; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck --check-prefix=GCN-NOHSA-SI %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck --check-prefix=GCN-HSA %s
; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck --check-prefix=GCN-NOHSA-VI %s
; RUN: llc -mtriple=r600 -mcpu=redwood -verify-machineinstrs < %s | FileCheck --check-prefix=EG %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX12 %s
define amdgpu_kernel void @constant_load_i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_short v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: flat_store_short v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 11, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT MSKOR T0.XW, T1.X
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: AND_INT T0.W, KC0[2].Y, literal.x,
; EG-NEXT: AND_INT * T1.W, T0.X, literal.y,
; EG-NEXT: 3(4.203895e-45), 65535(9.183409e-41)
; EG-NEXT: LSHL * T0.W, PV.W, literal.x,
; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00)
; EG-NEXT: LSHL T0.X, T1.W, PV.W,
; EG-NEXT: LSHL * T0.W, literal.x, PV.W,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: MOV T0.Y, 0.0,
; EG-NEXT: MOV * T0.Z, 0.0,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v0, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: global_load_u16 v1, v0, s[2:3]
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: global_store_b16 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load i16, ptr addrspace(4) %in
store i16 %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v2i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_v2i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dword s4, s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v2i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: flat_store_dword v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v2i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v2i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v2i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <2 x i16>, ptr addrspace(4) %in
store <2 x i16> %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v3i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_v3i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: buffer_store_short v0, off, s[0:3], 0 offset:4
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v3i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: s_add_u32 s4, s0, 4
; GCN-HSA-NEXT: s_addc_u32 s5, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s5
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s2
; GCN-HSA-NEXT: flat_store_short v[2:3], v4
; GCN-HSA-NEXT: flat_store_dword v[0:1], v5
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v3i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s0, 4
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s2
; GCN-NOHSA-VI-NEXT: flat_store_short v[2:3], v4
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v5
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v3i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 2 @6
; EG-NEXT: ALU 19, @13, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.X, T7.X, 0
; EG-NEXT: MEM_RAT MSKOR T5.XW, T8.X
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T6.X, T5.X, 0, #1
; EG-NEXT: VTX_READ_16 T7.X, T5.X, 2, #1
; EG-NEXT: VTX_READ_16 T5.X, T5.X, 4, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: MOV * T5.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 13:
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 4(5.605194e-45), 0(0.000000e+00)
; EG-NEXT: AND_INT T1.W, PV.W, literal.x,
; EG-NEXT: AND_INT * T2.W, T5.X, literal.y,
; EG-NEXT: 3(4.203895e-45), 65535(9.183409e-41)
; EG-NEXT: LSHL * T1.W, PV.W, literal.x,
; EG-NEXT: 3(4.203895e-45), 0(0.000000e+00)
; EG-NEXT: LSHL T5.X, T2.W, PV.W,
; EG-NEXT: LSHL * T5.W, literal.x, PV.W,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: MOV T5.Y, 0.0,
; EG-NEXT: MOV * T5.Z, 0.0,
; EG-NEXT: LSHR T8.X, T0.W, literal.x,
; EG-NEXT: LSHL T0.W, T7.X, literal.y,
; EG-NEXT: AND_INT * T1.W, T6.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: OR_INT T6.X, PV.W, PS,
; EG-NEXT: LSHR * T7.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v3i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b16 v0, v1, s[0:1] offset:4
; GFX12-NEXT: global_store_b32 v0, v2, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <3 x i16>, ptr addrspace(4) %in
store <3 x i16> %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v4i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_v4i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v4i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s3
; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v4i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v4i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T0.XY, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v4i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: v_mov_b32_e32 v2, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <4 x i16>, ptr addrspace(4) %in
store <4 x i16> %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v8i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_v8i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v8i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s7
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v8i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v8i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v8i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b128 s[4:7], s[2:3], 0x0
; GFX12-NEXT: v_mov_b32_e32 v4, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v3, s7
; GFX12-NEXT: v_dual_mov_b32 v1, s5 :: v_dual_mov_b32 v2, s6
; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <8 x i16>, ptr addrspace(4) %in
store <8 x i16> %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v16i16(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_load_v16i16:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v16i16:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[8:11], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GCN-HSA-NEXT: s_add_u32 s10, s8, 16
; GCN-HSA-NEXT: s_addc_u32 s11, s9, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s11
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s7
; GCN-HSA-NEXT: flat_store_dwordx4 v[6:7], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s9
; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v16i16:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s8, 16
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v6, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, s11
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[6:7], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v6, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v16i16:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @8
; EG-NEXT: ALU 3, @13, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T2.X, 0
; EG-NEXT: TEX 0 @10
; EG-NEXT: ALU 1, @17, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 8:
; EG-NEXT: VTX_READ_128 T1.XYZW, T0.X, 16, #1
; EG-NEXT: Fetch clause starting at 10:
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 13:
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: LSHR * T2.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 17:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v16i16:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[8:11], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b256 s[0:7], s[10:11], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v1, s5
; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v3, s7
; GFX12-NEXT: v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v5, s1
; GFX12-NEXT: v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v7, s3
; GFX12-NEXT: v_mov_b32_e32 v6, s2
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v8, v[0:3], s[8:9] offset:16
; GFX12-NEXT: global_store_b128 v8, v[4:7], s[8:9]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <16 x i16>, ptr addrspace(4) %in
store <16 x i16> %ld, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_load_v16i16_align2(ptr addrspace(4) %ptr0) #0 {
; GCN-NOHSA-SI-LABEL: constant_load_v16i16_align2:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v1, off, s[0:3], 0 offset:2
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v4, off, s[0:3], 0 offset:4
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v2, off, s[0:3], 0 offset:6
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v5, off, s[0:3], 0 offset:8
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v3, off, s[0:3], 0 offset:10
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v6, off, s[0:3], 0 offset:12
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v7, off, s[0:3], 0 offset:14
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v8, off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v9, off, s[0:3], 0 offset:18
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v10, off, s[0:3], 0 offset:20
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v11, off, s[0:3], 0 offset:22
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v12, off, s[0:3], 0 offset:24
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v13, off, s[0:3], 0 offset:26
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v14, off, s[0:3], 0 offset:28
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v15, off, s[0:3], 0 offset:30
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(8)
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v16, 16, v3
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v17, 16, v2
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v18, 16, v1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
; GCN-NOHSA-SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v3, v7, v6
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v2, v16, v5
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v1, v17, v4
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v0, v18, v0
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v7, v15, v14
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v6, v13, v12
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v5, v11, v10
; GCN-NOHSA-SI-NEXT: v_or_b32_e32 v4, v9, v8
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_load_v16i16_align2:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx2 s[0:1], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_add_u32 s0, s0, 16
; GCN-HSA-NEXT: s_addc_u32 s1, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_load_dwordx4 v[0:3], v[0:1]
; GCN-HSA-NEXT: flat_load_dwordx4 v[4:7], v[4:5]
; GCN-HSA-NEXT: s_waitcnt vmcnt(1)
; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[0:3]
; GCN-HSA-NEXT: s_waitcnt vmcnt(1)
; GCN-HSA-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_load_v16i16_align2:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[0:1], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 14
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 12
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 10
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 8
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v6, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 6
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v9, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v8, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 4
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v11, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v10, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 30
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v13, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v12, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 28
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v15, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v14, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 26
; GCN-NOHSA-VI-NEXT: flat_load_ushort v16, v[0:1]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v17, v[2:3]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v18, v[4:5]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v19, v[6:7]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v20, v[8:9]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v21, v[10:11]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v22, v[12:13]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v23, v[14:15]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 24
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 22
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 20
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v7, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v6, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 18
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v9, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v8, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v11, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v10, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 2
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v13, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v15, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v12, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v14, s0
; GCN-NOHSA-VI-NEXT: flat_load_ushort v0, v[0:1]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v24, v[2:3]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v4, v[4:5]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v5, v[6:7]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v8, v[8:9]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v9, v[10:11]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v10, v[12:13]
; GCN-NOHSA-VI-NEXT: flat_load_ushort v11, v[14:15]
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(14)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v1, 16, v16
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v3, v17, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(13)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v1, 16, v18
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(12)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v2, v19, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(11)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v1, 16, v20
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(10)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v1, v21, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(9)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v6, 16, v22
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(8)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v7, v23, v6
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(7)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(6)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v6, v24, v0
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(5)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 16, v4
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(4)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v5, v5, v0
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(3)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 16, v8
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(2)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v4, v9, v0
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(1)
; GCN-NOHSA-VI-NEXT: v_lshlrev_b32_e32 v0, 16, v10
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: v_or_b32_e32 v0, v11, v0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[0:1], v[4:7]
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[0:1], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_load_v16i16_align2:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @6
; EG-NEXT: ALU 1, @11, KC0[], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T1.XYZW, T2.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XYZW, T2.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T1.XYZW, T0.X, 16, #1
; EG-NEXT: VTX_READ_128 T0.XYZW, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T0.X, KC0[2].Y,
; EG-NEXT: ALU clause starting at 11:
; EG-NEXT: MOV * T2.X, literal.x,
; EG-NEXT: 0(0.000000e+00), 0(0.000000e+00)
;
; GFX12-LABEL: constant_load_v16i16_align2:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b64 s[0:1], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v8, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_clause 0x7
; GFX12-NEXT: global_load_u16 v3, v8, s[0:1] offset:28
; GFX12-NEXT: global_load_u16 v2, v8, s[0:1] offset:24
; GFX12-NEXT: global_load_u16 v1, v8, s[0:1] offset:20
; GFX12-NEXT: global_load_u16 v0, v8, s[0:1] offset:16
; GFX12-NEXT: global_load_u16 v7, v8, s[0:1] offset:12
; GFX12-NEXT: global_load_u16 v6, v8, s[0:1] offset:8
; GFX12-NEXT: global_load_u16 v5, v8, s[0:1] offset:4
; GFX12-NEXT: global_load_u16 v4, v8, s[0:1]
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v3, v8, s[0:1] offset:30
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v2, v8, s[0:1] offset:26
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v1, v8, s[0:1] offset:22
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v0, v8, s[0:1] offset:18
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v7, v8, s[0:1] offset:14
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v6, v8, s[0:1] offset:10
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v5, v8, s[0:1] offset:6
; GFX12-NEXT: s_wait_loadcnt 0x7
; GFX12-NEXT: global_load_d16_hi_b16 v4, v8, s[0:1] offset:2
; GFX12-NEXT: s_wait_loadcnt 0x4
; GFX12-NEXT: global_store_b128 v[0:1], v[0:3], off
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: global_store_b128 v[0:1], v[4:7], off
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <16 x i16>, ptr addrspace(4) %ptr0, align 2
store <16 x i16> %ld, ptr addrspace(1) undef, align 32
ret void
}
define amdgpu_kernel void @constant_zextload_i16_to_i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_i16_to_i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_i16_to_i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dword v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_i16_to_i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_i16_to_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_i16_to_i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_u16 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%a = load i16, ptr addrspace(4) %in
%ext = zext i16 %a to i32
store i32 %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_i16_to_i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_i16_to_i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_sshort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_i16_to_i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_sshort v2, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dword v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_i16_to_i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_sshort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_i16_to_i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
;
; GFX12-LABEL: constant_sextload_i16_to_i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_i16 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%a = load i16, ptr addrspace(4) %in
%ext = sext i16 %a to i32
store i32 %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v1i16_to_v1i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v1i16_to_v1i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v1i16_to_v1i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_ushort v2, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dword v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v1i16_to_v1i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v1i16_to_v1i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 1, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v1i16_to_v1i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_u16 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <1 x i16>, ptr addrspace(4) %in
%ext = zext <1 x i16> %load to <1 x i32>
store <1 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v1i16_to_v1i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v1i16_to_v1i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_sshort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v1i16_to_v1i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_sshort v2, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dword v[0:1], v2
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v1i16_to_v1i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_sshort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: flat_store_dword v[0:1], v2
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v1i16_to_v1i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.X, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
;
; GFX12-LABEL: constant_sextload_v1i16_to_v1i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_i16 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: global_store_b32 v0, v1, s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <1 x i16>, ptr addrspace(4) %in
%ext = sext <1 x i16> %load to <1 x i32>
store <1 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v2i16_to_v2i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v2i16_to_v2i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s4, s2, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s4
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v2i16_to_v2i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s0, s2, 16
; GCN-HSA-NEXT: s_and_b32 s1, s2, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v2i16_to_v2i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s2, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v2i16_to_v2i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XY, T5.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T4.X, T4.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T4.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T4.Y, T4.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T4.X, T4.X, literal.x,
; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
;
; GFX12-LABEL: constant_zextload_v2i16_to_v2i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s3, s2, 0xffff
; GFX12-NEXT: s_lshr_b32 s2, s2, 16
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <2 x i16>, ptr addrspace(4) %in
%ext = zext <2 x i16> %load to <2 x i32>
store <2 x i32> %ext, ptr addrspace(1) %out
ret void
}
; TODO: We should use ASHR instead of LSHR + BFE
define amdgpu_kernel void @constant_sextload_v2i16_to_v2i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v2i16_to_v2i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s4, s2, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s4
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v2i16_to_v2i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s0, s2, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s1, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v2i16_to_v2i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s0, s2, 16
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s1, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v2i16_to_v2i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 5, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XY, T4.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T4.X, T4.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T4.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT T5.X, T4.X, 0.0, literal.x,
; EG-NEXT: LSHR T0.W, T4.X, literal.x,
; EG-NEXT: LSHR * T4.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
; EG-NEXT: BFE_INT * T5.Y, PV.W, 0.0, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
;
; GFX12-LABEL: constant_sextload_v2i16_to_v2i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_sext_i32_i16 s3, s2
; GFX12-NEXT: s_ashr_i32 s2, s2, 16
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <2 x i16>, ptr addrspace(4) %in
%ext = sext <2 x i16> %load to <2 x i32>
store <2 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v3i16_to_v3i32(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_zextload_v3i16_to_v3i32:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s6, s4, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:8
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s6
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v3i16_to_v3i32:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s0, s2, 16
; GCN-HSA-NEXT: s_and_b32 s1, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s1
; GCN-HSA-NEXT: flat_store_dwordx3 v[3:4], v[0:2]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v3i16_to_v3i32:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_and_b32 s0, s3, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s1, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx3 v[3:4], v[0:2]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v3i16_to_v3i32:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 4, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 2 @6
; EG-NEXT: ALU 2, @17, KC0[], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T2.X, T4.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T3.XY, T0.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T2.X, T1.X, 4, #1
; EG-NEXT: VTX_READ_16 T3.X, T1.X, 0, #1
; EG-NEXT: VTX_READ_16 T1.X, T1.X, 2, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: LSHR T0.X, KC0[2].Y, literal.x,
; EG-NEXT: MOV * T1.X, KC0[2].Z,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 8(1.121039e-44), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 17:
; EG-NEXT: LSHR T4.X, T0.W, literal.x,
; EG-NEXT: MOV * T3.Y, T1.X,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v3i16_to_v3i32:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: s_and_b32 s4, s2, 0xffff
; GFX12-NEXT: s_lshr_b32 s2, s2, 16
; GFX12-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v0, s4
; GFX12-NEXT: v_dual_mov_b32 v1, s2 :: v_dual_mov_b32 v2, s3
; GFX12-NEXT: global_store_b96 v3, v[0:2], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <3 x i16>, ptr addrspace(4) %in
%ext = zext <3 x i16> %ld to <3 x i32>
store <3 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v3i16_to_v3i32(ptr addrspace(1) %out, ptr addrspace(4) %in) {
; GCN-NOHSA-SI-LABEL: constant_sextload_v3i16_to_v3i32:
; GCN-NOHSA-SI: ; %bb.0: ; %entry
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s6, s4, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dword v0, off, s[0:3], 0 offset:8
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s6
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v3i16_to_v3i32:
; GCN-HSA: ; %bb.0: ; %entry
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s0, s2, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s1, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s2, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s1
; GCN-HSA-NEXT: flat_store_dwordx3 v[3:4], v[0:2]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v3i16_to_v3i32:
; GCN-NOHSA-VI: ; %bb.0: ; %entry
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s0, s2, 16
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s1, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: flat_store_dwordx3 v[3:4], v[0:2]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v3i16_to_v3i32:
; EG: ; %bb.0: ; %entry
; EG-NEXT: ALU 0, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 2 @6
; EG-NEXT: ALU 9, @13, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T2.X, T3.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T1.X, T0.X, 2, #1
; EG-NEXT: VTX_READ_16 T2.X, T0.X, 4, #1
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 13:
; EG-NEXT: BFE_INT * T0.Y, T1.X, 0.0, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
; EG-NEXT: BFE_INT T2.X, T2.X, 0.0, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 8(1.121039e-44)
; EG-NEXT: LSHR * T3.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_sextload_v3i16_to_v3i32:
; GFX12: ; %bb.0: ; %entry
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s4, s2, 16
; GFX12-NEXT: s_sext_i32_i16 s2, s2
; GFX12-NEXT: s_sext_i32_i16 s3, s3
; GFX12-NEXT: v_dual_mov_b32 v3, 0 :: v_dual_mov_b32 v0, s2
; GFX12-NEXT: v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s3
; GFX12-NEXT: global_store_b96 v3, v[0:2], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
entry:
%ld = load <3 x i16>, ptr addrspace(4) %in
%ext = sext <3 x i16> %ld to <3 x i32>
store <3 x i32> %ext, ptr addrspace(1) %out
ret void
}
; v4i16 is naturally 8 byte aligned
; TODO: This should use LD, but for some there are redundant MOVs
define amdgpu_kernel void @constant_zextload_v4i16_to_v4i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v4i16_to_v4i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s6, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s7, s4, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s6
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v4i16_to_v4i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s0, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s1, s2, 16
; GCN-HSA-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v4i16_to_v4i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s3, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s3, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s3, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v4i16_to_v4i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 8, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T6.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T5.XY, T5.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T5.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T5.W, T5.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT * T5.Z, T5.Y, literal.x,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: LSHR * T5.Y, T5.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T5.X, T5.X, literal.x,
; EG-NEXT: LSHR * T6.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
;
; GFX12-LABEL: constant_zextload_v4i16_to_v4i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s4, s3, 16
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: s_and_b32 s5, s2, 0xffff
; GFX12-NEXT: s_lshr_b32 s2, s2, 16
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s2
; GFX12-NEXT: v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v3, s4
; GFX12-NEXT: v_mov_b32_e32 v2, s3
; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <4 x i16>, ptr addrspace(4) %in
%ext = zext <4 x i16> %load to <4 x i32>
store <4 x i32> %ext, ptr addrspace(1) %out
ret void
}
; v4i16 is naturally 8 byte aligned
; TODO: This should use LD, but for some there are redundant MOVs
; TODO: We should use ASHR instead of LSHR + BFE
define amdgpu_kernel void @constant_sextload_v4i16_to_v4i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v4i16_to_v4i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s8, s4, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[6:7], s[4:5], 48
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s6
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v4i16_to_v4i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i64 s[0:1], s[2:3], 48
; GCN-HSA-NEXT: s_ashr_i32 s4, s2, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s1, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s2, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v4i16_to_v4i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s0, s3, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s1, s2, 16
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s3, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v4i16_to_v4i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 10, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T5.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T5.XY, T5.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T5.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT * T6.Z, T5.Y, 0.0, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T6.X, T5.X, 0.0, literal.x,
; EG-NEXT: LSHR * T0.W, T5.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T6.W, PV.W, 0.0, literal.x,
; EG-NEXT: LSHR * T0.W, T5.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T5.X, KC0[2].Y, literal.x,
; EG-NEXT: BFE_INT * T6.Y, PS, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
;
; GFX12-LABEL: constant_sextload_v4i16_to_v4i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s4, s3, 16
; GFX12-NEXT: s_ashr_i32 s5, s2, 16
; GFX12-NEXT: s_sext_i32_i16 s2, s2
; GFX12-NEXT: s_sext_i32_i16 s3, s3
; GFX12-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s5
; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, s4
; GFX12-NEXT: v_mov_b32_e32 v2, s3
; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <4 x i16>, ptr addrspace(4) %in
%ext = sext <4 x i16> %load to <4 x i32>
store <4 x i32> %ext, ptr addrspace(1) %out
ret void
}
; v8i16 is naturally 16 byte aligned
; TODO: These should use LSHR instead of BFE_UINT
; TODO: This should use DST, but for some there are redundant MOVs
define amdgpu_kernel void @constant_zextload_v8i16_to_v8i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v8i16_to_v8i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s8, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s9, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s10, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s11, s6, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s10
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s8
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v8i16_to_v8i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s8, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s9, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s2, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s3, s6, 16
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v8i16_to_v8i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s8, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s9, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s7, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v8i16_to_v8i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 17, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T10.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T7.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T7.XYZW, T7.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T7.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T8.W, T7.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT * T8.Z, T7.Y, literal.x,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: LSHR T8.Y, T7.X, literal.x,
; EG-NEXT: LSHR * T9.W, T7.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T8.X, T7.X, literal.x,
; EG-NEXT: AND_INT T9.Z, T7.W, literal.x,
; EG-NEXT: LSHR * T7.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
; EG-NEXT: LSHR * T9.Y, T7.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T9.X, T7.Z, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: LSHR * T10.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v8i16_to_v8i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b128 s[4:7], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s8, s7, 16
; GFX12-NEXT: s_and_b32 s7, s7, 0xffff
; GFX12-NEXT: s_and_b32 s9, s6, 0xffff
; GFX12-NEXT: s_lshr_b32 s6, s6, 16
; GFX12-NEXT: s_lshr_b32 s2, s5, 16
; GFX12-NEXT: s_and_b32 s3, s5, 0xffff
; GFX12-NEXT: s_lshr_b32 s5, s4, 16
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v1, s6
; GFX12-NEXT: v_dual_mov_b32 v0, s9 :: v_dual_mov_b32 v3, s8
; GFX12-NEXT: v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v5, s5
; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s2
; GFX12-NEXT: v_mov_b32_e32 v6, s3
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <8 x i16>, ptr addrspace(4) %in
%ext = zext <8 x i16> %load to <8 x i32>
store <8 x i32> %ext, ptr addrspace(1) %out
ret void
}
; v8i16 is naturally 16 byte aligned
; TODO: 4 of these should use ASHR instead of LSHR + BFE_INT
; TODO: This should use DST, but for some there are redundant MOVs
define amdgpu_kernel void @constant_sextload_v8i16_to_v8i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v8i16_to_v8i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s8, s5, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s9, s4, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s10, s7, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s11, s6, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s10
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s8
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v8i16_to_v8i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s8, s5, 16
; GCN-HSA-NEXT: s_ashr_i32 s9, s4, 16
; GCN-HSA-NEXT: s_ashr_i32 s2, s7, 16
; GCN-HSA-NEXT: s_ashr_i32 s3, s6, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: s_sext_i32_i16 s7, s7
; GCN-HSA-NEXT: s_sext_i32_i16 s6, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_sext_i32_i16 s5, s5
; GCN-HSA-NEXT: s_sext_i32_i16 s4, s4
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v8i16_to_v8i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s8, s5, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s9, s4, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s2, s7, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s3, s6, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v8i16_to_v8i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 19, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T10.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T7.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_128 T7.XYZW, T7.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T7.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT * T8.Z, T7.Y, 0.0, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T8.X, T7.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T9.Z, T7.W, 0.0, literal.x,
; EG-NEXT: LSHR * T0.W, T7.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T9.X, T7.Z, 0.0, literal.x,
; EG-NEXT: LSHR T0.Z, T7.W, literal.x,
; EG-NEXT: BFE_INT T8.W, PV.W, 0.0, literal.x,
; EG-NEXT: LSHR * T0.W, T7.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T7.X, KC0[2].Y, literal.x,
; EG-NEXT: BFE_INT T8.Y, PS, 0.0, literal.y,
; EG-NEXT: LSHR T1.Z, T7.Z, literal.y,
; EG-NEXT: BFE_INT T9.W, PV.Z, 0.0, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T10.X, PS, literal.x,
; EG-NEXT: BFE_INT * T9.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
;
; GFX12-LABEL: constant_sextload_v8i16_to_v8i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b128 s[4:7], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s8, s7, 16
; GFX12-NEXT: s_ashr_i32 s9, s6, 16
; GFX12-NEXT: s_sext_i32_i16 s6, s6
; GFX12-NEXT: s_sext_i32_i16 s7, s7
; GFX12-NEXT: s_ashr_i32 s2, s5, 16
; GFX12-NEXT: s_ashr_i32 s3, s4, 16
; GFX12-NEXT: s_sext_i32_i16 s5, s5
; GFX12-NEXT: s_sext_i32_i16 s4, s4
; GFX12-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v1, s9
; GFX12-NEXT: v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v3, s8
; GFX12-NEXT: v_dual_mov_b32 v2, s7 :: v_dual_mov_b32 v5, s3
; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s2
; GFX12-NEXT: v_mov_b32_e32 v6, s5
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <8 x i16>, ptr addrspace(4) %in
%ext = sext <8 x i16> %load to <8 x i32>
store <8 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v16i16_to_v16i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v16i16_to_v16i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s12, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s13, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s14, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s15, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s16, s9, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s17, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s18, s11, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s19, s10, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s18
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s16
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s14
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s12
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v16i16_to_v16i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s12, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s13, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s14, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s15, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s16, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s17, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s2, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s3, s10, 16
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-HSA-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s11
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s17
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s16
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s15
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v16i16_to_v16i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s12, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s13, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s14, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s15, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s16, s9, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s17, s8, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s11, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s11, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s11, s10, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s16
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v16i16_to_v16i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @8
; EG-NEXT: ALU 35, @13, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T18.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T11.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T16.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T12.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 8:
; EG-NEXT: VTX_READ_128 T12.XYZW, T11.X, 0, #1
; EG-NEXT: VTX_READ_128 T11.XYZW, T11.X, 16, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: MOV * T11.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 13:
; EG-NEXT: LSHR * T13.W, T12.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT * T13.Z, T12.Y, literal.x,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: LSHR T13.Y, T12.X, literal.x,
; EG-NEXT: LSHR * T14.W, T12.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T13.X, T12.X, literal.x,
; EG-NEXT: AND_INT T14.Z, T12.W, literal.x,
; EG-NEXT: LSHR * T12.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
; EG-NEXT: LSHR T14.Y, T12.Z, literal.x,
; EG-NEXT: LSHR * T15.W, T11.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T14.X, T12.Z, literal.x,
; EG-NEXT: AND_INT T15.Z, T11.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: LSHR T16.X, PV.W, literal.x,
; EG-NEXT: LSHR T15.Y, T11.X, literal.y,
; EG-NEXT: LSHR T17.W, T11.W, literal.y,
; EG-NEXT: AND_INT * T15.X, T11.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T17.Z, T11.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 32(4.484155e-44)
; EG-NEXT: LSHR T11.X, PV.W, literal.x,
; EG-NEXT: LSHR T17.Y, T11.Z, literal.y,
; EG-NEXT: AND_INT * T17.X, T11.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00)
; EG-NEXT: LSHR * T18.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v16i16_to_v16i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b256 s[4:11], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s16, s11, 16
; GFX12-NEXT: s_and_b32 s11, s11, 0xffff
; GFX12-NEXT: s_and_b32 s17, s10, 0xffff
; GFX12-NEXT: s_lshr_b32 s10, s10, 16
; GFX12-NEXT: s_lshr_b32 s14, s9, 16
; GFX12-NEXT: s_and_b32 s9, s9, 0xffff
; GFX12-NEXT: s_lshr_b32 s15, s8, 16
; GFX12-NEXT: s_and_b32 s8, s8, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v1, s10
; GFX12-NEXT: s_lshr_b32 s12, s7, 16
; GFX12-NEXT: s_and_b32 s7, s7, 0xffff
; GFX12-NEXT: s_lshr_b32 s13, s6, 16
; GFX12-NEXT: s_and_b32 s6, s6, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v0, s17 :: v_dual_mov_b32 v3, s16
; GFX12-NEXT: v_dual_mov_b32 v2, s11 :: v_dual_mov_b32 v5, s15
; GFX12-NEXT: s_lshr_b32 s2, s5, 16
; GFX12-NEXT: s_and_b32 s3, s5, 0xffff
; GFX12-NEXT: s_lshr_b32 s5, s4, 16
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v4, s8 :: v_dual_mov_b32 v7, s14
; GFX12-NEXT: v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v9, s13
; GFX12-NEXT: v_dual_mov_b32 v8, s6 :: v_dual_mov_b32 v11, s12
; GFX12-NEXT: v_dual_mov_b32 v10, s7 :: v_dual_mov_b32 v13, s5
; GFX12-NEXT: v_dual_mov_b32 v12, s4 :: v_dual_mov_b32 v15, s2
; GFX12-NEXT: v_mov_b32_e32 v14, s3
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:48
; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1] offset:32
; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <16 x i16>, ptr addrspace(4) %in
%ext = zext <16 x i16> %load to <16 x i32>
store <16 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v16i16_to_v16i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v16i16_to_v16i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s12, s5, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s13, s4, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s14, s7, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s15, s6, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s16, s9, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s17, s8, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s9, s9
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s18, s11, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s19, s10, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s18
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s16
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s14
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s12
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v16i16_to_v16i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s12, s5, 16
; GCN-HSA-NEXT: s_ashr_i32 s13, s4, 16
; GCN-HSA-NEXT: s_ashr_i32 s14, s7, 16
; GCN-HSA-NEXT: s_ashr_i32 s15, s6, 16
; GCN-HSA-NEXT: s_ashr_i32 s16, s9, 16
; GCN-HSA-NEXT: s_ashr_i32 s17, s8, 16
; GCN-HSA-NEXT: s_ashr_i32 s2, s11, 16
; GCN-HSA-NEXT: s_ashr_i32 s3, s10, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s11, s11
; GCN-HSA-NEXT: s_sext_i32_i16 s10, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s11
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s9, s9
; GCN-HSA-NEXT: s_sext_i32_i16 s8, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s17
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s16
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: s_sext_i32_i16 s7, s7
; GCN-HSA-NEXT: s_sext_i32_i16 s6, s6
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s15
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_sext_i32_i16 s5, s5
; GCN-HSA-NEXT: s_sext_i32_i16 s4, s4
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v16i16_to_v16i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s12, s5, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s13, s4, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s14, s7, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s15, s6, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s16, s9, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s17, s8, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s2, s11, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s3, s10, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s9, s9
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s16
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v16i16_to_v16i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @12, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @8
; EG-NEXT: ALU 39, @13, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T12.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T11.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T14.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T13.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 8:
; EG-NEXT: VTX_READ_128 T12.XYZW, T11.X, 16, #1
; EG-NEXT: VTX_READ_128 T11.XYZW, T11.X, 0, #1
; EG-NEXT: ALU clause starting at 12:
; EG-NEXT: MOV * T11.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 13:
; EG-NEXT: LSHR T13.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T14.X, PV.W, literal.x,
; EG-NEXT: BFE_INT * T15.Z, T11.Y, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: BFE_INT T15.X, T11.X, 0.0, literal.x,
; EG-NEXT: LSHR T0.Y, T12.W, literal.x,
; EG-NEXT: BFE_INT T16.Z, T11.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR T0.W, T12.Y, literal.x,
; EG-NEXT: LSHR * T1.W, T11.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T16.X, T11.Z, 0.0, literal.x,
; EG-NEXT: LSHR T1.Y, T11.W, literal.x,
; EG-NEXT: BFE_INT T17.Z, T12.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T15.W, PS, 0.0, literal.x,
; EG-NEXT: LSHR * T1.W, T11.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T17.X, T12.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T15.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T18.Z, T12.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T16.W, PV.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T1.W, T11.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T18.X, T12.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T16.Y, PS, 0.0, literal.x,
; EG-NEXT: LSHR T0.Z, T12.X, literal.x,
; EG-NEXT: BFE_INT T17.W, T0.W, 0.0, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 32(4.484155e-44)
; EG-NEXT: LSHR T11.X, PS, literal.x,
; EG-NEXT: BFE_INT T17.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: LSHR T0.Z, T12.Z, literal.y,
; EG-NEXT: BFE_INT T18.W, T0.Y, 0.0, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T12.X, PS, literal.x,
; EG-NEXT: BFE_INT * T18.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
;
; GFX12-LABEL: constant_sextload_v16i16_to_v16i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b256 s[4:11], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s16, s11, 16
; GFX12-NEXT: s_ashr_i32 s17, s10, 16
; GFX12-NEXT: s_sext_i32_i16 s10, s10
; GFX12-NEXT: s_sext_i32_i16 s11, s11
; GFX12-NEXT: s_ashr_i32 s14, s9, 16
; GFX12-NEXT: s_ashr_i32 s15, s8, 16
; GFX12-NEXT: s_sext_i32_i16 s9, s9
; GFX12-NEXT: s_sext_i32_i16 s8, s8
; GFX12-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v1, s17
; GFX12-NEXT: s_ashr_i32 s12, s7, 16
; GFX12-NEXT: s_ashr_i32 s13, s6, 16
; GFX12-NEXT: s_sext_i32_i16 s7, s7
; GFX12-NEXT: s_sext_i32_i16 s6, s6
; GFX12-NEXT: v_dual_mov_b32 v0, s10 :: v_dual_mov_b32 v3, s16
; GFX12-NEXT: v_dual_mov_b32 v2, s11 :: v_dual_mov_b32 v5, s15
; GFX12-NEXT: s_ashr_i32 s2, s5, 16
; GFX12-NEXT: s_ashr_i32 s3, s4, 16
; GFX12-NEXT: s_sext_i32_i16 s5, s5
; GFX12-NEXT: s_sext_i32_i16 s4, s4
; GFX12-NEXT: v_dual_mov_b32 v4, s8 :: v_dual_mov_b32 v7, s14
; GFX12-NEXT: v_dual_mov_b32 v6, s9 :: v_dual_mov_b32 v9, s13
; GFX12-NEXT: v_dual_mov_b32 v8, s6 :: v_dual_mov_b32 v11, s12
; GFX12-NEXT: v_dual_mov_b32 v10, s7 :: v_dual_mov_b32 v13, s3
; GFX12-NEXT: v_dual_mov_b32 v12, s4 :: v_dual_mov_b32 v15, s2
; GFX12-NEXT: v_mov_b32_e32 v14, s5
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:48
; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1] offset:32
; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <16 x i16>, ptr addrspace(4) %in
%ext = sext <16 x i16> %load to <16 x i32>
store <16 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v32i16_to_v32i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v32i16_to_v32i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s19, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s18, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s20, s1, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s21, s0, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s22, s3, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s23, s2, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s24, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s25, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s26, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s27, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s28, s9, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s29, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s30, s11, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s31, s10, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s33, s13, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s34, s12, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s35, s15, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s36, s14, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s15, s15, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s36
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s30
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s28
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s27
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s26
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s25
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s24
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s22
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s21
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s20
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v32i16_to_v32i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s18, s1, 16
; GCN-HSA-NEXT: s_lshr_b32 s19, s0, 16
; GCN-HSA-NEXT: s_lshr_b32 s20, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s21, s2, 16
; GCN-HSA-NEXT: s_lshr_b32 s22, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s23, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s24, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s25, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s26, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s27, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s28, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s29, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s30, s13, 16
; GCN-HSA-NEXT: s_lshr_b32 s31, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s33, s15, 16
; GCN-HSA-NEXT: s_lshr_b32 s34, s14, 16
; GCN-HSA-NEXT: s_and_b32 s35, s1, 0xffff
; GCN-HSA-NEXT: s_and_b32 s36, s0, 0xffff
; GCN-HSA-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-HSA-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-HSA-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-HSA-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-HSA-NEXT: s_and_b32 s0, s15, 0xffff
; GCN-HSA-NEXT: s_and_b32 s1, s14, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s1
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s33
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s30
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s28
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s26
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s24
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s22
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s36
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v32i16_to_v32i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s18, s1, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s19, s1, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s20, s0, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s21, s0, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s22, s3, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s23, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s24, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s25, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s26, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s27, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s28, s9, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s29, s8, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s30, s11, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s31, s10, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s33, s13, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s34, s12, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s15, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s15, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s15, s14, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x70
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s30
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s28
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s26
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s24
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s17
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v32i16_to_v32i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @20, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @12
; EG-NEXT: ALU 71, @21, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T34.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T21.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T32.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T22.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T29.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T19.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T26.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T20.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 12:
; EG-NEXT: VTX_READ_128 T20.XYZW, T19.X, 0, #1
; EG-NEXT: VTX_READ_128 T21.XYZW, T19.X, 48, #1
; EG-NEXT: VTX_READ_128 T22.XYZW, T19.X, 32, #1
; EG-NEXT: VTX_READ_128 T19.XYZW, T19.X, 16, #1
; EG-NEXT: ALU clause starting at 20:
; EG-NEXT: MOV * T19.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 21:
; EG-NEXT: LSHR * T23.W, T20.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT * T23.Z, T20.Y, literal.x,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: LSHR T23.Y, T20.X, literal.x,
; EG-NEXT: LSHR * T24.W, T20.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T23.X, T20.X, literal.x,
; EG-NEXT: AND_INT T24.Z, T20.W, literal.x,
; EG-NEXT: LSHR * T20.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
; EG-NEXT: LSHR T24.Y, T20.Z, literal.x,
; EG-NEXT: LSHR * T25.W, T19.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T24.X, T20.Z, literal.x,
; EG-NEXT: AND_INT T25.Z, T19.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: LSHR T26.X, PV.W, literal.x,
; EG-NEXT: LSHR T25.Y, T19.X, literal.y,
; EG-NEXT: LSHR T27.W, T19.W, literal.y,
; EG-NEXT: AND_INT * T25.X, T19.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T27.Z, T19.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 32(4.484155e-44)
; EG-NEXT: LSHR T19.X, PV.W, literal.x,
; EG-NEXT: LSHR T27.Y, T19.Z, literal.y,
; EG-NEXT: LSHR T28.W, T22.Y, literal.y,
; EG-NEXT: AND_INT * T27.X, T19.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T28.Z, T22.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 48(6.726233e-44)
; EG-NEXT: LSHR T29.X, PV.W, literal.x,
; EG-NEXT: LSHR T28.Y, T22.X, literal.y,
; EG-NEXT: LSHR T30.W, T22.W, literal.y,
; EG-NEXT: AND_INT * T28.X, T22.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T30.Z, T22.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 64(8.968310e-44)
; EG-NEXT: LSHR T22.X, PV.W, literal.x,
; EG-NEXT: LSHR T30.Y, T22.Z, literal.y,
; EG-NEXT: LSHR T31.W, T21.Y, literal.y,
; EG-NEXT: AND_INT * T30.X, T22.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T31.Z, T21.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 80(1.121039e-43)
; EG-NEXT: LSHR T32.X, PV.W, literal.x,
; EG-NEXT: LSHR T31.Y, T21.X, literal.y,
; EG-NEXT: LSHR T33.W, T21.W, literal.y,
; EG-NEXT: AND_INT * T31.X, T21.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T33.Z, T21.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 96(1.345247e-43)
; EG-NEXT: LSHR T21.X, PV.W, literal.x,
; EG-NEXT: LSHR T33.Y, T21.Z, literal.y,
; EG-NEXT: AND_INT * T33.X, T21.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00)
; EG-NEXT: LSHR * T34.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v32i16_to_v32i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[16:19], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b512 s[0:15], s[18:19], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s33, s15, 16
; GFX12-NEXT: s_and_b32 s15, s15, 0xffff
; GFX12-NEXT: s_and_b32 s34, s14, 0xffff
; GFX12-NEXT: s_lshr_b32 s14, s14, 16
; GFX12-NEXT: s_lshr_b32 s30, s13, 16
; GFX12-NEXT: s_and_b32 s13, s13, 0xffff
; GFX12-NEXT: s_lshr_b32 s31, s12, 16
; GFX12-NEXT: s_and_b32 s12, s12, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s14
; GFX12-NEXT: v_dual_mov_b32 v0, s34 :: v_dual_mov_b32 v3, s33
; GFX12-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s31
; GFX12-NEXT: s_lshr_b32 s29, s10, 16
; GFX12-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s30
; GFX12-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s29
; GFX12-NEXT: s_lshr_b32 s28, s11, 16
; GFX12-NEXT: s_and_b32 s11, s11, 0xffff
; GFX12-NEXT: s_and_b32 s10, s10, 0xffff
; GFX12-NEXT: s_lshr_b32 s26, s9, 16
; GFX12-NEXT: s_and_b32 s9, s9, 0xffff
; GFX12-NEXT: s_lshr_b32 s27, s8, 16
; GFX12-NEXT: s_and_b32 s8, s8, 0xffff
; GFX12-NEXT: s_lshr_b32 s24, s7, 16
; GFX12-NEXT: s_and_b32 s7, s7, 0xffff
; GFX12-NEXT: s_lshr_b32 s25, s6, 16
; GFX12-NEXT: s_and_b32 s6, s6, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s28
; GFX12-NEXT: v_mov_b32_e32 v10, s11
; GFX12-NEXT: s_lshr_b32 s22, s5, 16
; GFX12-NEXT: s_and_b32 s5, s5, 0xffff
; GFX12-NEXT: s_lshr_b32 s23, s4, 16
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:112
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:96
; GFX12-NEXT: v_dual_mov_b32 v1, s27 :: v_dual_mov_b32 v0, s8
; GFX12-NEXT: v_dual_mov_b32 v3, s26 :: v_dual_mov_b32 v2, s9
; GFX12-NEXT: v_mov_b32_e32 v5, s25
; GFX12-NEXT: s_lshr_b32 s20, s3, 16
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: s_lshr_b32 s21, s2, 16
; GFX12-NEXT: s_and_b32 s2, s2, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v4, s6 :: v_dual_mov_b32 v7, s24
; GFX12-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v13, s23
; GFX12-NEXT: s_lshr_b32 s18, s1, 16
; GFX12-NEXT: s_and_b32 s1, s1, 0xffff
; GFX12-NEXT: s_lshr_b32 s19, s0, 16
; GFX12-NEXT: s_and_b32 s0, s0, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v12, s4 :: v_dual_mov_b32 v15, s22
; GFX12-NEXT: v_dual_mov_b32 v14, s5 :: v_dual_mov_b32 v17, s21
; GFX12-NEXT: v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s20
; GFX12-NEXT: v_dual_mov_b32 v18, s3 :: v_dual_mov_b32 v21, s19
; GFX12-NEXT: v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v23, s18
; GFX12-NEXT: v_mov_b32_e32 v22, s1
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[16:17] offset:80
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:64
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:48
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[16:17] offset:32
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[16:17] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[16:17]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <32 x i16>, ptr addrspace(4) %in
%ext = zext <32 x i16> %load to <32 x i32>
store <32 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v32i16_to_v32i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v32i16_to_v32i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s19, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s18, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s20, s1, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s21, s0, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s1, s1
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s0, s0
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s22, s3, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s23, s2, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s3, s3
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s2, s2
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s24, s5, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s25, s4, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s26, s7, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s27, s6, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s28, s9, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s29, s8, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s9, s9
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s30, s11, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s31, s10, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s33, s13, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s34, s12, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s13, s13
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s35, s15, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s36, s14, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s15, s15
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s14, s14
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s12, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s36
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s30
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s28
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s27
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s26
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s25
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s24
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s22
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s21
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s20
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v32i16_to_v32i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s18, s1, 16
; GCN-HSA-NEXT: s_ashr_i32 s19, s0, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s20, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s21, s0
; GCN-HSA-NEXT: s_ashr_i32 s22, s3, 16
; GCN-HSA-NEXT: s_ashr_i32 s23, s2, 16
; GCN-HSA-NEXT: s_ashr_i32 s24, s5, 16
; GCN-HSA-NEXT: s_ashr_i32 s25, s4, 16
; GCN-HSA-NEXT: s_ashr_i32 s26, s7, 16
; GCN-HSA-NEXT: s_ashr_i32 s27, s6, 16
; GCN-HSA-NEXT: s_ashr_i32 s28, s9, 16
; GCN-HSA-NEXT: s_ashr_i32 s29, s8, 16
; GCN-HSA-NEXT: s_ashr_i32 s30, s11, 16
; GCN-HSA-NEXT: s_ashr_i32 s31, s10, 16
; GCN-HSA-NEXT: s_ashr_i32 s33, s13, 16
; GCN-HSA-NEXT: s_ashr_i32 s34, s12, 16
; GCN-HSA-NEXT: s_ashr_i32 s0, s15, 16
; GCN-HSA-NEXT: s_ashr_i32 s1, s14, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s1
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s15, s15
; GCN-HSA-NEXT: s_sext_i32_i16 s14, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s15
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s13, s13
; GCN-HSA-NEXT: s_sext_i32_i16 s12, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s33
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s11, s11
; GCN-HSA-NEXT: s_sext_i32_i16 s10, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s30
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s9, s9
; GCN-HSA-NEXT: s_sext_i32_i16 s8, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s28
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s7, s7
; GCN-HSA-NEXT: s_sext_i32_i16 s6, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s26
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s5, s5
; GCN-HSA-NEXT: s_sext_i32_i16 s4, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s24
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: s_sext_i32_i16 s3, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s2, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s22
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v32i16_to_v32i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s18, s1, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s19, s0, 16
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s20, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s21, s0
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s22, s3, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s23, s2, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s24, s5, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s25, s4, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s26, s7, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s27, s6, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s28, s9, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s29, s8, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s30, s11, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s31, s10, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s33, s13, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s34, s12, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s0, s15, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s1, s14, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x70
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s15, s15
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s14, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s13, s13
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s12, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s30
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s9, s9
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s28
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s26
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s16, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s24
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s17, 0
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s3, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s2, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s17
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v32i16_to_v32i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 8, @20, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @12
; EG-NEXT: ALU 73, @29, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T24.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T22.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T28.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T27.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T26.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T21.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T20.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T19.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 12:
; EG-NEXT: VTX_READ_128 T23.XYZW, T22.X, 16, #1
; EG-NEXT: VTX_READ_128 T24.XYZW, T22.X, 32, #1
; EG-NEXT: VTX_READ_128 T25.XYZW, T22.X, 0, #1
; EG-NEXT: VTX_READ_128 T22.XYZW, T22.X, 48, #1
; EG-NEXT: ALU clause starting at 20:
; EG-NEXT: LSHR T19.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T20.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T21.X, PV.W, literal.x,
; EG-NEXT: MOV * T22.X, KC0[2].Z,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 29:
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T26.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T27.X, PV.W, literal.x,
; EG-NEXT: LSHR T0.W, T22.W, literal.y,
; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 80(1.121039e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T28.X, PS, literal.x,
; EG-NEXT: LSHR T0.Y, T22.Y, literal.y,
; EG-NEXT: BFE_INT T29.Z, T25.Y, 0.0, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: LSHR T1.W, T24.W, literal.y,
; EG-NEXT: LSHR * T2.W, T24.Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: BFE_INT T29.X, T25.X, 0.0, literal.x,
; EG-NEXT: LSHR T1.Y, T23.W, literal.x,
; EG-NEXT: BFE_INT T30.Z, T25.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR T3.W, T23.Y, literal.x,
; EG-NEXT: LSHR * T4.W, T25.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T30.X, T25.Z, 0.0, literal.x,
; EG-NEXT: LSHR T2.Y, T25.W, literal.x,
; EG-NEXT: BFE_INT T31.Z, T23.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T29.W, PS, 0.0, literal.x,
; EG-NEXT: LSHR * T4.W, T25.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T31.X, T23.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T29.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T32.Z, T23.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T30.W, PV.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T4.W, T25.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T32.X, T23.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T30.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T25.Z, T24.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T31.W, T3.W, 0.0, literal.x,
; EG-NEXT: LSHR * T3.W, T23.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T25.X, T24.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T31.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T33.Z, T24.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T32.W, T1.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T3.W, T23.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T33.X, T24.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T32.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T23.Z, T22.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T25.W, T2.W, 0.0, literal.x,
; EG-NEXT: LSHR * T2.W, T24.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T23.X, T22.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T25.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T34.Z, T22.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T33.W, T1.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T1.W, T24.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T34.X, T22.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T33.Y, PS, 0.0, literal.x,
; EG-NEXT: LSHR T0.Z, T22.X, literal.x,
; EG-NEXT: BFE_INT T23.W, T0.Y, 0.0, literal.x,
; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 96(1.345247e-43)
; EG-NEXT: LSHR T22.X, PS, literal.x,
; EG-NEXT: BFE_INT T23.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: LSHR T0.Z, T22.Z, literal.y,
; EG-NEXT: BFE_INT T34.W, T0.W, 0.0, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T24.X, PS, literal.x,
; EG-NEXT: BFE_INT * T34.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
;
; GFX12-LABEL: constant_sextload_v32i16_to_v32i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[16:19], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b512 s[0:15], s[18:19], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s33, s15, 16
; GFX12-NEXT: s_ashr_i32 s34, s14, 16
; GFX12-NEXT: s_sext_i32_i16 s14, s14
; GFX12-NEXT: s_sext_i32_i16 s15, s15
; GFX12-NEXT: s_ashr_i32 s30, s13, 16
; GFX12-NEXT: s_ashr_i32 s31, s12, 16
; GFX12-NEXT: s_sext_i32_i16 s13, s13
; GFX12-NEXT: s_sext_i32_i16 s12, s12
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s34
; GFX12-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s33
; GFX12-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s31
; GFX12-NEXT: s_ashr_i32 s29, s10, 16
; GFX12-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s30
; GFX12-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s29
; GFX12-NEXT: s_ashr_i32 s28, s11, 16
; GFX12-NEXT: s_sext_i32_i16 s11, s11
; GFX12-NEXT: s_sext_i32_i16 s10, s10
; GFX12-NEXT: s_ashr_i32 s26, s9, 16
; GFX12-NEXT: s_ashr_i32 s27, s8, 16
; GFX12-NEXT: s_sext_i32_i16 s9, s9
; GFX12-NEXT: s_sext_i32_i16 s8, s8
; GFX12-NEXT: s_ashr_i32 s24, s7, 16
; GFX12-NEXT: s_ashr_i32 s25, s6, 16
; GFX12-NEXT: s_sext_i32_i16 s7, s7
; GFX12-NEXT: s_sext_i32_i16 s6, s6
; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s28
; GFX12-NEXT: v_mov_b32_e32 v10, s11
; GFX12-NEXT: s_ashr_i32 s22, s5, 16
; GFX12-NEXT: s_ashr_i32 s23, s4, 16
; GFX12-NEXT: s_sext_i32_i16 s5, s5
; GFX12-NEXT: s_sext_i32_i16 s4, s4
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:112
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:96
; GFX12-NEXT: v_dual_mov_b32 v1, s27 :: v_dual_mov_b32 v0, s8
; GFX12-NEXT: v_dual_mov_b32 v3, s26 :: v_dual_mov_b32 v2, s9
; GFX12-NEXT: v_mov_b32_e32 v5, s25
; GFX12-NEXT: s_ashr_i32 s20, s3, 16
; GFX12-NEXT: s_ashr_i32 s21, s2, 16
; GFX12-NEXT: s_sext_i32_i16 s3, s3
; GFX12-NEXT: s_sext_i32_i16 s2, s2
; GFX12-NEXT: v_dual_mov_b32 v4, s6 :: v_dual_mov_b32 v7, s24
; GFX12-NEXT: v_dual_mov_b32 v6, s7 :: v_dual_mov_b32 v13, s23
; GFX12-NEXT: s_ashr_i32 s18, s1, 16
; GFX12-NEXT: s_ashr_i32 s19, s0, 16
; GFX12-NEXT: s_sext_i32_i16 s1, s1
; GFX12-NEXT: s_sext_i32_i16 s0, s0
; GFX12-NEXT: v_dual_mov_b32 v12, s4 :: v_dual_mov_b32 v15, s22
; GFX12-NEXT: v_dual_mov_b32 v14, s5 :: v_dual_mov_b32 v17, s21
; GFX12-NEXT: v_dual_mov_b32 v16, s2 :: v_dual_mov_b32 v19, s20
; GFX12-NEXT: v_dual_mov_b32 v18, s3 :: v_dual_mov_b32 v21, s19
; GFX12-NEXT: v_dual_mov_b32 v20, s0 :: v_dual_mov_b32 v23, s18
; GFX12-NEXT: v_mov_b32_e32 v22, s1
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[16:17] offset:80
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:64
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:48
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[16:17] offset:32
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[16:17] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[16:17]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <32 x i16>, ptr addrspace(4) %in
%ext = sext <32 x i16> %load to <32 x i32>
store <32 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v64i16_to_v64i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v64i16_to_v64i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[36:39], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x0
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x10
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s33, s1, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s34, s0, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s35, s3, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s38, s2, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s41, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s42, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s45, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s46, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s47, s9, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s48, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s49, s11, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s50, s10, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s51, s13, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s52, s12, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s53, s15, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s54, s14, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s39, s1, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s40, s0, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s43, s3, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s44, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s15, s15, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s55, s17, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s56, s16, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s57, s19, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s58, s18, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s59, s21, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s60, s20, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s61, s23, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s62, s22, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s63, s25, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s64, s24, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s65, s27, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s66, s26, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s67, s29, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s68, s28, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s69, s31, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s70, s30, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s17, s17, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s16, s16, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s19, s19, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s18, s18, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s21, s21, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s20, s20, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s23, s23, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s25, s25, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s24, s24, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s27, s27, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s29, s29, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s28, s28, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s31, s31, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s30, s30, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s26, s26, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s22, s22, 0xffff
; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s36
; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s37
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s30
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s70
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s69
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s28
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s68
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s67
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s26
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s66
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s27
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s65
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, s24
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, s64
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, s25
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:224
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, s63
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(3)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s62
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s23
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s61
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:176
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s20
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s60
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s59
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s58
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s57
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s56
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s55
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s54
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s53
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s52
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s51
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s50
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s49
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s48
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s47
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s46
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s45
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s42
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s41
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s44
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s38
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s43
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s40
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s39
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v64i16_to_v64i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s20, s1, 16
; GCN-HSA-NEXT: s_lshr_b32 s21, s0, 16
; GCN-HSA-NEXT: s_lshr_b32 s22, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s23, s2, 16
; GCN-HSA-NEXT: s_lshr_b32 s24, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s26, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s28, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s30, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s33, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s35, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s37, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s39, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s41, s13, 16
; GCN-HSA-NEXT: s_lshr_b32 s43, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s45, s15, 16
; GCN-HSA-NEXT: s_lshr_b32 s47, s14, 16
; GCN-HSA-NEXT: s_and_b32 s25, s1, 0xffff
; GCN-HSA-NEXT: s_and_b32 s27, s0, 0xffff
; GCN-HSA-NEXT: s_and_b32 s29, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s31, s2, 0xffff
; GCN-HSA-NEXT: s_and_b32 s34, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s36, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s38, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s40, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s42, s9, 0xffff
; GCN-HSA-NEXT: s_and_b32 s44, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s46, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s48, s10, 0xffff
; GCN-HSA-NEXT: s_and_b32 s49, s13, 0xffff
; GCN-HSA-NEXT: s_and_b32 s50, s12, 0xffff
; GCN-HSA-NEXT: s_and_b32 s51, s15, 0xffff
; GCN-HSA-NEXT: s_and_b32 s52, s14, 0xffff
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x10
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s18, s1, 16
; GCN-HSA-NEXT: s_lshr_b32 s19, s0, 16
; GCN-HSA-NEXT: s_lshr_b32 s53, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s54, s2, 16
; GCN-HSA-NEXT: s_lshr_b32 s55, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s56, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s57, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s58, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s59, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s60, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s61, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s62, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s63, s13, 16
; GCN-HSA-NEXT: s_lshr_b32 s64, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s65, s15, 16
; GCN-HSA-NEXT: s_lshr_b32 s66, s14, 16
; GCN-HSA-NEXT: s_and_b32 s67, s1, 0xffff
; GCN-HSA-NEXT: s_and_b32 s68, s0, 0xffff
; GCN-HSA-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-HSA-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-HSA-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-HSA-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-HSA-NEXT: s_and_b32 s15, s15, 0xffff
; GCN-HSA-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xf0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xe0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v25, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v24, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xd0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v27, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v26, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xc0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v29, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v28, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xb0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v31, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v30, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xa0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v33, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v32, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x90
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s64
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s63
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[24:25], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v25, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v24, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x80
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v35, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v34, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s66
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s15
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s65
; GCN-HSA-NEXT: v_mov_b32_e32 v8, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v9, s62
; GCN-HSA-NEXT: v_mov_b32_e32 v10, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v11, s61
; GCN-HSA-NEXT: v_mov_b32_e32 v12, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v13, s60
; GCN-HSA-NEXT: v_mov_b32_e32 v14, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v15, s59
; GCN-HSA-NEXT: v_mov_b32_e32 v16, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v17, s58
; GCN-HSA-NEXT: v_mov_b32_e32 v18, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v19, s57
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s68
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s67
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s18
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v20, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v21, s56
; GCN-HSA-NEXT: flat_store_dwordx4 v[22:23], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s55
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s54
; GCN-HSA-NEXT: flat_store_dwordx4 v[26:27], v[8:11]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s3
; GCN-HSA-NEXT: flat_store_dwordx4 v[28:29], v[12:15]
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s53
; GCN-HSA-NEXT: flat_store_dwordx4 v[30:31], v[16:19]
; GCN-HSA-NEXT: flat_store_dwordx4 v[32:33], v[20:23]
; GCN-HSA-NEXT: flat_store_dwordx4 v[24:25], v[0:3]
; GCN-HSA-NEXT: flat_store_dwordx4 v[34:35], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s52
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s47
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s51
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s45
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s50
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s43
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s49
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s41
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s48
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s39
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s46
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s37
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s44
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s42
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s33
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s40
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s30
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s38
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s28
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s36
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s26
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s24
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s22
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v64i16_to_v64i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[36:39], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x0
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x40
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s33, s1, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s34, s0, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s35, s3, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s40, s2, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s41, s5, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s42, s4, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s43, s7, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s44, s6, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s45, s9, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s46, s8, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s47, s11, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s48, s10, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s38, s13, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s39, s12, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s49, s15, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s50, s14, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s51, s17, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s52, s16, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s53, s19, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s54, s18, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s55, s21, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s56, s20, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s57, s23, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s58, s22, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s59, s25, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s60, s24, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s61, s27, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s62, s26, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s63, s29, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s64, s28, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s65, s31, 16
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s66, s30, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s67, s11, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s68, s10, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s15, s15, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s17, s17, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s16, s16, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s19, s19, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s18, s18, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s21, s21, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s20, s20, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s23, s23, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s22, s22, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s25, s25, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s24, s24, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s27, s27, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s26, s26, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s29, s29, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s28, s28, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s10, s31, 0xffff
; GCN-NOHSA-VI-NEXT: s_and_b32 s11, s30, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xf0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s11
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s66
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s65
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xe0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s63
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xd0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s62
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s61
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xc0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s59
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xb0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s58
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s57
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0xa0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s56
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s55
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0x90
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s54
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s53
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0x80
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s52
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s51
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0x70
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s49
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0x60
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s39
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s38
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s36, 0x50
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s68
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s67
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s47
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: s_add_u32 s8, s36, 64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: s_addc_u32 s9, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s46
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s45
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s9
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s36, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s44
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s43
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s36, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s42
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s41
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s36, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s40
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s36
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s37
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v64i16_to_v64i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @38, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @22
; EG-NEXT: ALU 55, @39, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 3 @30
; EG-NEXT: ALU 87, @95, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T65.XYZW, T66.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T63.XYZW, T49.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T62.XYZW, T64.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T60.XYZW, T50.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T59.XYZW, T61.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T57.XYZW, T51.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T56.XYZW, T58.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T54.XYZW, T52.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T53.XYZW, T55.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T39.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T48.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T40.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T46.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T41.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T36.XYZW, T43.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T38.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 22:
; EG-NEXT: VTX_READ_128 T38.XYZW, T37.X, 0, #1
; EG-NEXT: VTX_READ_128 T39.XYZW, T37.X, 48, #1
; EG-NEXT: VTX_READ_128 T40.XYZW, T37.X, 32, #1
; EG-NEXT: VTX_READ_128 T41.XYZW, T37.X, 16, #1
; EG-NEXT: Fetch clause starting at 30:
; EG-NEXT: VTX_READ_128 T49.XYZW, T37.X, 112, #1
; EG-NEXT: VTX_READ_128 T50.XYZW, T37.X, 96, #1
; EG-NEXT: VTX_READ_128 T51.XYZW, T37.X, 80, #1
; EG-NEXT: VTX_READ_128 T52.XYZW, T37.X, 64, #1
; EG-NEXT: ALU clause starting at 38:
; EG-NEXT: MOV * T37.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 39:
; EG-NEXT: LSHR * T35.W, T38.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT * T35.Z, T38.Y, literal.x,
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: LSHR T35.Y, T38.X, literal.x,
; EG-NEXT: LSHR * T36.W, T38.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T35.X, T38.X, literal.x,
; EG-NEXT: AND_INT T36.Z, T38.W, literal.x,
; EG-NEXT: LSHR * T38.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
; EG-NEXT: LSHR T36.Y, T38.Z, literal.x,
; EG-NEXT: LSHR * T42.W, T41.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T36.X, T38.Z, literal.x,
; EG-NEXT: AND_INT T42.Z, T41.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: LSHR T43.X, PV.W, literal.x,
; EG-NEXT: LSHR T42.Y, T41.X, literal.y,
; EG-NEXT: LSHR T44.W, T41.W, literal.y,
; EG-NEXT: AND_INT * T42.X, T41.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T44.Z, T41.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 32(4.484155e-44)
; EG-NEXT: LSHR T41.X, PV.W, literal.x,
; EG-NEXT: LSHR T44.Y, T41.Z, literal.y,
; EG-NEXT: LSHR T45.W, T40.Y, literal.y,
; EG-NEXT: AND_INT * T44.X, T41.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T45.Z, T40.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 48(6.726233e-44)
; EG-NEXT: LSHR T46.X, PV.W, literal.x,
; EG-NEXT: LSHR T45.Y, T40.X, literal.y,
; EG-NEXT: LSHR T47.W, T40.W, literal.y,
; EG-NEXT: AND_INT * T45.X, T40.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T47.Z, T40.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 64(8.968310e-44)
; EG-NEXT: LSHR T40.X, PV.W, literal.x,
; EG-NEXT: LSHR T47.Y, T40.Z, literal.y,
; EG-NEXT: AND_INT * T47.X, T40.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: ADD_INT T0.W, KC0[2].Y, literal.x,
; EG-NEXT: LSHR * T37.W, T39.Y, literal.y,
; EG-NEXT: 80(1.121039e-43), 16(2.242078e-44)
; EG-NEXT: LSHR T48.X, PV.W, literal.x,
; EG-NEXT: AND_INT * T37.Z, T39.Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 65535(9.183409e-41)
; EG-NEXT: ALU clause starting at 95:
; EG-NEXT: LSHR T37.Y, T39.X, literal.x,
; EG-NEXT: LSHR * T53.W, T39.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T37.X, T39.X, literal.x,
; EG-NEXT: AND_INT T53.Z, T39.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 96(1.345247e-43)
; EG-NEXT: LSHR T39.X, PV.W, literal.x,
; EG-NEXT: LSHR T53.Y, T39.Z, literal.y,
; EG-NEXT: LSHR T54.W, T52.Y, literal.y,
; EG-NEXT: AND_INT * T53.X, T39.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T54.Z, T52.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 112(1.569454e-43)
; EG-NEXT: LSHR T55.X, PV.W, literal.x,
; EG-NEXT: LSHR T54.Y, T52.X, literal.y,
; EG-NEXT: LSHR T56.W, T52.W, literal.y,
; EG-NEXT: AND_INT * T54.X, T52.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T56.Z, T52.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 128(1.793662e-43)
; EG-NEXT: LSHR T52.X, PV.W, literal.x,
; EG-NEXT: LSHR T56.Y, T52.Z, literal.y,
; EG-NEXT: LSHR T57.W, T51.Y, literal.y,
; EG-NEXT: AND_INT * T56.X, T52.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T57.Z, T51.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 144(2.017870e-43)
; EG-NEXT: LSHR T58.X, PV.W, literal.x,
; EG-NEXT: LSHR T57.Y, T51.X, literal.y,
; EG-NEXT: LSHR T59.W, T51.W, literal.y,
; EG-NEXT: AND_INT * T57.X, T51.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T59.Z, T51.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 160(2.242078e-43)
; EG-NEXT: LSHR T51.X, PV.W, literal.x,
; EG-NEXT: LSHR T59.Y, T51.Z, literal.y,
; EG-NEXT: LSHR T60.W, T50.Y, literal.y,
; EG-NEXT: AND_INT * T59.X, T51.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T60.Z, T50.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 176(2.466285e-43)
; EG-NEXT: LSHR T61.X, PV.W, literal.x,
; EG-NEXT: LSHR T60.Y, T50.X, literal.y,
; EG-NEXT: LSHR T62.W, T50.W, literal.y,
; EG-NEXT: AND_INT * T60.X, T50.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T62.Z, T50.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 192(2.690493e-43)
; EG-NEXT: LSHR T50.X, PV.W, literal.x,
; EG-NEXT: LSHR T62.Y, T50.Z, literal.y,
; EG-NEXT: LSHR T63.W, T49.Y, literal.y,
; EG-NEXT: AND_INT * T62.X, T50.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T63.Z, T49.Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 208(2.914701e-43)
; EG-NEXT: LSHR T64.X, PV.W, literal.x,
; EG-NEXT: LSHR T63.Y, T49.X, literal.y,
; EG-NEXT: LSHR T65.W, T49.W, literal.y,
; EG-NEXT: AND_INT * T63.X, T49.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: AND_INT T65.Z, T49.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 224(3.138909e-43)
; EG-NEXT: LSHR T49.X, PV.W, literal.x,
; EG-NEXT: LSHR T65.Y, T49.Z, literal.y,
; EG-NEXT: AND_INT * T65.X, T49.Z, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 65535(9.183409e-41), 0(0.000000e+00)
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 240(3.363116e-43), 0(0.000000e+00)
; EG-NEXT: LSHR * T66.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v64i16_to_v64i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[36:39], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: s_load_b512 s[16:31], s[38:39], 0x0
; GFX12-NEXT: s_load_b512 s[0:15], s[38:39], 0x40
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s49, s31, 16
; GFX12-NEXT: s_lshr_b32 s65, s15, 16
; GFX12-NEXT: s_lshr_b32 s66, s14, 16
; GFX12-NEXT: s_and_b32 s14, s14, 0xffff
; GFX12-NEXT: s_and_b32 s15, s15, 0xffff
; GFX12-NEXT: s_lshr_b32 s63, s13, 16
; GFX12-NEXT: s_lshr_b32 s64, s12, 16
; GFX12-NEXT: s_and_b32 s13, s13, 0xffff
; GFX12-NEXT: s_and_b32 s12, s12, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s66
; GFX12-NEXT: s_lshr_b32 s61, s11, 16
; GFX12-NEXT: s_lshr_b32 s62, s10, 16
; GFX12-NEXT: s_and_b32 s11, s11, 0xffff
; GFX12-NEXT: s_and_b32 s10, s10, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s65
; GFX12-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s64
; GFX12-NEXT: s_lshr_b32 s59, s9, 16
; GFX12-NEXT: s_lshr_b32 s60, s8, 16
; GFX12-NEXT: s_and_b32 s9, s9, 0xffff
; GFX12-NEXT: s_and_b32 s8, s8, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s63
; GFX12-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s62
; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s61
; GFX12-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v13, s60
; GFX12-NEXT: s_lshr_b32 s57, s7, 16
; GFX12-NEXT: s_lshr_b32 s58, s6, 16
; GFX12-NEXT: s_and_b32 s7, s7, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v12, s8 :: v_dual_mov_b32 v15, s59
; GFX12-NEXT: v_mov_b32_e32 v14, s9
; GFX12-NEXT: s_and_b32 s6, s6, 0xffff
; GFX12-NEXT: s_lshr_b32 s55, s5, 16
; GFX12-NEXT: s_lshr_b32 s56, s4, 16
; GFX12-NEXT: s_and_b32 s5, s5, 0xffff
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: s_lshr_b32 s53, s3, 16
; GFX12-NEXT: s_lshr_b32 s54, s2, 16
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: s_and_b32 s2, s2, 0xffff
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:240
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:224
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:208
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:192
; GFX12-NEXT: v_dual_mov_b32 v1, s58 :: v_dual_mov_b32 v0, s6
; GFX12-NEXT: v_dual_mov_b32 v3, s57 :: v_dual_mov_b32 v2, s7
; GFX12-NEXT: v_mov_b32_e32 v5, s56
; GFX12-NEXT: s_lshr_b32 s51, s1, 16
; GFX12-NEXT: s_lshr_b32 s52, s0, 16
; GFX12-NEXT: s_and_b32 s1, s1, 0xffff
; GFX12-NEXT: s_and_b32 s0, s0, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s55
; GFX12-NEXT: v_dual_mov_b32 v6, s5 :: v_dual_mov_b32 v9, s54
; GFX12-NEXT: s_lshr_b32 s50, s30, 16
; GFX12-NEXT: s_and_b32 s31, s31, 0xffff
; GFX12-NEXT: s_and_b32 s30, s30, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v11, s53
; GFX12-NEXT: v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v13, s52
; GFX12-NEXT: s_lshr_b32 s45, s27, 16
; GFX12-NEXT: s_lshr_b32 s46, s26, 16
; GFX12-NEXT: s_lshr_b32 s47, s29, 16
; GFX12-NEXT: s_lshr_b32 s48, s28, 16
; GFX12-NEXT: s_and_b32 s27, s27, 0xffff
; GFX12-NEXT: s_and_b32 s26, s26, 0xffff
; GFX12-NEXT: s_and_b32 s29, s29, 0xffff
; GFX12-NEXT: s_and_b32 s28, s28, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v15, s51
; GFX12-NEXT: v_dual_mov_b32 v14, s1 :: v_dual_mov_b32 v17, s50
; GFX12-NEXT: s_lshr_b32 s43, s25, 16
; GFX12-NEXT: s_lshr_b32 s44, s24, 16
; GFX12-NEXT: s_and_b32 s25, s25, 0xffff
; GFX12-NEXT: s_and_b32 s24, s24, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v16, s30 :: v_dual_mov_b32 v19, s49
; GFX12-NEXT: v_dual_mov_b32 v18, s31 :: v_dual_mov_b32 v21, s48
; GFX12-NEXT: s_lshr_b32 s41, s23, 16
; GFX12-NEXT: s_lshr_b32 s42, s22, 16
; GFX12-NEXT: s_and_b32 s23, s23, 0xffff
; GFX12-NEXT: s_and_b32 s22, s22, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v20, s28 :: v_dual_mov_b32 v23, s47
; GFX12-NEXT: v_mov_b32_e32 v22, s29
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:176
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:160
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:144
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:128
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:112
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[36:37] offset:96
; GFX12-NEXT: v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v0, s26
; GFX12-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v2, s27
; GFX12-NEXT: v_mov_b32_e32 v5, s44
; GFX12-NEXT: s_lshr_b32 s39, s21, 16
; GFX12-NEXT: s_lshr_b32 s40, s20, 16
; GFX12-NEXT: s_and_b32 s21, s21, 0xffff
; GFX12-NEXT: s_and_b32 s20, s20, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v4, s24 :: v_dual_mov_b32 v7, s43
; GFX12-NEXT: v_dual_mov_b32 v6, s25 :: v_dual_mov_b32 v9, s42
; GFX12-NEXT: s_lshr_b32 s35, s19, 16
; GFX12-NEXT: s_lshr_b32 s38, s18, 16
; GFX12-NEXT: s_and_b32 s19, s19, 0xffff
; GFX12-NEXT: s_and_b32 s18, s18, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v11, s41
; GFX12-NEXT: v_dual_mov_b32 v10, s23 :: v_dual_mov_b32 v13, s40
; GFX12-NEXT: s_lshr_b32 s33, s17, 16
; GFX12-NEXT: s_lshr_b32 s34, s16, 16
; GFX12-NEXT: s_and_b32 s17, s17, 0xffff
; GFX12-NEXT: s_and_b32 s16, s16, 0xffff
; GFX12-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v15, s39
; GFX12-NEXT: v_dual_mov_b32 v14, s21 :: v_dual_mov_b32 v17, s38
; GFX12-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v19, s35
; GFX12-NEXT: v_dual_mov_b32 v18, s19 :: v_dual_mov_b32 v21, s34
; GFX12-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v23, s33
; GFX12-NEXT: v_mov_b32_e32 v22, s17
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:80
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:64
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:48
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:32
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[36:37]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <64 x i16>, ptr addrspace(4) %in
%ext = zext <64 x i16> %load to <64 x i32>
store <64 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v64i16_to_v64i32(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v64i16_to_v64i32:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[36:39], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x0
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x10
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s33, s17, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s34, s16, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s17, s17
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s16, s16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s35, s19, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s38, s18, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s19, s19
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s18, s18
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s39, s21, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s40, s20, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s21, s21
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s20, s20
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s41, s23, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s42, s22, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s23, s23
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s22, s22
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s43, s25, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s44, s24, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s25, s25
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s24, s24
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s45, s27, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s46, s26, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s27, s27
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s26, s26
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s47, s29, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s48, s28, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s29, s29
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s28, s28
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s49, s31, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s50, s30, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s31, s31
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s30, s30
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s51, s1, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s52, s0, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s53, s1
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s54, s0
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s55, s3, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s56, s2, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s57, s3
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s58, s2
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s59, s5, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s60, s4, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s61, s7, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s62, s6, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s63, s8, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s64, s9
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s65, s11, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s66, s10, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s67, s13, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s68, s12, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s13, s13
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s12, s12
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s69, s15, 16
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s70, s14, 16
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s15, s15
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s14, s14
; GCN-NOHSA-SI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-SI-NEXT: s_ashr_i32 s9, s9, 16
; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s36
; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s37
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s70
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s69
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s68
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s67
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s66
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s65
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, s63
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, s64
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:224
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:208
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, s9
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:192
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(3)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s62
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s61
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:176
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s60
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s59
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:160
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s58
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s56
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s57
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s55
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:144
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s54
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s52
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s53
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s51
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:128
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s30
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s50
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s49
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s48
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s47
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s26
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s46
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s27
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s45
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s24
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s44
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s25
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s43
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s42
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s23
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s41
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s20
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s40
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s39
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s38
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v64i16_to_v64i32:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s20, s1, 16
; GCN-HSA-NEXT: s_ashr_i32 s21, s0, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s22, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s23, s0
; GCN-HSA-NEXT: s_ashr_i32 s24, s3, 16
; GCN-HSA-NEXT: s_ashr_i32 s25, s2, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s26, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s27, s2
; GCN-HSA-NEXT: s_ashr_i32 s28, s5, 16
; GCN-HSA-NEXT: s_ashr_i32 s29, s4, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s30, s5
; GCN-HSA-NEXT: s_sext_i32_i16 s31, s4
; GCN-HSA-NEXT: s_ashr_i32 s33, s7, 16
; GCN-HSA-NEXT: s_ashr_i32 s34, s6, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s35, s7
; GCN-HSA-NEXT: s_sext_i32_i16 s36, s6
; GCN-HSA-NEXT: s_ashr_i32 s37, s9, 16
; GCN-HSA-NEXT: s_ashr_i32 s38, s8, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s39, s9
; GCN-HSA-NEXT: s_sext_i32_i16 s40, s8
; GCN-HSA-NEXT: s_ashr_i32 s41, s11, 16
; GCN-HSA-NEXT: s_ashr_i32 s42, s10, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s43, s11
; GCN-HSA-NEXT: s_sext_i32_i16 s44, s10
; GCN-HSA-NEXT: s_ashr_i32 s45, s13, 16
; GCN-HSA-NEXT: s_ashr_i32 s46, s12, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s47, s13
; GCN-HSA-NEXT: s_sext_i32_i16 s48, s12
; GCN-HSA-NEXT: s_ashr_i32 s49, s15, 16
; GCN-HSA-NEXT: s_ashr_i32 s50, s14, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s51, s15
; GCN-HSA-NEXT: s_sext_i32_i16 s52, s14
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x10
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_ashr_i32 s18, s1, 16
; GCN-HSA-NEXT: s_ashr_i32 s19, s0, 16
; GCN-HSA-NEXT: s_ashr_i32 s55, s3, 16
; GCN-HSA-NEXT: s_ashr_i32 s56, s2, 16
; GCN-HSA-NEXT: s_ashr_i32 s57, s5, 16
; GCN-HSA-NEXT: s_ashr_i32 s58, s4, 16
; GCN-HSA-NEXT: s_ashr_i32 s59, s7, 16
; GCN-HSA-NEXT: s_ashr_i32 s60, s6, 16
; GCN-HSA-NEXT: s_ashr_i32 s61, s9, 16
; GCN-HSA-NEXT: s_ashr_i32 s62, s8, 16
; GCN-HSA-NEXT: s_ashr_i32 s63, s11, 16
; GCN-HSA-NEXT: s_ashr_i32 s64, s10, 16
; GCN-HSA-NEXT: s_ashr_i32 s65, s13, 16
; GCN-HSA-NEXT: s_ashr_i32 s66, s12, 16
; GCN-HSA-NEXT: s_ashr_i32 s67, s15, 16
; GCN-HSA-NEXT: s_ashr_i32 s68, s14, 16
; GCN-HSA-NEXT: s_sext_i32_i16 s54, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xf0
; GCN-HSA-NEXT: s_sext_i32_i16 s53, s1
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xe0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v25, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v24, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xd0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v27, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v26, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xc0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v29, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v28, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xb0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v31, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v30, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0xa0
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v33, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s13, s13
; GCN-HSA-NEXT: s_sext_i32_i16 s12, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v32, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x90
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s66
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s65
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[24:25], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v25, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v24, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x80
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v35, s1
; GCN-HSA-NEXT: s_sext_i32_i16 s7, s7
; GCN-HSA-NEXT: s_sext_i32_i16 s6, s6
; GCN-HSA-NEXT: s_sext_i32_i16 s9, s9
; GCN-HSA-NEXT: s_sext_i32_i16 s8, s8
; GCN-HSA-NEXT: s_sext_i32_i16 s11, s11
; GCN-HSA-NEXT: s_sext_i32_i16 s10, s10
; GCN-HSA-NEXT: s_sext_i32_i16 s15, s15
; GCN-HSA-NEXT: s_sext_i32_i16 s14, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v34, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x70
; GCN-HSA-NEXT: s_sext_i32_i16 s3, s3
; GCN-HSA-NEXT: s_sext_i32_i16 s2, s2
; GCN-HSA-NEXT: s_sext_i32_i16 s5, s5
; GCN-HSA-NEXT: s_sext_i32_i16 s4, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s68
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s15
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s67
; GCN-HSA-NEXT: v_mov_b32_e32 v8, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v9, s64
; GCN-HSA-NEXT: v_mov_b32_e32 v10, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v11, s63
; GCN-HSA-NEXT: v_mov_b32_e32 v12, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v13, s62
; GCN-HSA-NEXT: v_mov_b32_e32 v14, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v15, s61
; GCN-HSA-NEXT: v_mov_b32_e32 v16, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v17, s60
; GCN-HSA-NEXT: v_mov_b32_e32 v18, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v19, s59
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s54
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s53
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s18
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v20, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v21, s58
; GCN-HSA-NEXT: flat_store_dwordx4 v[22:23], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s57
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s56
; GCN-HSA-NEXT: flat_store_dwordx4 v[26:27], v[8:11]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s3
; GCN-HSA-NEXT: flat_store_dwordx4 v[28:29], v[12:15]
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s55
; GCN-HSA-NEXT: flat_store_dwordx4 v[30:31], v[16:19]
; GCN-HSA-NEXT: flat_store_dwordx4 v[32:33], v[20:23]
; GCN-HSA-NEXT: flat_store_dwordx4 v[24:25], v[0:3]
; GCN-HSA-NEXT: flat_store_dwordx4 v[34:35], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s52
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s50
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s51
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s49
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s48
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s46
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s47
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s45
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s44
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s42
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s43
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s41
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s40
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s38
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s39
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s37
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s36
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s33
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: s_add_u32 s0, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s30
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s28
; GCN-HSA-NEXT: s_addc_u32 s1, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s26
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s24
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s22
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v64i16_to_v64i32:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[36:39], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[16:31], s[38:39], 0x0
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[38:39], 0x40
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s33, s17, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s34, s16, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s35, s19, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s38, s18, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s39, s21, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s40, s20, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s41, s23, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s42, s22, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s43, s25, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s44, s24, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s45, s27, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s46, s26, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s47, s29, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s48, s28, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s49, s31, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s50, s30, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s51, s1, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s52, s0, 16
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s53, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s54, s0
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s55, s3, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s56, s2, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s57, s5, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s58, s4, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s59, s7, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s60, s6, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s61, s9, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s62, s8, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s63, s11, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s64, s10, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s65, s13, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s66, s12, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s0, s15, 16
; GCN-NOHSA-VI-NEXT: s_ashr_i32 s1, s14, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xf0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s15, s15
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s14, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xe0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s13, s13
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s12, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xd0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s66
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s65
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s11, s11
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s10, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xc0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s63
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s9, s9
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s8, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xb0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s62
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s61
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s7, s7
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s6, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0xa0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s59
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s5, s5
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0x90
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s58
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s57
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s3, s3
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0x80
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s56
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s55
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0x70
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s54
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s52
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s53
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s51
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s31, s31
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s30, s30
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0x60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s30
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s49
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s29, s29
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s28, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s47
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s27, s27
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s26, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s46
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s45
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s25, s25
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s24, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s44
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s43
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s23, s23
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s22, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s42
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s41
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s21, s21
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s20, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_add_u32 s0, s36, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s40
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s39
; GCN-NOHSA-VI-NEXT: s_addc_u32 s1, s37, 0
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s19, s19
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s18, s18
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s38
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s17, s17
; GCN-NOHSA-VI-NEXT: s_sext_i32_i16 s16, s16
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s36
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s33
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s37
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v64i16_to_v64i32:
; EG: ; %bb.0:
; EG-NEXT: ALU 17, @38, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 7 @22
; EG-NEXT: ALU 75, @56, KC0[CB0:0-32], KC1[]
; EG-NEXT: ALU 71, @132, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T66.XYZW, T48.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T41.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T65.XYZW, T56.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T46.XYZW, T55.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T64.XYZW, T54.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T53.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T63.XYZW, T52.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T43.XYZW, T51.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T62.XYZW, T50.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T49.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T61.XYZW, T40.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T39.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T60.XYZW, T38.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T59.XYZW, T37.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T58.XYZW, T36.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T57.XYZW, T35.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 22:
; EG-NEXT: VTX_READ_128 T42.XYZW, T41.X, 16, #1
; EG-NEXT: VTX_READ_128 T43.XYZW, T41.X, 32, #1
; EG-NEXT: VTX_READ_128 T44.XYZW, T41.X, 0, #1
; EG-NEXT: VTX_READ_128 T45.XYZW, T41.X, 48, #1
; EG-NEXT: VTX_READ_128 T46.XYZW, T41.X, 64, #1
; EG-NEXT: VTX_READ_128 T47.XYZW, T41.X, 80, #1
; EG-NEXT: VTX_READ_128 T48.XYZW, T41.X, 96, #1
; EG-NEXT: VTX_READ_128 T41.XYZW, T41.X, 112, #1
; EG-NEXT: ALU clause starting at 38:
; EG-NEXT: LSHR T35.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T36.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T37.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR T38.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T39.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43)
; EG-NEXT: LSHR T40.X, PV.W, literal.x,
; EG-NEXT: MOV * T41.X, KC0[2].Z,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 56:
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 96(1.345247e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T49.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43)
; EG-NEXT: LSHR T50.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43)
; EG-NEXT: LSHR T51.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43)
; EG-NEXT: LSHR T52.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43)
; EG-NEXT: LSHR T53.X, PV.W, literal.x,
; EG-NEXT: LSHR T0.Y, T41.W, literal.y,
; EG-NEXT: LSHR T0.Z, T41.Y, literal.y,
; EG-NEXT: LSHR T0.W, T48.W, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: ADD_INT * T1.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 176(2.466285e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T54.X, PS, literal.x,
; EG-NEXT: LSHR T1.Y, T48.Y, literal.y,
; EG-NEXT: LSHR T1.Z, T47.W, literal.y,
; EG-NEXT: LSHR T1.W, T47.Y, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: ADD_INT * T2.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 192(2.690493e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T55.X, PS, literal.x,
; EG-NEXT: LSHR T2.Y, T46.W, literal.y,
; EG-NEXT: LSHR T2.Z, T46.Y, literal.y,
; EG-NEXT: LSHR T2.W, T45.W, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: ADD_INT * T3.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 208(2.914701e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T56.X, PS, literal.x,
; EG-NEXT: LSHR T3.Y, T45.Y, literal.y,
; EG-NEXT: BFE_INT T57.Z, T44.Y, 0.0, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: LSHR T3.W, T43.W, literal.y,
; EG-NEXT: LSHR * T4.W, T43.Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: BFE_INT T57.X, T44.X, 0.0, literal.x,
; EG-NEXT: LSHR T4.Y, T42.W, literal.x,
; EG-NEXT: BFE_INT T58.Z, T44.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR T5.W, T42.Y, literal.x,
; EG-NEXT: LSHR * T6.W, T44.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T58.X, T44.Z, 0.0, literal.x,
; EG-NEXT: LSHR T5.Y, T44.W, literal.x,
; EG-NEXT: BFE_INT T59.Z, T42.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T57.W, PS, 0.0, literal.x,
; EG-NEXT: LSHR * T6.W, T44.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T59.X, T42.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T57.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T60.Z, T42.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T58.W, PV.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T6.W, T44.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T60.X, T42.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T58.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T44.Z, T43.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T59.W, T5.W, 0.0, literal.x,
; EG-NEXT: LSHR * T5.W, T42.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T44.X, T43.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T59.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T61.Z, T43.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T60.W, T4.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T5.W, T42.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T61.X, T43.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T60.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T42.Z, T45.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT * T44.W, T4.W, 0.0, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 132:
; EG-NEXT: LSHR * T4.W, T43.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T42.X, T45.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T44.Y, PV.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T62.Z, T45.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T61.W, T3.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T3.W, T43.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T62.X, T45.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T61.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T43.Z, T46.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T42.W, T3.Y, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T3.W, T45.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T43.X, T46.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T42.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T63.Z, T46.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T62.W, T2.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T2.W, T45.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T63.X, T46.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T62.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T45.Z, T47.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T43.W, T2.Z, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T2.W, T46.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T45.X, T47.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T43.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T64.Z, T47.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T63.W, T2.Y, 0.0, literal.x,
; EG-NEXT: LSHR * T2.W, T46.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T64.X, T47.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T63.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T46.Z, T48.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T45.W, T1.W, 0.0, literal.x,
; EG-NEXT: LSHR * T1.W, T47.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T46.X, T48.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T45.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T65.Z, T48.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T64.W, T1.Z, 0.0, literal.x,
; EG-NEXT: LSHR * T1.W, T47.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T65.X, T48.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T64.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T47.Z, T41.Y, 0.0, literal.x,
; EG-NEXT: BFE_INT T46.W, T1.Y, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T1.W, T48.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T47.X, T41.X, 0.0, literal.x,
; EG-NEXT: BFE_INT T46.Y, PS, 0.0, literal.x,
; EG-NEXT: BFE_INT T66.Z, T41.W, 0.0, literal.x,
; EG-NEXT: BFE_INT T65.W, T0.W, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: LSHR * T0.W, T48.Z, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T66.X, T41.Z, 0.0, literal.x,
; EG-NEXT: BFE_INT T65.Y, PS, 0.0, literal.x,
; EG-NEXT: LSHR T1.Z, T41.X, literal.x,
; EG-NEXT: BFE_INT T47.W, T0.Z, 0.0, literal.x, BS:VEC_120/SCL_212
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 224(3.138909e-43)
; EG-NEXT: LSHR T41.X, PS, literal.x,
; EG-NEXT: BFE_INT T47.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: LSHR T0.Z, T41.Z, literal.y,
; EG-NEXT: BFE_INT T66.W, T0.Y, 0.0, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 240(3.363116e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T48.X, PS, literal.x,
; EG-NEXT: BFE_INT * T66.Y, PV.Z, 0.0, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
;
; GFX12-LABEL: constant_sextload_v64i16_to_v64i32:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[36:39], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: s_load_b512 s[0:15], s[38:39], 0x40
; GFX12-NEXT: s_load_b512 s[16:31], s[38:39], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_ashr_i32 s65, s15, 16
; GFX12-NEXT: s_ashr_i32 s66, s14, 16
; GFX12-NEXT: s_sext_i32_i16 s14, s14
; GFX12-NEXT: s_sext_i32_i16 s15, s15
; GFX12-NEXT: s_ashr_i32 s63, s13, 16
; GFX12-NEXT: s_ashr_i32 s64, s12, 16
; GFX12-NEXT: s_sext_i32_i16 s13, s13
; GFX12-NEXT: s_sext_i32_i16 s12, s12
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s66
; GFX12-NEXT: s_ashr_i32 s61, s11, 16
; GFX12-NEXT: s_ashr_i32 s62, s10, 16
; GFX12-NEXT: s_sext_i32_i16 s11, s11
; GFX12-NEXT: s_sext_i32_i16 s10, s10
; GFX12-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s65
; GFX12-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v5, s64
; GFX12-NEXT: s_ashr_i32 s59, s9, 16
; GFX12-NEXT: s_ashr_i32 s60, s8, 16
; GFX12-NEXT: s_sext_i32_i16 s9, s9
; GFX12-NEXT: s_sext_i32_i16 s8, s8
; GFX12-NEXT: v_dual_mov_b32 v4, s12 :: v_dual_mov_b32 v7, s63
; GFX12-NEXT: v_dual_mov_b32 v6, s13 :: v_dual_mov_b32 v9, s62
; GFX12-NEXT: v_dual_mov_b32 v8, s10 :: v_dual_mov_b32 v11, s61
; GFX12-NEXT: v_dual_mov_b32 v10, s11 :: v_dual_mov_b32 v13, s60
; GFX12-NEXT: s_ashr_i32 s57, s7, 16
; GFX12-NEXT: s_ashr_i32 s58, s6, 16
; GFX12-NEXT: s_sext_i32_i16 s7, s7
; GFX12-NEXT: v_dual_mov_b32 v12, s8 :: v_dual_mov_b32 v15, s59
; GFX12-NEXT: v_mov_b32_e32 v14, s9
; GFX12-NEXT: s_sext_i32_i16 s6, s6
; GFX12-NEXT: s_ashr_i32 s55, s5, 16
; GFX12-NEXT: s_ashr_i32 s56, s4, 16
; GFX12-NEXT: s_sext_i32_i16 s5, s5
; GFX12-NEXT: s_sext_i32_i16 s4, s4
; GFX12-NEXT: s_ashr_i32 s53, s3, 16
; GFX12-NEXT: s_ashr_i32 s54, s2, 16
; GFX12-NEXT: s_sext_i32_i16 s3, s3
; GFX12-NEXT: s_sext_i32_i16 s2, s2
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:240
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:224
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:208
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:192
; GFX12-NEXT: v_dual_mov_b32 v1, s58 :: v_dual_mov_b32 v0, s6
; GFX12-NEXT: v_dual_mov_b32 v3, s57 :: v_dual_mov_b32 v2, s7
; GFX12-NEXT: v_mov_b32_e32 v5, s56
; GFX12-NEXT: s_ashr_i32 s51, s1, 16
; GFX12-NEXT: s_ashr_i32 s52, s0, 16
; GFX12-NEXT: s_sext_i32_i16 s1, s1
; GFX12-NEXT: s_sext_i32_i16 s0, s0
; GFX12-NEXT: v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v7, s55
; GFX12-NEXT: v_dual_mov_b32 v6, s5 :: v_dual_mov_b32 v9, s54
; GFX12-NEXT: s_ashr_i32 s49, s31, 16
; GFX12-NEXT: s_ashr_i32 s50, s30, 16
; GFX12-NEXT: s_sext_i32_i16 s31, s31
; GFX12-NEXT: s_sext_i32_i16 s30, s30
; GFX12-NEXT: v_dual_mov_b32 v8, s2 :: v_dual_mov_b32 v11, s53
; GFX12-NEXT: v_dual_mov_b32 v10, s3 :: v_dual_mov_b32 v13, s52
; GFX12-NEXT: s_ashr_i32 s45, s27, 16
; GFX12-NEXT: s_ashr_i32 s46, s26, 16
; GFX12-NEXT: s_sext_i32_i16 s27, s27
; GFX12-NEXT: s_sext_i32_i16 s26, s26
; GFX12-NEXT: s_ashr_i32 s47, s29, 16
; GFX12-NEXT: s_ashr_i32 s48, s28, 16
; GFX12-NEXT: s_sext_i32_i16 s29, s29
; GFX12-NEXT: s_sext_i32_i16 s28, s28
; GFX12-NEXT: v_dual_mov_b32 v12, s0 :: v_dual_mov_b32 v15, s51
; GFX12-NEXT: v_dual_mov_b32 v14, s1 :: v_dual_mov_b32 v17, s50
; GFX12-NEXT: s_ashr_i32 s43, s25, 16
; GFX12-NEXT: s_ashr_i32 s44, s24, 16
; GFX12-NEXT: s_sext_i32_i16 s25, s25
; GFX12-NEXT: s_sext_i32_i16 s24, s24
; GFX12-NEXT: v_dual_mov_b32 v16, s30 :: v_dual_mov_b32 v19, s49
; GFX12-NEXT: v_dual_mov_b32 v18, s31 :: v_dual_mov_b32 v21, s48
; GFX12-NEXT: s_ashr_i32 s41, s23, 16
; GFX12-NEXT: s_ashr_i32 s42, s22, 16
; GFX12-NEXT: s_sext_i32_i16 s23, s23
; GFX12-NEXT: s_sext_i32_i16 s22, s22
; GFX12-NEXT: v_dual_mov_b32 v20, s28 :: v_dual_mov_b32 v23, s47
; GFX12-NEXT: v_mov_b32_e32 v22, s29
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:176
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:160
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:144
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:128
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:112
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[36:37] offset:96
; GFX12-NEXT: v_dual_mov_b32 v1, s46 :: v_dual_mov_b32 v0, s26
; GFX12-NEXT: v_dual_mov_b32 v3, s45 :: v_dual_mov_b32 v2, s27
; GFX12-NEXT: v_mov_b32_e32 v5, s44
; GFX12-NEXT: s_ashr_i32 s39, s21, 16
; GFX12-NEXT: s_ashr_i32 s40, s20, 16
; GFX12-NEXT: s_sext_i32_i16 s21, s21
; GFX12-NEXT: s_sext_i32_i16 s20, s20
; GFX12-NEXT: v_dual_mov_b32 v4, s24 :: v_dual_mov_b32 v7, s43
; GFX12-NEXT: v_dual_mov_b32 v6, s25 :: v_dual_mov_b32 v9, s42
; GFX12-NEXT: s_ashr_i32 s35, s19, 16
; GFX12-NEXT: s_ashr_i32 s38, s18, 16
; GFX12-NEXT: s_sext_i32_i16 s19, s19
; GFX12-NEXT: s_sext_i32_i16 s18, s18
; GFX12-NEXT: v_dual_mov_b32 v8, s22 :: v_dual_mov_b32 v11, s41
; GFX12-NEXT: v_dual_mov_b32 v10, s23 :: v_dual_mov_b32 v13, s40
; GFX12-NEXT: s_ashr_i32 s33, s17, 16
; GFX12-NEXT: s_ashr_i32 s34, s16, 16
; GFX12-NEXT: s_sext_i32_i16 s17, s17
; GFX12-NEXT: s_sext_i32_i16 s16, s16
; GFX12-NEXT: v_dual_mov_b32 v12, s20 :: v_dual_mov_b32 v15, s39
; GFX12-NEXT: v_dual_mov_b32 v14, s21 :: v_dual_mov_b32 v17, s38
; GFX12-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v19, s35
; GFX12-NEXT: v_dual_mov_b32 v18, s19 :: v_dual_mov_b32 v21, s34
; GFX12-NEXT: v_dual_mov_b32 v20, s16 :: v_dual_mov_b32 v23, s33
; GFX12-NEXT: v_mov_b32_e32 v22, s17
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[36:37] offset:80
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[36:37] offset:64
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[36:37] offset:48
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[36:37] offset:32
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[36:37] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[36:37]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <64 x i16>, ptr addrspace(4) %in
%ext = sext <64 x i16> %load to <64 x i32>
store <64 x i32> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_i16_to_i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_i16_to_i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_i16_to_i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_ushort v0, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_i16_to_i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, 0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_i16_to_i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: MOV * T0.Y, 0.0,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_i16_to_i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v1, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: global_load_u16 v0, v1, s[2:3]
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX12-NEXT: global_store_b64 v1, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%a = load i16, ptr addrspace(4) %in
%ext = zext i16 %a to i64
store i64 %ext, ptr addrspace(1) %out
ret void
}
; FIXME: Need to optimize this sequence to avoid extra bfe:
; t28: i32,ch = load<LD2[%in(addrspace=1)], anyext from i16> t12, t27, undef:i64
; t31: i64 = any_extend t28
; t33: i64 = sign_extend_inreg t31, ValueType:ch:i16
; TODO: These could be expanded earlier using ASHR 15
define amdgpu_kernel void @constant_sextload_i16_to_i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_i16_to_i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_sshort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_i16_to_i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_sshort v0, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_i16_to_i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v2, 0, 16
; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_i16_to_i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
; EG-NEXT: ASHR * T0.Y, PV.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
;
; GFX12-LABEL: constant_sextload_i16_to_i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v2, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: global_load_u16 v0, v2, s[2:3]
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%a = load i16, ptr addrspace(4) %in
%ext = sext i16 %a to i64
store i64 %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v1i16_to_v1i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v1i16_to_v1i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_ushort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v1i16_to_v1i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_ushort v0, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v1i16_to_v1i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, 0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: v_and_b32_e32 v2, 0xffff, v2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v1i16_to_v1i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 2, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: MOV * T0.Y, 0.0,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v1i16_to_v1i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v1, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: global_load_u16 v0, v1, s[2:3]
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_and_b32_e32 v0, 0xffff, v0
; GFX12-NEXT: global_store_b64 v1, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <1 x i16>, ptr addrspace(4) %in
%ext = zext <1 x i16> %load to <1 x i64>
store <1 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v1i16_to_v1i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v1i16_to_v1i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s7, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, -1
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s6
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, s7
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s2
; GCN-NOHSA-SI-NEXT: s_mov_b32 s9, s3
; GCN-NOHSA-SI-NEXT: buffer_load_sshort v0, off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s4, s0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s5, s1
; GCN-NOHSA-SI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-SI-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v1i16_to_v1i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: flat_load_sshort v0, v[0:1]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: s_waitcnt vmcnt(0)
; GCN-HSA-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GCN-HSA-NEXT: flat_store_dwordx2 v[2:3], v[0:1]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v1i16_to_v1i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: flat_load_ushort v2, v[0:1]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt vmcnt(0)
; GCN-NOHSA-VI-NEXT: v_bfe_i32 v2, v2, 0, 16
; GCN-NOHSA-VI-NEXT: v_ashrrev_i32_e32 v3, 31, v2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx2 v[0:1], v[2:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v1i16_to_v1i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 4, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T0.XY, T1.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_16 T0.X, T0.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T0.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: BFE_INT T0.X, T0.X, 0.0, literal.x,
; EG-NEXT: LSHR * T1.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
; EG-NEXT: ASHR * T0.Y, PV.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
;
; GFX12-LABEL: constant_sextload_v1i16_to_v1i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: v_mov_b32_e32 v2, 0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: global_load_u16 v0, v2, s[2:3]
; GFX12-NEXT: s_wait_loadcnt 0x0
; GFX12-NEXT: v_bfe_i32 v0, v0, 0, 16
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT: v_ashrrev_i32_e32 v1, 31, v0
; GFX12-NEXT: global_store_b64 v2, v[0:1], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <1 x i16>, ptr addrspace(4) %in
%ext = sext <1 x i16> %load to <1 x i64>
store <1 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v2i16_to_v2i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v2i16_to_v2i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s4, s2, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v2i16_to_v2i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s0, s2, 16
; GCN-HSA-NEXT: s_and_b32 s1, s2, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v2i16_to_v2i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s2, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v2i16_to_v2i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 6, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T5.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T4.X, T4.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T4.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T4.Z, T4.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T4.X, T4.X, literal.x,
; EG-NEXT: MOV T4.Y, 0.0,
; EG-NEXT: MOV T4.W, 0.0,
; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y,
; EG-NEXT: 65535(9.183409e-41), 2(2.802597e-45)
;
; GFX12-LABEL: constant_zextload_v2i16_to_v2i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s3, 0xffff, s2
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s3
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s2, 0
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <2 x i16>, ptr addrspace(4) %in
%ext = zext <2 x i16> %load to <2 x i64>
store <2 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v2i16_to_v2i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v2i16_to_v2i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s4, s2, 16
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[6:7], s[2:3], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v2i16_to_v2i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s0, s2, 16
; GCN-HSA-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[0:1], s[0:1], 0x100000
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v2i16_to_v2i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dword s2, s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[0:1], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s2, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v2i16_to_v2i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 8, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T4.XYZW, T5.X, 1
; EG-NEXT: CF_END
; EG-NEXT: PAD
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_32 T4.X, T4.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T4.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: ASHR * T4.W, T4.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: ASHR * T4.Z, T4.X, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T4.X, T4.X, 0.0, literal.x,
; EG-NEXT: LSHR * T5.X, KC0[2].Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 2(2.802597e-45)
; EG-NEXT: ASHR * T4.Y, PV.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
;
; GFX12-LABEL: constant_sextload_v2i16_to_v2i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b32 s2, s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s4, s2, 16
; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v4, 0 :: v_dual_mov_b32 v1, s3
; GFX12-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v3, s5
; GFX12-NEXT: v_mov_b32_e32 v2, s4
; GFX12-NEXT: global_store_b128 v4, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <2 x i16>, ptr addrspace(4) %in
%ext = sext <2 x i16> %load to <2 x i64>
store <2 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v4i16_to_v4i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v4i16_to_v4i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s6, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s7, s4, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v4i16_to_v4i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s4, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s5, s2, 16
; GCN-HSA-NEXT: s_and_b32 s6, s2, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s3, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v4i16_to_v4i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s2, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s3, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v4i16_to_v4i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 14, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T6.XYZW, T8.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T7.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T5.XY, T5.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T5.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: LSHR * T6.Z, T5.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T6.X, T5.Y, literal.x,
; EG-NEXT: MOV T6.Y, 0.0,
; EG-NEXT: LSHR T5.Z, T5.X, literal.y,
; EG-NEXT: AND_INT * T5.X, T5.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T5.Y, 0.0,
; EG-NEXT: MOV T6.W, 0.0,
; EG-NEXT: MOV * T5.W, 0.0,
; EG-NEXT: LSHR T7.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR * T8.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v4i16_to_v4i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s4, 0xffff, s2
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s4
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s2, 0
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, v1
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s3, 0
; GFX12-NEXT: s_and_b32 s3, 0xffff, s3
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1]
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <4 x i16>, ptr addrspace(4) %in
%ext = zext <4 x i16> %load to <4 x i64>
store <4 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v4i16_to_v4i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v4i16_to_v4i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx2 s[4:5], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s6, s5
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s8, s4, 16
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[10:11], s[4:5], 0x100000
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[4:5], s[4:5], 48
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v4i16_to_v4i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_mov_b32 s4, s3
; GCN-HSA-NEXT: s_lshr_b32 s6, s2, 16
; GCN-HSA-NEXT: s_bfe_i64 s[8:9], s[2:3], 0x100000
; GCN-HSA-NEXT: s_ashr_i64 s[2:3], s[2:3], 48
; GCN-HSA-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v4i16_to_v4i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx2 s[2:3], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[4:5], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s2, 16
; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, s3
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s8, s3, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s7
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v4i16_to_v4i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @8, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @6
; EG-NEXT: ALU 16, @9, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T8.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T5.XYZW, T6.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 6:
; EG-NEXT: VTX_READ_64 T5.XY, T5.X, 0, #1
; EG-NEXT: ALU clause starting at 8:
; EG-NEXT: MOV * T5.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 9:
; EG-NEXT: ASHR * T5.W, T5.X, literal.x,
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T6.X, KC0[2].Y, literal.x,
; EG-NEXT: ASHR T5.Z, T5.X, literal.y,
; EG-NEXT: ASHR * T7.W, T5.Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T5.X, T5.X, 0.0, literal.x,
; EG-NEXT: ASHR * T7.Z, T5.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T7.X, T5.Y, 0.0, literal.x,
; EG-NEXT: ASHR T5.Y, PV.X, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: LSHR T8.X, PV.W, literal.x,
; EG-NEXT: ASHR * T7.Y, PV.X, literal.y,
; EG-NEXT: 2(2.802597e-45), 31(4.344025e-44)
;
; GFX12-LABEL: constant_sextload_v4i16_to_v4i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b64 s[2:3], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s6, s3
; GFX12-NEXT: s_lshr_b32 s8, s3, 16
; GFX12-NEXT: s_bfe_i64 s[4:5], s[2:3], 0x100000
; GFX12-NEXT: s_lshr_b32 s2, s2, 16
; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v8, 0 :: v_dual_mov_b32 v1, s5
; GFX12-NEXT: v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v5, s7
; GFX12-NEXT: v_dual_mov_b32 v4, s6 :: v_dual_mov_b32 v7, s9
; GFX12-NEXT: v_dual_mov_b32 v6, s8 :: v_dual_mov_b32 v3, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v8, v[4:7], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v8, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <4 x i16>, ptr addrspace(4) %in
%ext = sext <4 x i16> %load to <4 x i64>
store <4 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v8i16_to_v8i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v8i16_to_v8i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s8, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s9, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s10, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s11, s4, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s11
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v8i16_to_v8i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s8, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s2, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s9, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s10, s4, 16
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s3, s7, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s8
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v8i16_to_v8i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s8, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s9, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s10, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s7, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 32
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v8i16_to_v8i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @8
; EG-NEXT: ALU 30, @11, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T8.XYZW, T14.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T9.XYZW, T13.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T12.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T7.XYZW, T11.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 8:
; EG-NEXT: VTX_READ_128 T7.XYZW, T7.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T7.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 11:
; EG-NEXT: LSHR * T8.Z, T7.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T8.X, T7.W, literal.x,
; EG-NEXT: MOV T8.Y, 0.0,
; EG-NEXT: LSHR T9.Z, T7.Z, literal.y,
; EG-NEXT: AND_INT * T9.X, T7.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T9.Y, 0.0,
; EG-NEXT: LSHR * T10.Z, T7.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T10.X, T7.Y, literal.x,
; EG-NEXT: MOV T10.Y, 0.0,
; EG-NEXT: LSHR T7.Z, T7.X, literal.y,
; EG-NEXT: AND_INT * T7.X, T7.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T7.Y, 0.0,
; EG-NEXT: MOV T8.W, 0.0,
; EG-NEXT: MOV * T9.W, 0.0,
; EG-NEXT: MOV T10.W, 0.0,
; EG-NEXT: MOV * T7.W, 0.0,
; EG-NEXT: LSHR T11.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T12.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T13.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR * T14.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v8i16_to_v8i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b128 s[4:7], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s2, 0xffff, s7
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
; GFX12-NEXT: s_pack_hl_b32_b16 s3, s7, 0
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s6, 0
; GFX12-NEXT: v_dual_mov_b32 v2, s3 :: v_dual_mov_b32 v3, v1
; GFX12-NEXT: s_and_b32 s3, 0xffff, s6
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:48
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s5, 0
; GFX12-NEXT: s_and_b32 s3, 0xffff, s5
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:32
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_pack_hl_b32_b16 s2, s4, 0
; GFX12-NEXT: s_and_b32 s3, 0xffff, s4
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1] offset:16
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <8 x i16>, ptr addrspace(4) %in
%ext = zext <8 x i16> %load to <8 x i64>
store <8 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v8i16_to_v8i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v8i16_to_v8i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s8, s7
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, s5
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s12, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s14, s4, 16
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[16:17], s[4:5], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[18:19], s[6:7], 0x100000
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[4:5], s[4:5], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[6:7], s[6:7], 48
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s16
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s13
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s15
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v8i16_to_v8i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_mov_b32 s2, s7
; GCN-HSA-NEXT: s_mov_b32 s8, s5
; GCN-HSA-NEXT: s_lshr_b32 s10, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s12, s4, 16
; GCN-HSA-NEXT: s_bfe_i64 s[14:15], s[4:5], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[16:17], s[6:7], 0x100000
; GCN-HSA-NEXT: s_ashr_i64 s[4:5], s[4:5], 48
; GCN-HSA-NEXT: s_ashr_i64 s[6:7], s[6:7], 48
; GCN-HSA-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s7
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s5
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s17
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s15
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v8i16_to_v8i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[4:7], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[12:13], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s6, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[14:15], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, s7
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[2:3], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s4, 16
; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s5
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s10, s5, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[16:17], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s7, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s0, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s0, 32
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s15
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s0, 16
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v8i16_to_v8i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @10, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @8
; EG-NEXT: ALU 33, @11, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T7.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T11.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T9.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T10.XYZW, T8.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 8:
; EG-NEXT: VTX_READ_128 T7.XYZW, T7.X, 0, #1
; EG-NEXT: ALU clause starting at 10:
; EG-NEXT: MOV * T7.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 11:
; EG-NEXT: LSHR T8.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T9.X, PV.W, literal.x,
; EG-NEXT: ADD_INT T0.W, KC0[2].Y, literal.y,
; EG-NEXT: ASHR * T10.W, T7.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T11.X, PV.W, literal.x,
; EG-NEXT: ASHR T10.Z, T7.X, literal.y,
; EG-NEXT: ASHR * T12.W, T7.Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T10.X, T7.X, 0.0, literal.x,
; EG-NEXT: ASHR T12.Z, T7.Y, literal.x,
; EG-NEXT: ASHR * T13.W, T7.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T12.X, T7.Y, 0.0, literal.x,
; EG-NEXT: ASHR T10.Y, PV.X, literal.y,
; EG-NEXT: ASHR T13.Z, T7.Z, literal.x,
; EG-NEXT: ASHR * T14.W, T7.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T13.X, T7.Z, 0.0, literal.x,
; EG-NEXT: ASHR T12.Y, PV.X, literal.y,
; EG-NEXT: ASHR * T14.Z, T7.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T14.X, T7.W, 0.0, literal.x,
; EG-NEXT: ASHR T13.Y, PV.X, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: 48(6.726233e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T7.X, PV.W, literal.x,
; EG-NEXT: ASHR * T14.Y, PV.X, literal.y,
; EG-NEXT: 2(2.802597e-45), 31(4.344025e-44)
;
; GFX12-LABEL: constant_sextload_v8i16_to_v8i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[0:3], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b128 s[4:7], s[2:3], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s14, s7
; GFX12-NEXT: s_lshr_b32 s16, s7, 16
; GFX12-NEXT: s_bfe_i64 s[12:13], s[6:7], 0x100000
; GFX12-NEXT: s_lshr_b32 s6, s6, 16
; GFX12-NEXT: s_bfe_i64 s[2:3], s[4:5], 0x100000
; GFX12-NEXT: s_mov_b32 s8, s5
; GFX12-NEXT: s_lshr_b32 s10, s5, 16
; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GFX12-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x100000
; GFX12-NEXT: s_lshr_b32 s4, s4, 16
; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v16, 0 :: v_dual_mov_b32 v1, s13
; GFX12-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GFX12-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v9, s15
; GFX12-NEXT: v_dual_mov_b32 v8, s14 :: v_dual_mov_b32 v11, s17
; GFX12-NEXT: v_dual_mov_b32 v10, s16 :: v_dual_mov_b32 v3, s7
; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v5, s3
; GFX12-NEXT: v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v13, s9
; GFX12-NEXT: v_dual_mov_b32 v12, s8 :: v_dual_mov_b32 v15, s11
; GFX12-NEXT: v_dual_mov_b32 v14, s10 :: v_dual_mov_b32 v7, s5
; GFX12-NEXT: v_mov_b32_e32 v6, s4
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v16, v[8:11], s[0:1] offset:48
; GFX12-NEXT: global_store_b128 v16, v[0:3], s[0:1] offset:32
; GFX12-NEXT: global_store_b128 v16, v[12:15], s[0:1] offset:16
; GFX12-NEXT: global_store_b128 v16, v[4:7], s[0:1]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <8 x i16>, ptr addrspace(4) %in
%ext = sext <8 x i16> %load to <8 x i64>
store <8 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v16i16_to_v16i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v16i16_to_v16i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s11, 0xf000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s10, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s12, s1, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s13, s3, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s14, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s15, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s16, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s17, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s18, s2, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s19, s0, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s14
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s12
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s16
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s18
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[8:11], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v16i16_to_v16i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s12, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s13, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s14, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s2, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s15, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s16, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s17, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s18, s4, 16
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-HSA-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s3, s9, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x70
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s14
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s13
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s12
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s15
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s16
; GCN-HSA-NEXT: s_addc_u32 s3, s1, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s17
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v16i16_to_v16i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s12, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s13, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s14, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s15, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s16, s10, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s17, s11, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s18, s8, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s9, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s9, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 64
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 0x70
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s18
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 0x60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s17
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s16
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s15
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s0, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s14
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s1, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v16i16_to_v16i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @16, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @12
; EG-NEXT: ALU 62, @17, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T13.XYZW, T26.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T14.XYZW, T25.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T15.XYZW, T24.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T12.XYZW, T23.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T16.XYZW, T22.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T17.XYZW, T21.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T18.XYZW, T20.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T19.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 12:
; EG-NEXT: VTX_READ_128 T12.XYZW, T11.X, 16, #1
; EG-NEXT: VTX_READ_128 T11.XYZW, T11.X, 0, #1
; EG-NEXT: ALU clause starting at 16:
; EG-NEXT: MOV * T11.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 17:
; EG-NEXT: LSHR * T13.Z, T12.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T13.X, T12.W, literal.x,
; EG-NEXT: MOV T13.Y, 0.0,
; EG-NEXT: LSHR T14.Z, T12.Z, literal.y,
; EG-NEXT: AND_INT * T14.X, T12.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T14.Y, 0.0,
; EG-NEXT: LSHR * T15.Z, T12.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T15.X, T12.Y, literal.x,
; EG-NEXT: MOV T15.Y, 0.0,
; EG-NEXT: LSHR T12.Z, T12.X, literal.y,
; EG-NEXT: AND_INT * T12.X, T12.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T12.Y, 0.0,
; EG-NEXT: LSHR * T16.Z, T11.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T16.X, T11.W, literal.x,
; EG-NEXT: MOV T16.Y, 0.0,
; EG-NEXT: LSHR T17.Z, T11.Z, literal.y,
; EG-NEXT: AND_INT * T17.X, T11.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T17.Y, 0.0,
; EG-NEXT: LSHR * T18.Z, T11.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T18.X, T11.Y, literal.x,
; EG-NEXT: MOV T18.Y, 0.0,
; EG-NEXT: LSHR T11.Z, T11.X, literal.y,
; EG-NEXT: AND_INT * T11.X, T11.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T11.Y, 0.0,
; EG-NEXT: MOV T13.W, 0.0,
; EG-NEXT: MOV * T14.W, 0.0,
; EG-NEXT: MOV T15.W, 0.0,
; EG-NEXT: MOV * T12.W, 0.0,
; EG-NEXT: MOV T16.W, 0.0,
; EG-NEXT: MOV * T17.W, 0.0,
; EG-NEXT: MOV T18.W, 0.0,
; EG-NEXT: MOV * T11.W, 0.0,
; EG-NEXT: LSHR T19.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T20.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T21.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR T22.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T23.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43)
; EG-NEXT: LSHR T24.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43)
; EG-NEXT: LSHR T25.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43)
; EG-NEXT: LSHR * T26.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v16i16_to_v16i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[8:11], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b256 s[0:7], s[10:11], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s10, s5, 0xffff
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s10
; GFX12-NEXT: s_lshr_b32 s5, s5, 16
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v3, v1
; GFX12-NEXT: s_lshr_b32 s5, s4, 16
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:80
; GFX12-NEXT: v_mov_b32_e32 v0, s4
; GFX12-NEXT: v_mov_b32_e32 v2, s5
; GFX12-NEXT: s_lshr_b32 s4, s7, 16
; GFX12-NEXT: s_and_b32 s5, s7, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:64
; GFX12-NEXT: v_mov_b32_e32 v0, s5
; GFX12-NEXT: v_mov_b32_e32 v2, s4
; GFX12-NEXT: s_lshr_b32 s4, s6, 16
; GFX12-NEXT: s_and_b32 s5, s6, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:112
; GFX12-NEXT: v_mov_b32_e32 v0, s5
; GFX12-NEXT: v_mov_b32_e32 v2, s4
; GFX12-NEXT: s_lshr_b32 s4, s3, 16
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:96
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s4
; GFX12-NEXT: s_lshr_b32 s3, s2, 16
; GFX12-NEXT: s_and_b32 s2, s2, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:48
; GFX12-NEXT: v_mov_b32_e32 v0, s2
; GFX12-NEXT: v_mov_b32_e32 v2, s3
; GFX12-NEXT: s_lshr_b32 s2, s1, 16
; GFX12-NEXT: s_and_b32 s1, s1, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:32
; GFX12-NEXT: v_mov_b32_e32 v0, s1
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_lshr_b32 s1, s0, 16
; GFX12-NEXT: s_and_b32 s0, s0, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9] offset:16
; GFX12-NEXT: v_mov_b32_e32 v0, s0
; GFX12-NEXT: v_mov_b32_e32 v2, s1
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[8:9]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <16 x i16>, ptr addrspace(4) %in
%ext = zext <16 x i16> %load to <16 x i64>
store <16 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v16i16_to_v16i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v16i16_to_v16i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[0:3], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx8 s[4:11], s[2:3], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s12, s11
; GCN-NOHSA-SI-NEXT: s_mov_b32 s14, s9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s16, s7
; GCN-NOHSA-SI-NEXT: s_mov_b32 s18, s5
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s20, s10, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s22, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s24, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s26, s4, 16
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[28:29], s[4:5], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[30:31], s[6:7], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[34:35], s[8:9], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[36:37], s[10:11], 0x100000
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[4:5], s[4:5], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[6:7], s[6:7], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[8:9], s[8:9], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[10:11], s[10:11], 48
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x100000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s11
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s36
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s37
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s34
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s35
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s30
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, s28
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s20
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s21
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:96
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s22
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s23
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:64
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s24
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s25
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, s26
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, s27
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v16i16_to_v16i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[0:3], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx8 s[8:15], s[2:3], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_mov_b32 s16, s15
; GCN-HSA-NEXT: s_mov_b32 s18, s13
; GCN-HSA-NEXT: s_mov_b32 s20, s11
; GCN-HSA-NEXT: s_mov_b32 s22, s9
; GCN-HSA-NEXT: s_lshr_b32 s24, s14, 16
; GCN-HSA-NEXT: s_lshr_b32 s26, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s28, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s4, s8, 16
; GCN-HSA-NEXT: s_bfe_i64 s[2:3], s[8:9], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[6:7], s[10:11], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[30:31], s[12:13], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[34:35], s[14:15], 0x100000
; GCN-HSA-NEXT: s_ashr_i64 s[8:9], s[8:9], 48
; GCN-HSA-NEXT: s_ashr_i64 s[10:11], s[10:11], 48
; GCN-HSA-NEXT: s_ashr_i64 s[12:13], s[12:13], 48
; GCN-HSA-NEXT: s_ashr_i64 s[14:15], s[14:15], 48
; GCN-HSA-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[22:23], s[22:23], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[16:17], s[16:17], 0x100000
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s14
; GCN-HSA-NEXT: s_add_u32 s14, s0, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s15
; GCN-HSA-NEXT: s_addc_u32 s15, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s17
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s15
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s12
; GCN-HSA-NEXT: s_add_u32 s12, s0, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s13
; GCN-HSA-NEXT: s_addc_u32 s13, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s13
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s10
; GCN-HSA-NEXT: s_add_u32 s10, s0, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s11
; GCN-HSA-NEXT: s_addc_u32 s11, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s11
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s8
; GCN-HSA-NEXT: s_add_u32 s8, s0, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s9
; GCN-HSA-NEXT: s_addc_u32 s9, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s22
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s9
; GCN-HSA-NEXT: s_add_u32 s8, s0, 0x60
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_addc_u32 s9, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s24
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s9
; GCN-HSA-NEXT: s_add_u32 s8, s0, 64
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_addc_u32 s9, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s30
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s26
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s9
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: s_add_u32 s6, s0, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s7
; GCN-HSA-NEXT: s_addc_u32 s7, s1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s28
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s7
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v16i16_to_v16i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[8:11], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx8 s[0:7], s[10:11], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_mov_b32 s12, s1
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s18, s1, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[16:17], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s2, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[22:23], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s6, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[28:29], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s4, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[14:15], s[12:13], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[12:13], s[18:19], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[18:19], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, s3
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[24:25], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, s7
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[30:31], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s5
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[10:11], s[0:1], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[20:21], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s3, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[26:27], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s7, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[34:35], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s5, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[0:1], s[0:1], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s8, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s35
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s8, 64
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s30
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s8, 0x70
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s8, 0x60
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s8, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s9, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s8, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s21
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s9, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s8, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s17
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s19
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s9, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s15
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s9
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v16i16_to_v16i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @16, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 1 @12
; EG-NEXT: ALU 65, @17, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T12.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T20.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T11.XYZW, T18.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T17.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T16.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T15.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T14.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T13.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 12:
; EG-NEXT: VTX_READ_128 T12.XYZW, T11.X, 16, #1
; EG-NEXT: VTX_READ_128 T11.XYZW, T11.X, 0, #1
; EG-NEXT: ALU clause starting at 16:
; EG-NEXT: MOV * T11.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 17:
; EG-NEXT: LSHR T13.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T14.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T15.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR T16.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T17.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43)
; EG-NEXT: LSHR T18.X, PV.W, literal.x,
; EG-NEXT: ADD_INT T0.W, KC0[2].Y, literal.y,
; EG-NEXT: ASHR * T19.W, T11.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T20.X, PV.W, literal.x,
; EG-NEXT: ASHR T19.Z, T11.X, literal.y,
; EG-NEXT: ASHR * T21.W, T11.Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T19.X, T11.X, 0.0, literal.x,
; EG-NEXT: ASHR T21.Z, T11.Y, literal.x,
; EG-NEXT: ASHR * T22.W, T11.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T21.X, T11.Y, 0.0, literal.x,
; EG-NEXT: ASHR T19.Y, PV.X, literal.y,
; EG-NEXT: ASHR T22.Z, T11.Z, literal.x,
; EG-NEXT: ASHR * T23.W, T11.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T22.X, T11.Z, 0.0, literal.x,
; EG-NEXT: ASHR T21.Y, PV.X, literal.y,
; EG-NEXT: ASHR T23.Z, T11.W, literal.x,
; EG-NEXT: ASHR * T24.W, T12.X, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T23.X, T11.W, 0.0, literal.x,
; EG-NEXT: ASHR T22.Y, PV.X, literal.y,
; EG-NEXT: ASHR T24.Z, T12.X, literal.x,
; EG-NEXT: ASHR * T11.W, T12.Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T24.X, T12.X, 0.0, literal.x,
; EG-NEXT: ASHR T23.Y, PV.X, literal.y,
; EG-NEXT: ASHR T11.Z, T12.Y, literal.x,
; EG-NEXT: ASHR * T25.W, T12.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T11.X, T12.Y, 0.0, literal.x,
; EG-NEXT: ASHR T24.Y, PV.X, literal.y,
; EG-NEXT: ASHR T25.Z, T12.Z, literal.x,
; EG-NEXT: ASHR * T26.W, T12.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T25.X, T12.Z, 0.0, literal.x,
; EG-NEXT: ASHR T11.Y, PV.X, literal.y,
; EG-NEXT: ASHR * T26.Z, T12.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T26.X, T12.W, 0.0, literal.x,
; EG-NEXT: ASHR T25.Y, PV.X, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: 112(1.569454e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T12.X, PV.W, literal.x,
; EG-NEXT: ASHR * T26.Y, PV.X, literal.y,
; EG-NEXT: 2(2.802597e-45), 31(4.344025e-44)
;
; GFX12-LABEL: constant_sextload_v16i16_to_v16i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[8:11], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b256 s[0:7], s[10:11], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_mov_b32 s30, s5
; GFX12-NEXT: s_lshr_b32 s34, s5, 16
; GFX12-NEXT: s_bfe_i64 s[28:29], s[4:5], 0x100000
; GFX12-NEXT: s_lshr_b32 s4, s4, 16
; GFX12-NEXT: s_bfe_i64 s[22:23], s[6:7], 0x100000
; GFX12-NEXT: s_mov_b32 s24, s7
; GFX12-NEXT: s_lshr_b32 s26, s7, 16
; GFX12-NEXT: s_bfe_i64 s[30:31], s[30:31], 0x100000
; GFX12-NEXT: s_bfe_i64 s[34:35], s[34:35], 0x100000
; GFX12-NEXT: s_lshr_b32 s6, s6, 16
; GFX12-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s29
; GFX12-NEXT: s_mov_b32 s18, s3
; GFX12-NEXT: s_lshr_b32 s20, s3, 16
; GFX12-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x100000
; GFX12-NEXT: s_bfe_i64 s[26:27], s[26:27], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v4, s22 :: v_dual_mov_b32 v9, s31
; GFX12-NEXT: v_dual_mov_b32 v8, s30 :: v_dual_mov_b32 v11, s35
; GFX12-NEXT: v_dual_mov_b32 v10, s34 :: v_dual_mov_b32 v3, s5
; GFX12-NEXT: s_bfe_i64 s[16:17], s[2:3], 0x100000
; GFX12-NEXT: s_lshr_b32 s2, s2, 16
; GFX12-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v0, s28 :: v_dual_mov_b32 v5, s23
; GFX12-NEXT: v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v13, s25
; GFX12-NEXT: s_mov_b32 s12, s1
; GFX12-NEXT: s_lshr_b32 s14, s1, 16
; GFX12-NEXT: s_bfe_i64 s[18:19], s[18:19], 0x100000
; GFX12-NEXT: s_bfe_i64 s[20:21], s[20:21], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v15, s27
; GFX12-NEXT: v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v7, s7
; GFX12-NEXT: s_bfe_i64 s[10:11], s[0:1], 0x100000
; GFX12-NEXT: s_lshr_b32 s0, s0, 16
; GFX12-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v6, s6 :: v_dual_mov_b32 v17, s19
; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v16, s18 :: v_dual_mov_b32 v19, s21
; GFX12-NEXT: v_mov_b32_e32 v18, s20
; GFX12-NEXT: s_bfe_i64 s[0:1], s[0:1], 0x100000
; GFX12-NEXT: s_clause 0x1
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[8:9] offset:80
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[8:9] offset:64
; GFX12-NEXT: v_dual_mov_b32 v1, s17 :: v_dual_mov_b32 v0, s16
; GFX12-NEXT: v_dual_mov_b32 v3, s3 :: v_dual_mov_b32 v2, s2
; GFX12-NEXT: v_dual_mov_b32 v9, s13 :: v_dual_mov_b32 v8, s12
; GFX12-NEXT: v_dual_mov_b32 v11, s15 :: v_dual_mov_b32 v10, s14
; GFX12-NEXT: v_dual_mov_b32 v21, s11 :: v_dual_mov_b32 v20, s10
; GFX12-NEXT: v_dual_mov_b32 v23, s1 :: v_dual_mov_b32 v22, s0
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[8:9] offset:112
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[8:9] offset:96
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[8:9] offset:48
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[8:9] offset:32
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[8:9] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[8:9]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <16 x i16>, ptr addrspace(4) %in
%ext = sext <16 x i16> %load to <16 x i64>
store <16 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_zextload_v32i16_to_v32i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_zextload_v32i16_to_v32i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s19, 0xf000
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-SI-NEXT: s_mov_b32 s18, -1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s20, s1, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s21, s3, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s22, s5, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s23, s7, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s24, s9, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s25, s11, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s26, s13, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s27, s15, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s28, s14, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s29, s12, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s30, s10, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s31, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s33, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s34, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s35, s2, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s36, s0, 16
; GCN-NOHSA-SI-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s15, s15, 0xffff
; GCN-NOHSA-SI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s15
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s27
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:240
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s13
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s26
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:208
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s25
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:176
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s24
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:144
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s7
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s23
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s22
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s1
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s20
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s28
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:224
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s29
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:192
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s30
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:160
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s31
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:128
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s33
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:96
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s34
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:64
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s35
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0 offset:32
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s36
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[16:19], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_zextload_v32i16_to_v32i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: v_mov_b32_e32 v1, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, v1
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_lshr_b32 s19, s1, 16
; GCN-HSA-NEXT: s_lshr_b32 s20, s3, 16
; GCN-HSA-NEXT: s_lshr_b32 s21, s5, 16
; GCN-HSA-NEXT: s_lshr_b32 s22, s7, 16
; GCN-HSA-NEXT: s_lshr_b32 s23, s9, 16
; GCN-HSA-NEXT: s_lshr_b32 s24, s11, 16
; GCN-HSA-NEXT: s_lshr_b32 s25, s13, 16
; GCN-HSA-NEXT: s_lshr_b32 s26, s15, 16
; GCN-HSA-NEXT: s_lshr_b32 s27, s14, 16
; GCN-HSA-NEXT: s_lshr_b32 s28, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s29, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s30, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s31, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s33, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s34, s2, 16
; GCN-HSA-NEXT: s_lshr_b32 s18, s0, 16
; GCN-HSA-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-HSA-NEXT: s_and_b32 s35, s2, 0xffff
; GCN-HSA-NEXT: s_and_b32 s4, s4, 0xffff
; GCN-HSA-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-HSA-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-HSA-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-HSA-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-HSA-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-HSA-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-HSA-NEXT: s_and_b32 s36, s3, 0xffff
; GCN-HSA-NEXT: s_and_b32 s5, s5, 0xffff
; GCN-HSA-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-HSA-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-HSA-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-HSA-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-HSA-NEXT: s_and_b32 s2, s15, 0xffff
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xf0
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xd0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s26
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xb0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s13
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s25
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0x90
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s11
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s24
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s9
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s23
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s7
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s22
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s5
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s21
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s36
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s20
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xe0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s19
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xc0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s27
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0xa0
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s28
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0x80
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s29
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s30
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s31
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s33
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_zextload_v32i16_to_v32i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, v1
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s18, s0, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s0, s0, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s19, s1, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s1, s1, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s20, s2, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s2, s2, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s21, s3, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s3, s3, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s22, s4, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s23, s4, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s24, s5, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s25, s5, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s26, s6, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s6, s6, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s27, s7, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s7, s7, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s28, s8, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s8, s8, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s29, s9, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s9, s9, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s30, s10, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s10, s10, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s31, s11, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s11, s11, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s33, s12, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s12, s12, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s34, s13, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s13, s13, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s35, s14, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s14, s14, 0xffff
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s15, 16
; GCN-NOHSA-VI-NEXT: s_and_b32 s5, s15, 0xffff
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xf0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s5
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xe0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s35
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xd0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s13
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xc0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s33
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xb0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s11
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0xa0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s30
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x90
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s9
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s29
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x80
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x70
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s7
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s27
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x60
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x50
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 64
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 48
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s16, 32
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s16, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s20
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s17
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_zextload_v32i16_to_v32i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @30, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 2 @22
; EG-NEXT: ALU 33, @31, KC0[], KC1[]
; EG-NEXT: TEX 0 @28
; EG-NEXT: ALU 92, @65, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T23.XYZW, T50.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T24.XYZW, T49.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T25.XYZW, T48.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T47.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T26.XYZW, T46.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T27.XYZW, T45.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T28.XYZW, T44.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T22.XYZW, T43.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T42.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T30.XYZW, T41.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T31.XYZW, T40.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T21.XYZW, T39.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T32.XYZW, T38.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T33.XYZW, T37.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T34.XYZW, T36.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T29.XYZW, T35.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 22:
; EG-NEXT: VTX_READ_128 T20.XYZW, T19.X, 48, #1
; EG-NEXT: VTX_READ_128 T21.XYZW, T19.X, 16, #1
; EG-NEXT: VTX_READ_128 T22.XYZW, T19.X, 32, #1
; EG-NEXT: Fetch clause starting at 28:
; EG-NEXT: VTX_READ_128 T29.XYZW, T19.X, 0, #1
; EG-NEXT: ALU clause starting at 30:
; EG-NEXT: MOV * T19.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 31:
; EG-NEXT: LSHR * T23.Z, T20.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T23.X, T20.W, literal.x,
; EG-NEXT: MOV T23.Y, 0.0,
; EG-NEXT: LSHR T24.Z, T20.Z, literal.y,
; EG-NEXT: AND_INT * T24.X, T20.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T24.Y, 0.0,
; EG-NEXT: LSHR * T25.Z, T20.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T25.X, T20.Y, literal.x,
; EG-NEXT: MOV T25.Y, 0.0,
; EG-NEXT: LSHR T20.Z, T20.X, literal.y,
; EG-NEXT: AND_INT * T20.X, T20.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T20.Y, 0.0,
; EG-NEXT: LSHR * T26.Z, T22.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T26.X, T22.W, literal.x,
; EG-NEXT: MOV T26.Y, 0.0,
; EG-NEXT: LSHR T27.Z, T22.Z, literal.y,
; EG-NEXT: AND_INT * T27.X, T22.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T27.Y, 0.0,
; EG-NEXT: LSHR * T28.Z, T22.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T28.X, T22.Y, literal.x,
; EG-NEXT: MOV T28.Y, 0.0,
; EG-NEXT: LSHR T22.Z, T22.X, literal.y,
; EG-NEXT: AND_INT * T22.X, T22.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T22.Y, 0.0,
; EG-NEXT: LSHR * T19.Z, T21.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: ALU clause starting at 65:
; EG-NEXT: AND_INT T19.X, T21.W, literal.x,
; EG-NEXT: MOV T19.Y, 0.0,
; EG-NEXT: LSHR T30.Z, T21.Z, literal.y,
; EG-NEXT: AND_INT * T30.X, T21.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T30.Y, 0.0,
; EG-NEXT: LSHR * T31.Z, T21.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T31.X, T21.Y, literal.x,
; EG-NEXT: MOV T31.Y, 0.0,
; EG-NEXT: LSHR T21.Z, T21.X, literal.y,
; EG-NEXT: AND_INT * T21.X, T21.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T21.Y, 0.0,
; EG-NEXT: LSHR * T32.Z, T29.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T32.X, T29.W, literal.x,
; EG-NEXT: MOV T32.Y, 0.0,
; EG-NEXT: LSHR T33.Z, T29.Z, literal.y,
; EG-NEXT: AND_INT * T33.X, T29.Z, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T33.Y, 0.0,
; EG-NEXT: LSHR * T34.Z, T29.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: AND_INT T34.X, T29.Y, literal.x,
; EG-NEXT: MOV T34.Y, 0.0,
; EG-NEXT: LSHR T29.Z, T29.X, literal.y,
; EG-NEXT: AND_INT * T29.X, T29.X, literal.x,
; EG-NEXT: 65535(9.183409e-41), 16(2.242078e-44)
; EG-NEXT: MOV T29.Y, 0.0,
; EG-NEXT: MOV T23.W, 0.0,
; EG-NEXT: MOV * T24.W, 0.0,
; EG-NEXT: MOV T25.W, 0.0,
; EG-NEXT: MOV * T20.W, 0.0,
; EG-NEXT: MOV T26.W, 0.0,
; EG-NEXT: MOV * T27.W, 0.0,
; EG-NEXT: MOV T28.W, 0.0,
; EG-NEXT: MOV * T22.W, 0.0,
; EG-NEXT: MOV T19.W, 0.0,
; EG-NEXT: MOV * T30.W, 0.0,
; EG-NEXT: MOV T31.W, 0.0,
; EG-NEXT: MOV * T21.W, 0.0,
; EG-NEXT: MOV T32.W, 0.0,
; EG-NEXT: MOV * T33.W, 0.0,
; EG-NEXT: MOV T34.W, 0.0,
; EG-NEXT: MOV * T29.W, 0.0,
; EG-NEXT: LSHR T35.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T36.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T37.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR T38.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T39.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43)
; EG-NEXT: LSHR T40.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43)
; EG-NEXT: LSHR T41.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43)
; EG-NEXT: LSHR T42.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43)
; EG-NEXT: LSHR T43.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43)
; EG-NEXT: LSHR T44.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43)
; EG-NEXT: LSHR T45.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43)
; EG-NEXT: LSHR T46.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 192(2.690493e-43)
; EG-NEXT: LSHR T47.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 208(2.914701e-43)
; EG-NEXT: LSHR T48.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 224(3.138909e-43)
; EG-NEXT: LSHR T49.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 240(3.363116e-43)
; EG-NEXT: LSHR * T50.X, PV.W, literal.x,
; EG-NEXT: 2(2.802597e-45), 0(0.000000e+00)
;
; GFX12-LABEL: constant_zextload_v32i16_to_v32i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[16:19], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b512 s[0:15], s[18:19], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_and_b32 s18, s15, 0xffff
; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s18
; GFX12-NEXT: s_lshr_b32 s15, s15, 16
; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) | instid1(SALU_CYCLE_1)
; GFX12-NEXT: v_dual_mov_b32 v2, s15 :: v_dual_mov_b32 v3, v1
; GFX12-NEXT: s_lshr_b32 s15, s14, 16
; GFX12-NEXT: s_and_b32 s14, s14, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:240
; GFX12-NEXT: v_mov_b32_e32 v0, s14
; GFX12-NEXT: v_mov_b32_e32 v2, s15
; GFX12-NEXT: s_lshr_b32 s14, s13, 16
; GFX12-NEXT: s_and_b32 s13, s13, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:224
; GFX12-NEXT: v_mov_b32_e32 v0, s13
; GFX12-NEXT: v_mov_b32_e32 v2, s14
; GFX12-NEXT: s_lshr_b32 s13, s12, 16
; GFX12-NEXT: s_and_b32 s12, s12, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:208
; GFX12-NEXT: v_mov_b32_e32 v0, s12
; GFX12-NEXT: v_mov_b32_e32 v2, s13
; GFX12-NEXT: s_lshr_b32 s12, s11, 16
; GFX12-NEXT: s_and_b32 s11, s11, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:192
; GFX12-NEXT: v_mov_b32_e32 v0, s11
; GFX12-NEXT: v_mov_b32_e32 v2, s12
; GFX12-NEXT: s_lshr_b32 s11, s10, 16
; GFX12-NEXT: s_and_b32 s10, s10, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:176
; GFX12-NEXT: v_mov_b32_e32 v0, s10
; GFX12-NEXT: v_mov_b32_e32 v2, s11
; GFX12-NEXT: s_lshr_b32 s10, s9, 16
; GFX12-NEXT: s_and_b32 s9, s9, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:160
; GFX12-NEXT: v_mov_b32_e32 v0, s9
; GFX12-NEXT: v_mov_b32_e32 v2, s10
; GFX12-NEXT: s_lshr_b32 s9, s8, 16
; GFX12-NEXT: s_and_b32 s8, s8, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:144
; GFX12-NEXT: v_mov_b32_e32 v0, s8
; GFX12-NEXT: v_mov_b32_e32 v2, s9
; GFX12-NEXT: s_lshr_b32 s8, s7, 16
; GFX12-NEXT: s_and_b32 s7, s7, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:128
; GFX12-NEXT: v_mov_b32_e32 v0, s7
; GFX12-NEXT: v_mov_b32_e32 v2, s8
; GFX12-NEXT: s_lshr_b32 s7, s6, 16
; GFX12-NEXT: s_and_b32 s6, s6, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:112
; GFX12-NEXT: v_mov_b32_e32 v0, s6
; GFX12-NEXT: v_mov_b32_e32 v2, s7
; GFX12-NEXT: s_lshr_b32 s6, s5, 16
; GFX12-NEXT: s_and_b32 s5, s5, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:96
; GFX12-NEXT: v_mov_b32_e32 v0, s5
; GFX12-NEXT: v_mov_b32_e32 v2, s6
; GFX12-NEXT: s_lshr_b32 s5, s4, 16
; GFX12-NEXT: s_and_b32 s4, s4, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:80
; GFX12-NEXT: v_mov_b32_e32 v0, s4
; GFX12-NEXT: v_mov_b32_e32 v2, s5
; GFX12-NEXT: s_lshr_b32 s4, s3, 16
; GFX12-NEXT: s_and_b32 s3, s3, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:64
; GFX12-NEXT: v_mov_b32_e32 v0, s3
; GFX12-NEXT: v_mov_b32_e32 v2, s4
; GFX12-NEXT: s_lshr_b32 s3, s2, 16
; GFX12-NEXT: s_and_b32 s2, s2, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:48
; GFX12-NEXT: v_mov_b32_e32 v0, s2
; GFX12-NEXT: v_mov_b32_e32 v2, s3
; GFX12-NEXT: s_lshr_b32 s2, s1, 16
; GFX12-NEXT: s_and_b32 s1, s1, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:32
; GFX12-NEXT: v_mov_b32_e32 v0, s1
; GFX12-NEXT: v_mov_b32_e32 v2, s2
; GFX12-NEXT: s_lshr_b32 s1, s0, 16
; GFX12-NEXT: s_and_b32 s0, s0, 0xffff
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17] offset:16
; GFX12-NEXT: v_mov_b32_e32 v0, s0
; GFX12-NEXT: v_mov_b32_e32 v2, s1
; GFX12-NEXT: global_store_b128 v1, v[0:3], s[16:17]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <32 x i16>, ptr addrspace(4) %in
%ext = zext <32 x i16> %load to <32 x i64>
store <32 x i64> %ext, ptr addrspace(1) %out
ret void
}
define amdgpu_kernel void @constant_sextload_v32i16_to_v32i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; GCN-NOHSA-SI-LABEL: constant_sextload_v32i16_to_v32i64:
; GCN-NOHSA-SI: ; %bb.0:
; GCN-NOHSA-SI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x9
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-SI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-SI-NEXT: s_mov_b32 s18, s15
; GCN-NOHSA-SI-NEXT: s_mov_b32 s20, s13
; GCN-NOHSA-SI-NEXT: s_mov_b32 s22, s11
; GCN-NOHSA-SI-NEXT: s_mov_b32 s30, s9
; GCN-NOHSA-SI-NEXT: s_mov_b32 s28, s7
; GCN-NOHSA-SI-NEXT: s_mov_b32 s36, s5
; GCN-NOHSA-SI-NEXT: s_mov_b32 s40, s3
; GCN-NOHSA-SI-NEXT: s_mov_b32 s44, s1
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s24, s14, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s26, s12, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s34, s10, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s38, s8, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s42, s6, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s46, s4, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s48, s2, 16
; GCN-NOHSA-SI-NEXT: s_lshr_b32 s52, s0, 16
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[62:63], s[20:21], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[64:65], s[18:19], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[66:67], s[30:31], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[68:69], s[22:23], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[18:19], s[0:1], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[20:21], s[2:3], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[22:23], s[4:5], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[30:31], s[6:7], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[50:51], s[8:9], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[54:55], s[10:11], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[56:57], s[12:13], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[58:59], s[14:15], 0x100000
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[60:61], s[0:1], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[70:71], s[2:3], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[72:73], s[4:5], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[2:3], s[8:9], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[4:5], s[12:13], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[8:9], s[14:15], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[10:11], s[10:11], 48
; GCN-NOHSA-SI-NEXT: s_ashr_i64 s[74:75], s[6:7], 48
; GCN-NOHSA-SI-NEXT: s_mov_b32 s0, s16
; GCN-NOHSA-SI-NEXT: s_mov_b32 s1, s17
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s64
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s65
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s62
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s63
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s5
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s68
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s69
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s11
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, s66
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, s67
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, s2
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, s3
; GCN-NOHSA-SI-NEXT: s_mov_b32 s3, 0xf000
; GCN-NOHSA-SI-NEXT: s_mov_b32 s2, -1
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[44:45], s[44:45], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[40:41], s[40:41], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[36:37], s[36:37], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[28:29], s[28:29], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[4:5], s[52:53], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[6:7], s[48:49], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[8:9], s[46:47], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[10:11], s[42:43], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[12:13], s[38:39], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[14:15], s[34:35], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[16:17], s[26:27], 0x100000
; GCN-NOHSA-SI-NEXT: s_bfe_i64 s[24:25], s[24:25], 0x100000
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:240
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:208
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:176
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:144
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(3)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s74
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s75
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:112
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s36
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s37
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s72
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s73
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:80
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s40
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s41
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s70
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s71
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:48
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s44
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s45
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s60
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s61
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:16
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s58
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s59
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s56
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s57
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s54
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s55
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v12, s50
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v13, s51
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v16, s30
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v17, s31
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s24
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s25
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:224
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s16
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s17
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:192
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v4, s20
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v5, s21
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s14
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s15
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0 offset:160
; GCN-NOHSA-SI-NEXT: s_waitcnt expcnt(0)
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v8, s18
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v9, s19
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v14, s12
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v15, s13
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[12:15], off, s[0:3], 0 offset:128
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v18, s10
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v19, s11
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[16:19], off, s[0:3], 0 offset:96
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[0:3], off, s[0:3], 0 offset:64
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v6, s6
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v7, s7
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[4:7], off, s[0:3], 0 offset:32
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v10, s4
; GCN-NOHSA-SI-NEXT: v_mov_b32_e32 v11, s5
; GCN-NOHSA-SI-NEXT: buffer_store_dwordx4 v[8:11], off, s[0:3], 0
; GCN-NOHSA-SI-NEXT: s_endpgm
;
; GCN-HSA-LABEL: constant_sextload_v32i16_to_v32i64:
; GCN-HSA: ; %bb.0:
; GCN-HSA-NEXT: s_load_dwordx4 s[16:19], s[6:7], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-HSA-NEXT: s_waitcnt lgkmcnt(0)
; GCN-HSA-NEXT: s_mov_b32 s42, s15
; GCN-HSA-NEXT: s_mov_b32 s48, s13
; GCN-HSA-NEXT: s_mov_b32 s50, s11
; GCN-HSA-NEXT: s_mov_b32 s52, s9
; GCN-HSA-NEXT: s_mov_b32 s54, s7
; GCN-HSA-NEXT: s_mov_b32 s56, s5
; GCN-HSA-NEXT: s_mov_b32 s46, s3
; GCN-HSA-NEXT: s_mov_b32 s58, s1
; GCN-HSA-NEXT: s_lshr_b32 s60, s14, 16
; GCN-HSA-NEXT: s_lshr_b32 s62, s12, 16
; GCN-HSA-NEXT: s_lshr_b32 s64, s10, 16
; GCN-HSA-NEXT: s_lshr_b32 s66, s8, 16
; GCN-HSA-NEXT: s_lshr_b32 s68, s6, 16
; GCN-HSA-NEXT: s_lshr_b32 s70, s4, 16
; GCN-HSA-NEXT: s_lshr_b32 s72, s2, 16
; GCN-HSA-NEXT: s_lshr_b32 s74, s0, 16
; GCN-HSA-NEXT: s_bfe_i64 s[18:19], s[0:1], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[20:21], s[2:3], 0x100000
; GCN-HSA-NEXT: s_ashr_i64 s[36:37], s[0:1], 48
; GCN-HSA-NEXT: s_ashr_i64 s[38:39], s[2:3], 48
; GCN-HSA-NEXT: s_ashr_i64 s[0:1], s[14:15], 48
; GCN-HSA-NEXT: s_bfe_i64 s[2:3], s[42:43], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[22:23], s[4:5], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[24:25], s[6:7], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[26:27], s[8:9], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[28:29], s[10:11], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[30:31], s[12:13], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[34:35], s[14:15], 0x100000
; GCN-HSA-NEXT: s_ashr_i64 s[40:41], s[4:5], 48
; GCN-HSA-NEXT: s_ashr_i64 s[44:45], s[6:7], 48
; GCN-HSA-NEXT: s_ashr_i64 s[76:77], s[8:9], 48
; GCN-HSA-NEXT: s_ashr_i64 s[78:79], s[10:11], 48
; GCN-HSA-NEXT: s_ashr_i64 s[80:81], s[12:13], 48
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s2
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: s_bfe_i64 s[0:1], s[74:75], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[2:3], s[72:73], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[4:5], s[70:71], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[6:7], s[68:69], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[8:9], s[66:67], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[10:11], s[64:65], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[12:13], s[62:63], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[14:15], s[60:61], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[42:43], s[58:59], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[46:47], s[46:47], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[56:57], s[56:57], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[52:53], s[52:53], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[50:51], s[50:51], 0x100000
; GCN-HSA-NEXT: s_bfe_i64 s[48:49], s[48:49], 0x100000
; GCN-HSA-NEXT: s_add_u32 s58, s16, 0xf0
; GCN-HSA-NEXT: s_addc_u32 s59, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s48
; GCN-HSA-NEXT: s_add_u32 s48, s16, 0xd0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s49
; GCN-HSA-NEXT: s_addc_u32 s49, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v24, s48
; GCN-HSA-NEXT: v_mov_b32_e32 v25, s49
; GCN-HSA-NEXT: s_add_u32 s48, s16, 0xb0
; GCN-HSA-NEXT: s_addc_u32 s49, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v26, s48
; GCN-HSA-NEXT: v_mov_b32_e32 v27, s49
; GCN-HSA-NEXT: s_add_u32 s48, s16, 0x90
; GCN-HSA-NEXT: s_addc_u32 s49, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v18, s44
; GCN-HSA-NEXT: s_add_u32 s44, s16, 0x70
; GCN-HSA-NEXT: v_mov_b32_e32 v19, s45
; GCN-HSA-NEXT: s_addc_u32 s45, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v30, s44
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s58
; GCN-HSA-NEXT: v_mov_b32_e32 v31, s45
; GCN-HSA-NEXT: s_add_u32 s44, s16, 0x50
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s59
; GCN-HSA-NEXT: s_addc_u32 s45, s17, 0
; GCN-HSA-NEXT: flat_store_dwordx4 v[22:23], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s80
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s38
; GCN-HSA-NEXT: s_add_u32 s38, s16, 48
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s81
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s39
; GCN-HSA-NEXT: s_addc_u32 s39, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v8, s50
; GCN-HSA-NEXT: v_mov_b32_e32 v9, s51
; GCN-HSA-NEXT: v_mov_b32_e32 v10, s78
; GCN-HSA-NEXT: v_mov_b32_e32 v11, s79
; GCN-HSA-NEXT: v_mov_b32_e32 v28, s48
; GCN-HSA-NEXT: flat_store_dwordx4 v[24:25], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v12, s52
; GCN-HSA-NEXT: v_mov_b32_e32 v6, s36
; GCN-HSA-NEXT: s_add_u32 s36, s16, 16
; GCN-HSA-NEXT: v_mov_b32_e32 v13, s53
; GCN-HSA-NEXT: v_mov_b32_e32 v14, s76
; GCN-HSA-NEXT: v_mov_b32_e32 v15, s77
; GCN-HSA-NEXT: v_mov_b32_e32 v29, s49
; GCN-HSA-NEXT: v_mov_b32_e32 v16, s54
; GCN-HSA-NEXT: v_mov_b32_e32 v17, s55
; GCN-HSA-NEXT: v_mov_b32_e32 v32, s44
; GCN-HSA-NEXT: flat_store_dwordx4 v[26:27], v[8:11]
; GCN-HSA-NEXT: v_mov_b32_e32 v7, s37
; GCN-HSA-NEXT: v_mov_b32_e32 v8, s38
; GCN-HSA-NEXT: s_addc_u32 s37, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v10, s36
; GCN-HSA-NEXT: v_mov_b32_e32 v20, s56
; GCN-HSA-NEXT: v_mov_b32_e32 v21, s57
; GCN-HSA-NEXT: v_mov_b32_e32 v33, s45
; GCN-HSA-NEXT: v_mov_b32_e32 v22, s40
; GCN-HSA-NEXT: v_mov_b32_e32 v23, s41
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s46
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s47
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s42
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s43
; GCN-HSA-NEXT: v_mov_b32_e32 v9, s39
; GCN-HSA-NEXT: flat_store_dwordx4 v[28:29], v[12:15]
; GCN-HSA-NEXT: v_mov_b32_e32 v11, s37
; GCN-HSA-NEXT: flat_store_dwordx4 v[30:31], v[16:19]
; GCN-HSA-NEXT: flat_store_dwordx4 v[32:33], v[20:23]
; GCN-HSA-NEXT: flat_store_dwordx4 v[8:9], v[0:3]
; GCN-HSA-NEXT: flat_store_dwordx4 v[10:11], v[4:7]
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s14
; GCN-HSA-NEXT: s_add_u32 s14, s16, 0xe0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s15
; GCN-HSA-NEXT: s_addc_u32 s15, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s14
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s34
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s35
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s15
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s12
; GCN-HSA-NEXT: s_add_u32 s12, s16, 0xc0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s13
; GCN-HSA-NEXT: s_addc_u32 s13, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s12
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s30
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s31
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s13
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s10
; GCN-HSA-NEXT: s_add_u32 s10, s16, 0xa0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s11
; GCN-HSA-NEXT: s_addc_u32 s11, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s10
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s28
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s29
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s11
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s8
; GCN-HSA-NEXT: s_add_u32 s8, s16, 0x80
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s9
; GCN-HSA-NEXT: s_addc_u32 s9, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s8
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s26
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s27
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s9
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s6
; GCN-HSA-NEXT: s_add_u32 s6, s16, 0x60
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s7
; GCN-HSA-NEXT: s_addc_u32 s7, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s6
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s24
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s25
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s7
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s4
; GCN-HSA-NEXT: s_add_u32 s4, s16, 64
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s5
; GCN-HSA-NEXT: s_addc_u32 s5, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s4
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s22
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s23
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s5
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_nop 0
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s2
; GCN-HSA-NEXT: s_add_u32 s2, s16, 32
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s3
; GCN-HSA-NEXT: s_addc_u32 s3, s17, 0
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s3
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s20
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s21
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s2
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: v_mov_b32_e32 v4, s16
; GCN-HSA-NEXT: v_mov_b32_e32 v0, s18
; GCN-HSA-NEXT: v_mov_b32_e32 v1, s19
; GCN-HSA-NEXT: v_mov_b32_e32 v2, s0
; GCN-HSA-NEXT: v_mov_b32_e32 v3, s1
; GCN-HSA-NEXT: v_mov_b32_e32 v5, s17
; GCN-HSA-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-HSA-NEXT: s_endpgm
;
; GCN-NOHSA-VI-LABEL: constant_sextload_v32i16_to_v32i64:
; GCN-NOHSA-VI: ; %bb.0:
; GCN-NOHSA-VI-NEXT: s_load_dwordx4 s[16:19], s[2:3], 0x24
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_load_dwordx16 s[0:15], s[18:19], 0x0
; GCN-NOHSA-VI-NEXT: s_waitcnt lgkmcnt(0)
; GCN-NOHSA-VI-NEXT: s_mov_b32 s20, s1
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s26, s1, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[24:25], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s2, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[30:31], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s4, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[38:39], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s6, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[44:45], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s8, s8, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[50:51], s[10:11], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s10, s10, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[56:57], s[12:13], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s12, s12, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[62:63], s[14:15], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s14, s14, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[22:23], s[20:21], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[20:21], s[26:27], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[26:27], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s2, s3
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[34:35], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s4, s5
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[40:41], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s6, s7
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[46:47], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s8, s9
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[52:53], s[10:11], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s10, s11
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[58:59], s[12:13], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s12, s13
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[64:65], s[14:15], 0x100000
; GCN-NOHSA-VI-NEXT: s_mov_b32 s14, s15
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[18:19], s[0:1], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s0, s0, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[28:29], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s2, s3, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[36:37], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s4, s5, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[42:43], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s6, s7, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[48:49], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s8, s9, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[54:55], s[10:11], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s10, s11, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[60:61], s[12:13], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s12, s13, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[66:67], s[14:15], 0x100000
; GCN-NOHSA-VI-NEXT: s_lshr_b32 s14, s15, 16
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[0:1], s[0:1], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[2:3], s[2:3], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[4:5], s[4:5], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[6:7], s[6:7], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[8:9], s[8:9], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[10:11], s[10:11], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GCN-NOHSA-VI-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s14
; GCN-NOHSA-VI-NEXT: s_add_u32 s14, s16, 0xf0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s15
; GCN-NOHSA-VI-NEXT: s_addc_u32 s15, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s66
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s67
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s15
; GCN-NOHSA-VI-NEXT: s_add_u32 s14, s16, 0xe0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s15, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s14
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s62
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s63
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s64
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s65
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s15
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s12
; GCN-NOHSA-VI-NEXT: s_add_u32 s12, s16, 0xd0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s13
; GCN-NOHSA-VI-NEXT: s_addc_u32 s13, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s60
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s61
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s13
; GCN-NOHSA-VI-NEXT: s_add_u32 s12, s16, 0xc0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s13, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s12
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s56
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s57
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s58
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s59
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s13
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s10
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s16, 0xb0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s11
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s54
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s55
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: s_add_u32 s10, s16, 0xa0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s11, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s10
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s51
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s52
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s53
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s11
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s8
; GCN-NOHSA-VI-NEXT: s_add_u32 s8, s16, 0x90
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s9
; GCN-NOHSA-VI-NEXT: s_addc_u32 s9, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s49
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s9
; GCN-NOHSA-VI-NEXT: s_add_u32 s8, s16, 0x80
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s9, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s8
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s44
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s45
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s46
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s47
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s9
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s6
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s16, 0x70
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s7
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s42
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s43
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: s_add_u32 s6, s16, 0x60
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s7, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s6
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s38
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s39
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s40
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s41
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s7
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s4
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 0x50
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s5
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s36
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s37
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: s_add_u32 s4, s16, 64
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_addc_u32 s5, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s4
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s30
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s31
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s34
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s35
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s5
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_nop 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s16, 48
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s3
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s17, 0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s16, 32
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s28
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s29
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: s_add_u32 s2, s16, 16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s24
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s25
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s26
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s27
; GCN-NOHSA-VI-NEXT: s_addc_u32 s3, s17, 0
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s3
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s22
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s23
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s20
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s21
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s2
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v4, s16
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v0, s18
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v1, s19
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v2, s0
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v3, s1
; GCN-NOHSA-VI-NEXT: v_mov_b32_e32 v5, s17
; GCN-NOHSA-VI-NEXT: flat_store_dwordx4 v[4:5], v[0:3]
; GCN-NOHSA-VI-NEXT: s_endpgm
;
; EG-LABEL: constant_sextload_v32i16_to_v32i64:
; EG: ; %bb.0:
; EG-NEXT: ALU 0, @30, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 0 @22
; EG-NEXT: ALU 55, @31, KC0[CB0:0-32], KC1[]
; EG-NEXT: TEX 2 @24
; EG-NEXT: ALU 74, @87, KC0[CB0:0-32], KC1[]
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T50.XYZW, T38.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T49.XYZW, T36.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T39.XYZW, T34.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T48.XYZW, T33.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T47.XYZW, T32.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T46.XYZW, T31.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T40.XYZW, T30.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T45.XYZW, T29.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T44.XYZW, T28.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T43.XYZW, T27.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T20.XYZW, T26.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T42.XYZW, T25.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T41.XYZW, T24.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T19.XYZW, T23.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T37.XYZW, T22.X, 0
; EG-NEXT: MEM_RAT_CACHELESS STORE_RAW T35.XYZW, T21.X, 1
; EG-NEXT: CF_END
; EG-NEXT: Fetch clause starting at 22:
; EG-NEXT: VTX_READ_128 T20.XYZW, T19.X, 0, #1
; EG-NEXT: Fetch clause starting at 24:
; EG-NEXT: VTX_READ_128 T38.XYZW, T19.X, 48, #1
; EG-NEXT: VTX_READ_128 T39.XYZW, T19.X, 32, #1
; EG-NEXT: VTX_READ_128 T40.XYZW, T19.X, 16, #1
; EG-NEXT: ALU clause starting at 30:
; EG-NEXT: MOV * T19.X, KC0[2].Z,
; EG-NEXT: ALU clause starting at 31:
; EG-NEXT: LSHR T21.X, KC0[2].Y, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: LSHR T22.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 32(4.484155e-44)
; EG-NEXT: LSHR T23.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 48(6.726233e-44)
; EG-NEXT: LSHR T24.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 64(8.968310e-44)
; EG-NEXT: LSHR T25.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 80(1.121039e-43)
; EG-NEXT: LSHR T26.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 96(1.345247e-43)
; EG-NEXT: LSHR T27.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 112(1.569454e-43)
; EG-NEXT: LSHR T28.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 128(1.793662e-43)
; EG-NEXT: LSHR T29.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 144(2.017870e-43)
; EG-NEXT: LSHR T30.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 160(2.242078e-43)
; EG-NEXT: LSHR T31.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 176(2.466285e-43)
; EG-NEXT: LSHR T32.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 192(2.690493e-43)
; EG-NEXT: LSHR T33.X, PV.W, literal.x,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.y,
; EG-NEXT: 2(2.802597e-45), 208(2.914701e-43)
; EG-NEXT: LSHR T34.X, PV.W, literal.x,
; EG-NEXT: ADD_INT T0.W, KC0[2].Y, literal.y,
; EG-NEXT: ASHR * T35.W, T20.X, literal.z,
; EG-NEXT: 2(2.802597e-45), 224(3.138909e-43)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: LSHR T36.X, PV.W, literal.x,
; EG-NEXT: ASHR T35.Z, T20.X, literal.y,
; EG-NEXT: ASHR * T37.W, T20.Y, literal.z,
; EG-NEXT: 2(2.802597e-45), 16(2.242078e-44)
; EG-NEXT: 31(4.344025e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T35.X, T20.X, 0.0, literal.x,
; EG-NEXT: ASHR * T37.Z, T20.Y, literal.x,
; EG-NEXT: 16(2.242078e-44), 0(0.000000e+00)
; EG-NEXT: BFE_INT T37.X, T20.Y, 0.0, literal.x,
; EG-NEXT: ASHR T35.Y, PV.X, literal.y,
; EG-NEXT: ASHR * T19.W, T20.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: ALU clause starting at 87:
; EG-NEXT: ASHR T19.Z, T20.Z, literal.x,
; EG-NEXT: ASHR * T41.W, T20.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T19.X, T20.Z, 0.0, literal.x,
; EG-NEXT: ASHR T37.Y, T37.X, literal.y,
; EG-NEXT: ASHR T41.Z, T20.W, literal.x,
; EG-NEXT: ASHR * T42.W, T40.X, literal.y, BS:VEC_120/SCL_212
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T41.X, T20.W, 0.0, literal.x,
; EG-NEXT: ASHR T19.Y, PV.X, literal.y,
; EG-NEXT: ASHR T42.Z, T40.X, literal.x,
; EG-NEXT: ASHR * T20.W, T40.Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T42.X, T40.X, 0.0, literal.x,
; EG-NEXT: ASHR T41.Y, PV.X, literal.y,
; EG-NEXT: ASHR T20.Z, T40.Y, literal.x,
; EG-NEXT: ASHR * T43.W, T40.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T20.X, T40.Y, 0.0, literal.x,
; EG-NEXT: ASHR T42.Y, PV.X, literal.y,
; EG-NEXT: ASHR T43.Z, T40.Z, literal.x,
; EG-NEXT: ASHR * T44.W, T40.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T43.X, T40.Z, 0.0, literal.x,
; EG-NEXT: ASHR T20.Y, PV.X, literal.y,
; EG-NEXT: ASHR T44.Z, T40.W, literal.x,
; EG-NEXT: ASHR * T45.W, T39.X, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T44.X, T40.W, 0.0, literal.x,
; EG-NEXT: ASHR T43.Y, PV.X, literal.y,
; EG-NEXT: ASHR T45.Z, T39.X, literal.x,
; EG-NEXT: ASHR * T40.W, T39.Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T45.X, T39.X, 0.0, literal.x,
; EG-NEXT: ASHR T44.Y, PV.X, literal.y,
; EG-NEXT: ASHR T40.Z, T39.Y, literal.x,
; EG-NEXT: ASHR * T46.W, T39.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T40.X, T39.Y, 0.0, literal.x,
; EG-NEXT: ASHR T45.Y, PV.X, literal.y,
; EG-NEXT: ASHR T46.Z, T39.Z, literal.x,
; EG-NEXT: ASHR * T47.W, T39.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T46.X, T39.Z, 0.0, literal.x,
; EG-NEXT: ASHR T40.Y, PV.X, literal.y,
; EG-NEXT: ASHR T47.Z, T39.W, literal.x,
; EG-NEXT: ASHR * T48.W, T38.X, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T47.X, T39.W, 0.0, literal.x,
; EG-NEXT: ASHR T46.Y, PV.X, literal.y,
; EG-NEXT: ASHR T48.Z, T38.X, literal.x,
; EG-NEXT: ASHR * T39.W, T38.Y, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T48.X, T38.X, 0.0, literal.x,
; EG-NEXT: ASHR T47.Y, PV.X, literal.y,
; EG-NEXT: ASHR T39.Z, T38.Y, literal.x,
; EG-NEXT: ASHR * T49.W, T38.Z, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T39.X, T38.Y, 0.0, literal.x,
; EG-NEXT: ASHR T48.Y, PV.X, literal.y,
; EG-NEXT: ASHR T49.Z, T38.Z, literal.x,
; EG-NEXT: ASHR * T50.W, T38.W, literal.y,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T49.X, T38.Z, 0.0, literal.x,
; EG-NEXT: ASHR T39.Y, PV.X, literal.y,
; EG-NEXT: ASHR * T50.Z, T38.W, literal.x,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: BFE_INT T50.X, T38.W, 0.0, literal.x,
; EG-NEXT: ASHR T49.Y, PV.X, literal.y,
; EG-NEXT: ADD_INT * T0.W, KC0[2].Y, literal.z,
; EG-NEXT: 16(2.242078e-44), 31(4.344025e-44)
; EG-NEXT: 240(3.363116e-43), 0(0.000000e+00)
; EG-NEXT: LSHR T38.X, PV.W, literal.x,
; EG-NEXT: ASHR * T50.Y, PV.X, literal.y,
; EG-NEXT: 2(2.802597e-45), 31(4.344025e-44)
;
; GFX12-LABEL: constant_sextload_v32i16_to_v32i64:
; GFX12: ; %bb.0:
; GFX12-NEXT: s_load_b128 s[16:19], s[2:3], 0x24
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_load_b512 s[0:15], s[18:19], 0x0
; GFX12-NEXT: s_wait_kmcnt 0x0
; GFX12-NEXT: s_lshr_b32 s28, s2, 16
; GFX12-NEXT: s_lshr_b32 s42, s5, 16
; GFX12-NEXT: s_lshr_b32 s52, s8, 16
; GFX12-NEXT: s_mov_b32 s60, s11
; GFX12-NEXT: s_lshr_b32 s22, s0, 16
; GFX12-NEXT: s_mov_b32 s24, s1
; GFX12-NEXT: s_lshr_b32 s26, s1, 16
; GFX12-NEXT: s_mov_b32 s30, s3
; GFX12-NEXT: s_lshr_b32 s36, s3, 16
; GFX12-NEXT: s_lshr_b32 s38, s4, 16
; GFX12-NEXT: s_mov_b32 s40, s5
; GFX12-NEXT: s_lshr_b32 s46, s6, 16
; GFX12-NEXT: s_mov_b32 s48, s7
; GFX12-NEXT: s_lshr_b32 s50, s7, 16
; GFX12-NEXT: s_mov_b32 s54, s9
; GFX12-NEXT: s_lshr_b32 s56, s9, 16
; GFX12-NEXT: s_bfe_i64 s[44:45], s[10:11], 0x100000
; GFX12-NEXT: s_lshr_b32 s58, s10, 16
; GFX12-NEXT: s_lshr_b32 s62, s11, 16
; GFX12-NEXT: s_bfe_i64 s[10:11], s[28:29], 0x100000
; GFX12-NEXT: s_bfe_i64 s[28:29], s[42:43], 0x100000
; GFX12-NEXT: s_bfe_i64 s[42:43], s[52:53], 0x100000
; GFX12-NEXT: s_bfe_i64 s[52:53], s[60:61], 0x100000
; GFX12-NEXT: s_lshr_b32 s60, s14, 16
; GFX12-NEXT: s_bfe_i64 s[64:65], s[14:15], 0x100000
; GFX12-NEXT: s_mov_b32 s14, s15
; GFX12-NEXT: s_lshr_b32 s66, s15, 16
; GFX12-NEXT: s_bfe_i64 s[18:19], s[0:1], 0x100000
; GFX12-NEXT: s_bfe_i64 s[0:1], s[2:3], 0x100000
; GFX12-NEXT: s_bfe_i64 s[2:3], s[4:5], 0x100000
; GFX12-NEXT: s_bfe_i64 s[20:21], s[6:7], 0x100000
; GFX12-NEXT: s_bfe_i64 s[34:35], s[8:9], 0x100000
; GFX12-NEXT: s_bfe_i64 s[4:5], s[22:23], 0x100000
; GFX12-NEXT: s_bfe_i64 s[8:9], s[24:25], 0x100000
; GFX12-NEXT: s_bfe_i64 s[6:7], s[26:27], 0x100000
; GFX12-NEXT: s_bfe_i64 s[24:25], s[30:31], 0x100000
; GFX12-NEXT: s_bfe_i64 s[22:23], s[36:37], 0x100000
; GFX12-NEXT: s_bfe_i64 s[26:27], s[38:39], 0x100000
; GFX12-NEXT: s_bfe_i64 s[30:31], s[40:41], 0x100000
; GFX12-NEXT: s_bfe_i64 s[36:37], s[46:47], 0x100000
; GFX12-NEXT: s_bfe_i64 s[40:41], s[48:49], 0x100000
; GFX12-NEXT: s_bfe_i64 s[38:39], s[50:51], 0x100000
; GFX12-NEXT: s_bfe_i64 s[46:47], s[54:55], 0x100000
; GFX12-NEXT: s_bfe_i64 s[48:49], s[56:57], 0x100000
; GFX12-NEXT: s_bfe_i64 s[50:51], s[58:59], 0x100000
; GFX12-NEXT: s_lshr_b32 s54, s12, 16
; GFX12-NEXT: s_bfe_i64 s[56:57], s[12:13], 0x100000
; GFX12-NEXT: s_mov_b32 s12, s13
; GFX12-NEXT: s_lshr_b32 s58, s13, 16
; GFX12-NEXT: s_bfe_i64 s[14:15], s[14:15], 0x100000
; GFX12-NEXT: s_bfe_i64 s[66:67], s[66:67], 0x100000
; GFX12-NEXT: s_bfe_i64 s[60:61], s[60:61], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v24, 0 :: v_dual_mov_b32 v1, s15
; GFX12-NEXT: s_bfe_i64 s[12:13], s[12:13], 0x100000
; GFX12-NEXT: s_bfe_i64 s[58:59], s[58:59], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v0, s14 :: v_dual_mov_b32 v3, s67
; GFX12-NEXT: v_dual_mov_b32 v2, s66 :: v_dual_mov_b32 v5, s65
; GFX12-NEXT: s_bfe_i64 s[54:55], s[54:55], 0x100000
; GFX12-NEXT: v_dual_mov_b32 v4, s64 :: v_dual_mov_b32 v7, s61
; GFX12-NEXT: v_dual_mov_b32 v6, s60 :: v_dual_mov_b32 v9, s13
; GFX12-NEXT: v_dual_mov_b32 v8, s12 :: v_dual_mov_b32 v11, s59
; GFX12-NEXT: v_dual_mov_b32 v10, s58 :: v_dual_mov_b32 v13, s57
; GFX12-NEXT: v_dual_mov_b32 v12, s56 :: v_dual_mov_b32 v15, s55
; GFX12-NEXT: v_mov_b32_e32 v14, s54
; GFX12-NEXT: s_bfe_i64 s[12:13], s[62:63], 0x100000
; GFX12-NEXT: s_clause 0x3
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:240
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:224
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[16:17] offset:208
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[16:17] offset:192
; GFX12-NEXT: v_dual_mov_b32 v1, s53 :: v_dual_mov_b32 v0, s52
; GFX12-NEXT: v_dual_mov_b32 v3, s13 :: v_dual_mov_b32 v2, s12
; GFX12-NEXT: v_dual_mov_b32 v5, s45 :: v_dual_mov_b32 v4, s44
; GFX12-NEXT: v_dual_mov_b32 v7, s51 :: v_dual_mov_b32 v6, s50
; GFX12-NEXT: v_dual_mov_b32 v9, s47 :: v_dual_mov_b32 v8, s46
; GFX12-NEXT: v_dual_mov_b32 v11, s49 :: v_dual_mov_b32 v10, s48
; GFX12-NEXT: v_dual_mov_b32 v13, s35 :: v_dual_mov_b32 v12, s34
; GFX12-NEXT: v_dual_mov_b32 v15, s43 :: v_dual_mov_b32 v14, s42
; GFX12-NEXT: v_dual_mov_b32 v17, s41 :: v_dual_mov_b32 v16, s40
; GFX12-NEXT: v_dual_mov_b32 v19, s39 :: v_dual_mov_b32 v18, s38
; GFX12-NEXT: v_dual_mov_b32 v21, s21 :: v_dual_mov_b32 v20, s20
; GFX12-NEXT: v_dual_mov_b32 v23, s37 :: v_dual_mov_b32 v22, s36
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:176
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:160
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[16:17] offset:144
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[16:17] offset:128
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[16:17] offset:112
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[16:17] offset:96
; GFX12-NEXT: v_dual_mov_b32 v1, s31 :: v_dual_mov_b32 v0, s30
; GFX12-NEXT: v_dual_mov_b32 v3, s29 :: v_dual_mov_b32 v2, s28
; GFX12-NEXT: v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v4, s2
; GFX12-NEXT: v_dual_mov_b32 v7, s27 :: v_dual_mov_b32 v6, s26
; GFX12-NEXT: v_dual_mov_b32 v9, s25 :: v_dual_mov_b32 v8, s24
; GFX12-NEXT: v_dual_mov_b32 v11, s23 :: v_dual_mov_b32 v10, s22
; GFX12-NEXT: v_dual_mov_b32 v13, s1 :: v_dual_mov_b32 v12, s0
; GFX12-NEXT: v_dual_mov_b32 v15, s11 :: v_dual_mov_b32 v14, s10
; GFX12-NEXT: v_dual_mov_b32 v17, s9 :: v_dual_mov_b32 v16, s8
; GFX12-NEXT: v_dual_mov_b32 v19, s7 :: v_dual_mov_b32 v18, s6
; GFX12-NEXT: v_dual_mov_b32 v21, s19 :: v_dual_mov_b32 v20, s18
; GFX12-NEXT: v_dual_mov_b32 v23, s5 :: v_dual_mov_b32 v22, s4
; GFX12-NEXT: s_clause 0x5
; GFX12-NEXT: global_store_b128 v24, v[0:3], s[16:17] offset:80
; GFX12-NEXT: global_store_b128 v24, v[4:7], s[16:17] offset:64
; GFX12-NEXT: global_store_b128 v24, v[8:11], s[16:17] offset:48
; GFX12-NEXT: global_store_b128 v24, v[12:15], s[16:17] offset:32
; GFX12-NEXT: global_store_b128 v24, v[16:19], s[16:17] offset:16
; GFX12-NEXT: global_store_b128 v24, v[20:23], s[16:17]
; GFX12-NEXT: s_nop 0
; GFX12-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
; GFX12-NEXT: s_endpgm
%load = load <32 x i16>, ptr addrspace(4) %in
%ext = sext <32 x i16> %load to <32 x i64>
store <32 x i64> %ext, ptr addrspace(1) %out
ret void
}
; These trigger undefined register machine verifier errors
; define amdgpu_kernel void @constant_zextload_v64i16_to_v64i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; %load = load <64 x i16>, ptr addrspace(4) %in
; %ext = zext <64 x i16> %load to <64 x i64>
; store <64 x i64> %ext, ptr addrspace(1) %out
; ret void
; }
; define amdgpu_kernel void @constant_sextload_v64i16_to_v64i64(ptr addrspace(1) %out, ptr addrspace(4) %in) #0 {
; %load = load <64 x i16>, ptr addrspace(4) %in
; %ext = sext <64 x i16> %load to <64 x i64>
; store <64 x i64> %ext, ptr addrspace(1) %out
; ret void
; }
attributes #0 = { nounwind }
Codebase Browser
llvm