, v9, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v8, v10, vcc_lo
; GFX11TRUE16-NEXT: v_add_f32_e32 v2, v2, v5
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11TRUE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11TRUE16-NEXT: v_dual_add_f32 v0, v0, v4 :: v_dual_add_f32 v1, v1, v3
; GFX11TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11TRUE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11TRUE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11TRUE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11TRUE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11TRUE16-NEXT: v_alignbit_b32 v1, v0, v1, 16
; GFX11TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11FAKE16-LABEL: v_fmuladd_v3bf16:
; GFX11FAKE16: ; %bb.0:
; GFX11FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v6, 16, v2
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v7, 16, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_dual_mul_f32 v0, v0, v2 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11FAKE16-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v10, 0x400000, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_add3_u32 v8, v8, v0, 0x7fff
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
; GFX11FAKE16-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX11FAKE16-NEXT: v_mul_f32_e32 v3, v7, v6
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_or_b32_e32 v9, 0x400000, v3
; GFX11FAKE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_add3_u32 v7, v7, v3, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v2, v6, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v5
; GFX11FAKE16-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX11FAKE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v7, v9, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v8, v10, vcc_lo
; GFX11FAKE16-NEXT: v_add_f32_e32 v2, v2, v5
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11FAKE16-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11FAKE16-NEXT: v_or_b32_e32 v7, 0x400000, v2
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11FAKE16-NEXT: v_dual_add_f32 v0, v0, v4 :: v_dual_add_f32 v1, v1, v3
; GFX11FAKE16-NEXT: v_bfe_u32 v3, v2, 16, 1
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11FAKE16-NEXT: v_bfe_u32 v5, v0, 16, 1
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11FAKE16-NEXT: v_bfe_u32 v4, v1, 16, 1
; GFX11FAKE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff
; GFX11FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11FAKE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
; GFX11FAKE16-NEXT: v_add3_u32 v4, v4, v1, 0x7fff
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v2, v3, v7, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v0, v5, v8, vcc_lo
; GFX11FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11FAKE16-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11FAKE16-NEXT: v_cndmask_b32_e32 v1, v4, v6, vcc_lo
; GFX11FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11FAKE16-NEXT: v_alignbit_b32 v1, s0, v1, 16
; GFX11FAKE16-NEXT: s_setpc_b64 s[30:31]
%op = call <3 x bfloat> @llvm.fmuladd.v3bf16(<3 x bfloat> %a, <3 x bfloat> %b, <3 x bfloat> %c)
ret <3 x bfloat> %op
}
define <4 x bfloat> @v_fmuladd_v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c) {
; GCN-LABEL: v_fmuladd_v4bf16:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GCN-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GCN-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GCN-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GCN-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GCN-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GCN-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GCN-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GCN-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GCN-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GCN-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v11, 0xffff0000, v11
; GCN-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
; GCN-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v9, 0xffff0000, v9
; GCN-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v8, 0xffff0000, v8
; GCN-NEXT: v_mul_f32_e32 v3, v3, v7
; GCN-NEXT: v_mul_f32_e32 v2, v2, v6
; GCN-NEXT: v_mul_f32_e32 v1, v1, v5
; GCN-NEXT: v_mul_f32_e32 v0, v0, v4
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_add_f32_e32 v3, v3, v11
; GCN-NEXT: v_add_f32_e32 v2, v2, v10
; GCN-NEXT: v_add_f32_e32 v1, v1, v9
; GCN-NEXT: v_add_f32_e32 v0, v0, v8
; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GCN-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GCN-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GCN-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_fmuladd_v4bf16:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT: v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT: v_mul_f32_e32 v5, 1.0, v5
; GFX7-NEXT: v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT: v_mul_f32_e32 v6, 1.0, v6
; GFX7-NEXT: v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT: v_mul_f32_e32 v7, 1.0, v7
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v7
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_mul_f32_e32 v8, 1.0, v8
; GFX7-NEXT: v_mul_f32_e32 v9, 1.0, v9
; GFX7-NEXT: v_mul_f32_e32 v10, 1.0, v10
; GFX7-NEXT: v_mul_f32_e32 v11, 1.0, v11
; GFX7-NEXT: v_mul_f32_e32 v3, v3, v7
; GFX7-NEXT: v_mul_f32_e32 v2, v2, v6
; GFX7-NEXT: v_mul_f32_e32 v1, v1, v5
; GFX7-NEXT: v_mul_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: v_and_b32_e32 v7, 0xffff0000, v11
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v6, 0xffff0000, v10
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v5, 0xffff0000, v9
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v4, 0xffff0000, v8
; GFX7-NEXT: v_add_f32_e32 v3, v3, v7
; GFX7-NEXT: v_add_f32_e32 v2, v2, v6
; GFX7-NEXT: v_add_f32_e32 v1, v1, v5
; GFX7-NEXT: v_add_f32_e32 v0, v0, v4
; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX7-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_fmuladd_v4bf16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT: v_lshlrev_b32_e32 v6, 16, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v1
; GFX8-NEXT: v_mul_f32_e32 v6, v7, v6
; GFX8-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT: v_add_u32_e32 v7, vcc, 0x7fff, v7
; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX8-NEXT: v_lshlrev_b32_e32 v7, 16, v5
; GFX8-NEXT: v_add_f32_e32 v6, v6, v7
; GFX8-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX8-NEXT: s_movk_i32 s4, 0x7fff
; GFX8-NEXT: v_add_u32_e32 v7, vcc, v7, v6
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_add_u32_e32 v7, vcc, s4, v7
; GFX8-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX8-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v7, vcc
; GFX8-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX8-NEXT: v_add_f32_e32 v1, v1, v3
; GFX8-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, v3, v1
; GFX8-NEXT: v_add_u32_e32 v3, vcc, s4, v3
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX8-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX8-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX8-NEXT: v_add_f32_e32 v3, v3, v5
; GFX8-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX8-NEXT: v_add_u32_e32 v5, vcc, v5, v3
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_add_u32_e32 v5, vcc, s4, v5
; GFX8-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, s4, v2
; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX8-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX8-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX8-NEXT: v_add_f32_e32 v0, v0, v2
; GFX8-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX8-NEXT: v_add_u32_e32 v2, vcc, v2, v0
; GFX8-NEXT: v_add_u32_e32 v2, vcc, 0x7fff, v2
; GFX8-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX8-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX8-NEXT: v_lshrrev_b32_e32 v1, 16, v1
; GFX8-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX8-NEXT: v_alignbit_b32 v0, v0, v3, 16
; GFX8-NEXT: v_alignbit_b32 v1, v1, v6, 16
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX9-LABEL: v_fmuladd_v4bf16:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_lshlrev_b32_e32 v6, 16, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v1
; GFX9-NEXT: v_mul_f32_e32 v6, v7, v6
; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT: s_movk_i32 s4, 0x7fff
; GFX9-NEXT: v_add3_u32 v7, v7, v6, s4
; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX9-NEXT: v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT: v_lshlrev_b32_e32 v7, 16, v5
; GFX9-NEXT: v_add_f32_e32 v6, v6, v7
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX9-NEXT: v_add3_u32 v7, v7, v6, s4
; GFX9-NEXT: v_or_b32_e32 v8, 0x400000, v6
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v7, vcc
; GFX9-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v5
; GFX9-NEXT: v_add_f32_e32 v1, v1, v3
; GFX9-NEXT: v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT: v_add3_u32 v3, v3, v1, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v1
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT: v_cndmask_b32_e32 v1, v3, v5, vcc
; GFX9-NEXT: v_lshlrev_b32_e32 v3, 16, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v0
; GFX9-NEXT: v_mul_f32_e32 v3, v5, v3
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX9-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT: v_lshlrev_b32_e32 v5, 16, v4
; GFX9-NEXT: v_add_f32_e32 v3, v3, v5
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX9-NEXT: v_add3_u32 v5, v5, v3, s4
; GFX9-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_cndmask_b32_e32 v3, v5, v7, vcc
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v5, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v5, vcc
; GFX9-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT: v_and_b32_e32 v2, 0xffff0000, v4
; GFX9-NEXT: v_add_f32_e32 v0, v0, v2
; GFX9-NEXT: v_bfe_u32 v2, v0, 16, 1
; GFX9-NEXT: v_add3_u32 v2, v2, v0, s4
; GFX9-NEXT: v_or_b32_e32 v4, 0x400000, v0
; GFX9-NEXT: v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT: v_cndmask_b32_e32 v0, v2, v4, vcc
; GFX9-NEXT: s_mov_b32 s4, 0x7060302
; GFX9-NEXT: v_perm_b32 v0, v0, v3, s4
; GFX9-NEXT: v_perm_b32 v1, v1, v6, s4
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: v_fmuladd_v4bf16:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v1
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_mul_f32_e32 v6, v7, v6
; GFX10-NEXT: v_lshlrev_b32_e32 v7, 16, v2
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_mul_f32_e32 v1, v1, v3
; GFX10-NEXT: v_lshlrev_b32_e32 v8, 16, v5
; GFX10-NEXT: v_bfe_u32 v10, v6, 16, 1
; GFX10-NEXT: v_or_b32_e32 v3, 0x400000, v6
; GFX10-NEXT: v_mul_f32_e32 v7, v9, v7
; GFX10-NEXT: v_mul_f32_e32 v0, v0, v2
; GFX10-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX10-NEXT: v_add3_u32 v10, v10, v6, 0x7fff
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX10-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX10-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX10-NEXT: v_bfe_u32 v11, v0, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v10, v3, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX10-NEXT: v_add3_u32 v9, v9, v7, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v12, 0x400000, v0
; GFX10-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX10-NEXT: v_cndmask_b32_e32 v1, v2, v6, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX10-NEXT: v_add3_u32 v11, v11, v0, 0x7fff
; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v5
; GFX10-NEXT: v_add_f32_e32 v3, v3, v8
; GFX10-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT: v_cndmask_b32_e32 v2, v9, v10, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_lshlrev_b32_e32 v6, 16, v4
; GFX10-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX10-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX10-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
; GFX10-NEXT: v_cndmask_b32_e32 v0, v11, v12, vcc_lo
; GFX10-NEXT: v_add_f32_e32 v1, v1, v5
; GFX10-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT: v_add_f32_e32 v2, v2, v6
; GFX10-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX10-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX10-NEXT: v_add_f32_e32 v0, v0, v4
; GFX10-NEXT: v_add3_u32 v4, v7, v3, 0x7fff
; GFX10-NEXT: v_bfe_u32 v7, v2, 16, 1
; GFX10-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX10-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc_lo
; GFX10-NEXT: v_add3_u32 v4, v6, v1, 0x7fff
; GFX10-NEXT: v_add3_u32 v5, v7, v2, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX10-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX10-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX10-NEXT: v_cndmask_b32_e32 v1, v4, v9, vcc_lo
; GFX10-NEXT: v_perm_b32 v1, v1, v3, 0x7060302
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_fmuladd_v4bf16:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT: v_lshlrev_b32_e32 v9, 16, v0
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v1
; GFX11-NEXT: v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT: v_lshlrev_b32_e32 v8, 16, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v6, 16, v3
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_dual_mul_f32 v6, v7, v6 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX11-NEXT: v_lshlrev_b32_e32 v7, 16, v2
; GFX11-NEXT: v_dual_mul_f32 v1, v1, v3 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v10, v6, 16, 1
; GFX11-NEXT: v_mul_f32_e32 v7, v9, v7
; GFX11-NEXT: v_or_b32_e32 v3, 0x400000, v6
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v10, v10, v6, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v1
; GFX11-NEXT: v_bfe_u32 v9, v7, 16, 1
; GFX11-NEXT: v_dual_cndmask_b32 v3, v10, v3 :: v_dual_mul_f32 v0, v0, v2
; GFX11-NEXT: v_bfe_u32 v2, v1, 16, 1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: v_or_b32_e32 v10, 0x400000, v7
; GFX11-NEXT: v_add3_u32 v9, v9, v7, 0x7fff
; GFX11-NEXT: v_bfe_u32 v11, v0, 16, 1
; GFX11-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v12, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_add3_u32 v11, v11, v0, 0x7fff
; GFX11-NEXT: v_dual_cndmask_b32 v1, v2, v6 :: v_dual_lshlrev_b32 v6, 16, v4
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-NEXT: v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_dual_cndmask_b32 v2, v9, v10 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_dual_add_f32 v1, v1, v5 :: v_dual_and_b32 v2, 0xffff0000, v2
; GFX11-NEXT: v_cndmask_b32_e32 v0, v11, v12, vcc_lo
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_or_b32_e32 v9, 0x400000, v1
; GFX11-NEXT: v_add_f32_e32 v2, v2, v6
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-NEXT: v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-NEXT: v_bfe_u32 v6, v1, 16, 1
; GFX11-NEXT: v_add_f32_e32 v0, v0, v4
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT: v_add_f32_e32 v3, v3, v8
; GFX11-NEXT: v_bfe_u32 v8, v0, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_bfe_u32 v7, v3, 16, 1
; GFX11-NEXT: v_or_b32_e32 v5, 0x400000, v3
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT: v_add3_u32 v4, v7, v3, 0x7fff
; GFX11-NEXT: v_bfe_u32 v7, v2, 16, 1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v3, v4, v5, vcc_lo
; GFX11-NEXT: v_add3_u32 v4, v6, v1, 0x7fff
; GFX11-NEXT: v_add3_u32 v5, v7, v2, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v6, 0x400000, v2
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT: v_add3_u32 v7, v8, v0, 0x7fff
; GFX11-NEXT: v_or_b32_e32 v8, 0x400000, v0
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT: v_cndmask_b32_e32 v2, v5, v6, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-NEXT: v_cndmask_b32_e32 v0, v7, v8, vcc_lo
; GFX11-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_perm_b32 v0, v0, v2, 0x7060302
; GFX11-NEXT: v_cndmask_b32_e32 v1, v4, v9, vcc_lo
; GFX11-NEXT: v_perm_b32 v1, v1, v3, 0x7060302
; GFX11-NEXT: s_setpc_b64 s[30:31]
%op = call <4 x bfloat> @llvm.fmuladd.v4bf16(<4 x bfloat> %a, <4 x bfloat> %b, <4 x bfloat> %c)
ret <4 x bfloat> %op
}
Codebase Browser
llvm