global_atomics.ll | Explore in Territory

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -amdgpu-atomic-optimizer-strategy=None -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI %s
; RUN: llc -mtriple=amdgcn -mcpu=tonga -amdgpu-atomic-optimizer-strategy=None -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -amdgpu-atomic-optimizer-strategy=None -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX9 %s

define amdgpu_kernel void @atomic_add_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_add_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_max_neg_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_add_i32_max_neg_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffff000
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_max_neg_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffff000
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_max_neg_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1] offset:-4096
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 -1024
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_soffset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_add_i32_soffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_mov_b32 s5, 0x8ca0
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_add v0, off, s[0:3], s5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_soffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s5, 0x8ca0
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_add v0, off, s[0:3], s5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_soffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0x8000
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1] offset:3232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 9000
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_huge_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_add_i32_huge_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xdeac
; SI-NEXT:    v_mov_b32_e32 v1, 0xabcd
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_huge_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xdeac
; VI-NEXT:    s_addc_u32 s1, s1, 0xabcd
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_huge_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_add_u32 s0, s0, 0xdeac
; GFX9-NEXT:    s_addc_u32 s1, s1, 0xabcd
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 47224239175595

  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_add_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_add v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_add v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_add v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_add_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_add_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_add v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_add_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile add ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_add_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_add v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_add v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile add ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_add_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_add v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile add ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_add_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_add v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_add_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_add v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_add_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_add v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile add ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_and_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_and v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_and_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_and v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_and v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_and v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_and_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_and v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_and v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_and v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_and_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_and v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_and v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_and v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_and_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_and v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile and ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_and_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_and v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_and v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile and ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_and_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_and v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_and v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_and v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile and ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_and_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_and v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_and_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_and v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_and_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_and v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile and ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_sub_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_sub v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_sub_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_sub v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_sub v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_sub v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_sub_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_sub v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_sub v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_sub v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_sub_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_sub v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_sub v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_sub v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_sub_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_sub v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile sub ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_sub_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_sub v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_sub v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile sub ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_sub_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_sub v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_sub v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_sub v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile sub ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_sub_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_sub v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_sub_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_sub v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_sub_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_sub v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile sub ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_max_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_max_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_smax v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_max_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smax v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_max_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_max_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile max ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_max_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_smax v0, off, s[0:3], 0 glc
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile max ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_max_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smax v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_smax v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile max ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_max_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_smax v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_max_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smax v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_max_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_smax v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile max ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_umax_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_umax_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_umax v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_umax_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umax v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_umax_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_umax_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umax ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_umax_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_umax v0, off, s[0:3], 0 glc
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umax ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_umax_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umax v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_umax v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umax ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_umax_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_umax v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umax_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umax v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umax_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_umax v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umax ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_min_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_min_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_smin v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_min_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smin v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_min_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_min_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile min ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_min_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_smin v0, off, s[0:3], 0 glc
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile min ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_min_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smin v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_smin v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile min ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_min_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_smin v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_min_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_smin v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_min_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_smin v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile min ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_umin_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umin ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_umin_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_umin v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umin ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_umin_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umin v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umin ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_umin_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umin ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_umin_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umin ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_umin_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_umin v0, off, s[0:3], 0 glc
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umin ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_umin_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umin v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_umin v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umin ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_umin_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_umin v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_umin_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_umin v0, v[0:1], v2 glc
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_umin_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_umin v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umin ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_or_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_or v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile or ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_or_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_or v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_or v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_or v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile or ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_or_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_or v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_or v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_or v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile or ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_or_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_or v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_or v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_or v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile or ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_or_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_or v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile or ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_or_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_or v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_or v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile or ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_or_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_or v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_or v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_or v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile or ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_or_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_or v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_or_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_or v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_or_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_or v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile or ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_xchg_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile xchg ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_f32_offset(ptr addrspace(1) %out, float %in) {
; SI-LABEL: atomic_xchg_f32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_f32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_f32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr float, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile xchg ptr addrspace(1) %gep, float %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_xchg_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_swap v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_swap v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile xchg ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_xchg_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_swap v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_swap v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile xchg ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_xchg_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_swap v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_swap v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile xchg ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_xchg_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_xchg_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_swap v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_swap v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xchg ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_xchg_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_swap v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_swap v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile xchg ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_xchg_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_swap v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xchg_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_swap v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xchg_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_swap v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile xchg ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_offset(ptr addrspace(1) %out, i32 %in, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s0
; SI-NEXT:    s_mov_b32 s5, s1
; SI-NEXT:    v_mov_b32_e32 v0, s2
; SI-NEXT:    v_mov_b32_e32 v1, s3
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s2
; VI-NEXT:    s_mov_b32 s4, s0
; VI-NEXT:    s_mov_b32 s5, s1
; VI-NEXT:    v_mov_b32_e32 v1, s3
; VI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v0, s6
; GFX9-NEXT:    v_mov_b32_e32 v1, s7
; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[4:5] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-NEXT:    v_mov_b32_e32 v1, s1
; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v2, v0, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %extract0 = extractvalue { i32, i1 } %val, 0
  store i32 %extract0, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s7, s[2:3], 0xf
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, s6
; SI-NEXT:    v_mov_b32_e32 v1, s7
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    v_mov_b32_e32 v3, s5
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], v[2:3], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dword s6, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x3c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    v_mov_b32_e32 v0, s6
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v3, s1
; VI-NEXT:    v_mov_b32_e32 v1, s2
; VI-NEXT:    v_mov_b32_e32 v2, s0
; VI-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_load_dword s2, s[2:3], 0x3c
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v0, s6
; GFX9-NEXT:    v_mov_b32_e32 v1, s2
; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s10, s[2:3], 0xd
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0x11
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s10
; SI-NEXT:    v_mov_b32_e32 v1, s2
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_mov_b32_e32 v3, s9
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], v[2:3], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x44
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v3, s1
; VI-NEXT:    v_mov_b32_e32 v1, s2
; VI-NEXT:    v_mov_b32_e32 v2, s0
; VI-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    s_load_dword s9, s[2:3], 0x44
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v0, s8
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v2, v0, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = cmpxchg volatile ptr addrspace(1) %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %extract0 = extractvalue { i32, i1 } %val, 0
  store i32 %extract0, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32(ptr addrspace(1) %out, i32 %in, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s0
; SI-NEXT:    s_mov_b32 s5, s1
; SI-NEXT:    v_mov_b32_e32 v0, s2
; SI-NEXT:    v_mov_b32_e32 v1, s3
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s2
; VI-NEXT:    s_mov_b32 s4, s0
; VI-NEXT:    s_mov_b32 s5, s1
; VI-NEXT:    v_mov_b32_e32 v1, s3
; VI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[4:7], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v0, s6
; GFX9-NEXT:    v_mov_b32_e32 v1, s7
; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = cmpxchg volatile ptr addrspace(1) %out, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    buffer_atomic_cmpswap v[0:1], off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-NEXT:    v_mov_b32_e32 v1, s1
; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v2, v0, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = cmpxchg volatile ptr addrspace(1) %out, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %extract0 = extractvalue { i32, i1 } %val, 0
  store i32 %extract0, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s7, s[2:3], 0xf
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, s6
; SI-NEXT:    v_mov_b32_e32 v1, s7
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    v_mov_b32_e32 v3, s5
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], v[2:3], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dword s6, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x3c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    v_mov_b32_e32 v0, s6
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v3, s1
; VI-NEXT:    v_mov_b32_e32 v1, s2
; VI-NEXT:    v_mov_b32_e32 v2, s0
; VI-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_load_dword s2, s[2:3], 0x3c
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v0, s6
; GFX9-NEXT:    v_mov_b32_e32 v1, s2
; GFX9-NEXT:    global_atomic_cmpswap v2, v[0:1], s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = cmpxchg volatile ptr addrspace(1) %ptr, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index, i32 %old) {
; SI-LABEL: atomic_cmpxchg_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s10, s[2:3], 0xd
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0x11
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s10
; SI-NEXT:    v_mov_b32_e32 v1, s2
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_mov_b32_e32 v3, s9
; SI-NEXT:    buffer_atomic_cmpswap v[0:1], v[2:3], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_cmpxchg_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x44
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v3, s1
; VI-NEXT:    v_mov_b32_e32 v1, s2
; VI-NEXT:    v_mov_b32_e32 v2, s0
; VI-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_cmpxchg_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    s_load_dword s9, s[2:3], 0x44
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v0, s8
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    global_atomic_cmpswap v0, v2, v[0:1], s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v2, v0, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = cmpxchg volatile ptr addrspace(1) %ptr, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %extract0 = extractvalue { i32, i1 } %val, 0
  store i32 %extract0, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_xor_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_xor v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile xor ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_xor_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_xor v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_xor v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_xor v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile xor ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_xor_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_xor v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_xor v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_xor v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile xor ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_xor_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_xor v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_xor v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_xor v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile xor ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_xor_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_xor v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xor ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_xor_i32_ret:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s4
; SI-NEXT:    s_mov_b32 s1, s5
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_ret:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s8, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s0, s4
; VI-NEXT:    s_mov_b32 s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s8
; VI-NEXT:    buffer_atomic_xor v0, off, s[0:3], 0 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_ret:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_xor v1, v0, v1, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xor ptr addrspace(1) %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_xor_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_xor v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_xor v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_xor v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile xor ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_xor_i32_ret_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_xor v2, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_xor_i32_ret_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_xor v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_xor_i32_ret_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_xor v1, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile xor ptr addrspace(1) %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_load_i32_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %in, i64 4
  %val = load atomic i32, ptr addrspace(1) %gep  seq_cst, align 4
  store i32 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32_negoffset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i32_negoffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffffe00
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    buffer_load_dword v0, v[0:1], s[0:3], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i32_negoffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffffe00
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i32_negoffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] offset:-512 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %in, i64 -128
  %val = load atomic i32, ptr addrspace(1) %gep  seq_cst, align 4
  store i32 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_f32_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_f32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_f32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_f32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr float, ptr addrspace(1) %in, i64 4
  %val = load atomic float, ptr addrspace(1) %gep  seq_cst, align 4
  store float %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s0
; SI-NEXT:    s_mov_b32 s5, s1
; SI-NEXT:    buffer_load_dword v0, off, s[4:7], 0 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_dword v1, v0, s[4:5] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %val = load atomic i32, ptr addrspace(1) %in syncscope("agent") seq_cst, align 4
  store i32 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32_addr64_offset(ptr addrspace(1) %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_load_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; VI-NEXT:    s_add_u32 s0, s0, s4
; VI-NEXT:    s_addc_u32 s1, s1, s5
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    global_load_dword v1, v0, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %in, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = load atomic i32, ptr addrspace(1) %gep seq_cst, align 4
  store i32 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32_addr64(ptr addrspace(1) %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_load_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; VI-NEXT:    s_add_u32 s0, s0, s4
; VI-NEXT:    s_addc_u32 s1, s1, s5
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    global_load_dword v1, v0, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %in, i64 %index
  %val = load atomic i32, ptr addrspace(1) %ptr seq_cst, align 4
  store i32 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_f32_addr64_offset(ptr addrspace(1) %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_load_f32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_f32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; VI-NEXT:    s_add_u32 s0, s0, s4
; VI-NEXT:    s_addc_u32 s1, s1, s5
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_dword v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_f32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    global_load_dword v1, v0, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr float, ptr addrspace(1) %in, i64 %index
  %gep = getelementptr float, ptr addrspace(1) %ptr, i64 4
  %val = load atomic float, ptr addrspace(1) %gep seq_cst, align 4
  store float %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_store_i32_offset(i32 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  store atomic i32 %in, ptr addrspace(1) %gep  seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32(i32 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  store atomic i32 %in, ptr addrspace(1) %out seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32(float %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_f32:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_f32:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_f32:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  store atomic float %in, ptr addrspace(1) %out seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32_addr64_offset(i32 %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_store_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0xb
; SI-NEXT:    s_load_dword s2, s[2:3], 0x9
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s0
; SI-NEXT:    v_mov_b32_e32 v1, s1
; SI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  store atomic i32 %in, ptr addrspace(1) %gep seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32_addr64_offset(float %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_store_f32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0xb
; SI-NEXT:    s_load_dword s2, s[2:3], 0x9
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s0
; SI-NEXT:    v_mov_b32_e32 v1, s1
; SI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_f32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_f32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr float, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr float, ptr addrspace(1) %ptr, i64 4
  store atomic float %in, ptr addrspace(1) %gep seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32_addr64(i32 %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_store_i32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0xb
; SI-NEXT:    s_load_dword s8, s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_mov_b32_e32 v0, s6
; SI-NEXT:    v_mov_b32_e32 v1, s7
; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  store atomic i32 %in, ptr addrspace(1) %ptr seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32_addr64(float %in, ptr addrspace(1) %out, i64 %index) {
; SI-LABEL: atomic_store_f32_addr64:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0xb
; SI-NEXT:    s_load_dword s8, s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[6:7], s[6:7], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b64 s[0:1], s[4:5]
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_mov_b32_e32 v0, s6
; SI-NEXT:    v_mov_b32_e32 v1, s7
; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_f32_addr64:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_dword v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_f32_addr64:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x2c
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[6:7], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr float, ptr addrspace(1) %out, i64 %index
  store atomic float %in, ptr addrspace(1) %ptr seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_load_i8_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i8_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_ubyte v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i8_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    s_mov_b32 s2, s6
; VI-NEXT:    s_mov_b32 s3, s7
; VI-NEXT:    buffer_load_ubyte v0, off, s[0:3], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i8_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ubyte v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_byte v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i8, ptr addrspace(1) %in, i64 16
  %val = load atomic i8, ptr addrspace(1) %gep  seq_cst, align 1
  store i8 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i8_negoffset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i8_negoffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffffe00
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    buffer_load_ubyte v0, v[0:1], s[0:3], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i8_negoffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffffe00
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_ubyte v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i8_negoffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ubyte v1, v0, s[4:5] offset:-512 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_byte v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i8, ptr addrspace(1) %in, i64 -512
  %val = load atomic i8, ptr addrspace(1) %gep  seq_cst, align 1
  store i8 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_store_i8_offset(i8 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i8_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_byte v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i8_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_byte v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i8_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_byte v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i8, ptr addrspace(1) %out, i64 16
  store atomic i8 %in, ptr addrspace(1) %gep  seq_cst, align 1
  ret void
}

define amdgpu_kernel void @atomic_store_i8(i8 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i8:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_byte v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i8:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_byte v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i8:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_byte v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  store atomic i8 %in, ptr addrspace(1) %out seq_cst, align 1
  ret void
}

define amdgpu_kernel void @atomic_load_i16_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i16_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i16_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    s_mov_b32 s2, s6
; VI-NEXT:    s_mov_b32 s3, s7
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i16_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i16, ptr addrspace(1) %in, i64 8
  %val = load atomic i16, ptr addrspace(1) %gep  seq_cst, align 2
  store i16 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_i16_negoffset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_i16_negoffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffffe00
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    buffer_load_ushort v0, v[0:1], s[0:3], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_i16_negoffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffffe00
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_ushort v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_i16_negoffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:-512 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i16, ptr addrspace(1) %in, i64 -256
  %val = load atomic i16, ptr addrspace(1) %gep  seq_cst, align 2
  store i16 %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_store_i16_offset(i16 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i16_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i16_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i16_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i16, ptr addrspace(1) %out, i64 8
  store atomic i16 %in, ptr addrspace(1) %gep  seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_i16(i16 %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_i16:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_i16:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_i16:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  store atomic i16 %in, ptr addrspace(1) %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_f16_offset(half %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_f16_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_f16_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_f16_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr half, ptr addrspace(1) %out, i64 8
  store atomic half %in, ptr addrspace(1) %gep  seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_f16(half %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_f16:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_f16:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_f16:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
entry:
  store atomic half %in, ptr addrspace(1) %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_bf16_offset(bfloat %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_bf16_offset:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_bf16_offset:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_bf16_offset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_endpgm
  %gep = getelementptr bfloat, ptr addrspace(1) %out, i64 8
  store atomic bfloat %in, ptr addrspace(1) %gep seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_bf16(bfloat %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_store_bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dword s4, s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_store_bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; VI-NEXT:    s_load_dword s2, s[2:3], 0x24
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_store_short v[0:1], v2
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_store_bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x24
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_store_short v0, v1, s[0:1]
; GFX9-NEXT:    s_endpgm
  store atomic bfloat %in, ptr addrspace(1) %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_inc_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_inc v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_inc v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_inc v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_max_neg_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_inc_i32_max_neg_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffff000
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_inc v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_max_neg_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffff000
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_inc v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_max_neg_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_inc v0, v1, s[0:1] offset:-4096
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 -1024
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_soffset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_inc_i32_soffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_mov_b32 s5, 0x8ca0
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_inc v0, off, s[0:3], s5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_soffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s5, 0x8ca0
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_inc v0, off, s[0:3], s5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_soffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0x8000
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_inc v0, v1, s[0:1] offset:3232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 9000
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_huge_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_inc_i32_huge_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xdeac
; SI-NEXT:    v_mov_b32_e32 v1, 0xabcd
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_inc v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_huge_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xdeac
; VI-NEXT:    s_addc_u32 s1, s1, 0xabcd
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_inc v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_huge_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_add_u32 s0, s0, 0xdeac
; GFX9-NEXT:    s_addc_u32 s1, s1, 0xabcd
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_inc v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 47224239175595
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_inc_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_inc v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_inc v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_inc v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_inc_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_inc v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_inc v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_inc v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_inc_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_inc v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_inc_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_inc v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_inc_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_inc v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile uinc_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_dec_i32_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_dec v0, off, s[0:3], 0 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_dec v0, off, s[0:3], 0 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_dec v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_max_neg_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_dec_i32_max_neg_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffff000
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_dec v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_max_neg_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffff000
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_dec v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_max_neg_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_dec v0, v1, s[0:1] offset:-4096
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 -1024
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_soffset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_dec_i32_soffset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_mov_b32 s5, 0x8ca0
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    buffer_atomic_dec v0, off, s[0:3], s5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_soffset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dword s4, s[2:3], 0x2c
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s5, 0x8ca0
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    buffer_atomic_dec v0, off, s[0:3], s5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_soffset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0x8000
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_dec v0, v1, s[0:1] offset:3232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 9000
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_huge_offset(ptr addrspace(1) %out, i32 %in) {
; SI-LABEL: atomic_dec_i32_huge_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_load_dword s4, s[2:3], 0xb
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v0, 0xdeac
; SI-NEXT:    v_mov_b32_e32 v1, 0xabcd
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_atomic_dec v2, v[0:1], s[0:3], 0 addr64
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_huge_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xdeac
; VI-NEXT:    s_addc_u32 s1, s1, 0xabcd
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_dec v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_huge_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s4, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_add_u32 s0, s0, 0xdeac
; GFX9-NEXT:    s_addc_u32 s1, s1, 0xabcd
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    global_atomic_dec v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 47224239175595
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; SI-LABEL: atomic_dec_i32_ret_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dword s8, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_mov_b32 s6, s2
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    buffer_atomic_dec v0, off, s[4:7], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_ret_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s0, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s11, 0xf000
; VI-NEXT:    s_mov_b32 s10, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s8, s6
; VI-NEXT:    s_mov_b32 s9, s7
; VI-NEXT:    s_mov_b32 s6, s10
; VI-NEXT:    s_mov_b32 s7, s11
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    buffer_atomic_dec v0, off, s[4:7], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_dword v0, off, s[8:11], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_ret_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dword s0, s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-NEXT:    global_atomic_dec v1, v0, v1, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; SI-LABEL: atomic_dec_i32_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0xd
; SI-NEXT:    s_load_dword s6, s[2:3], 0xb
; SI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_lshl_b64 s[4:5], s[4:5], 2
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    buffer_atomic_dec v2, v[0:1], s[0:3], 0 addr64 offset:16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; VI-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x2c
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_dec v[0:1], v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s6, s[2:3], 0x2c
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    global_atomic_dec v0, v1, s[0:1] offset:16
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; SI-LABEL: atomic_dec_i32_ret_addr64_offset:
; SI:       ; %bb.0: ; %entry
; SI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
; SI-NEXT:    s_load_dwordx2 s[8:9], s[2:3], 0xf
; SI-NEXT:    s_load_dword s2, s[2:3], 0xd
; SI-NEXT:    s_mov_b32 s3, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s0, s6
; SI-NEXT:    s_mov_b32 s1, s7
; SI-NEXT:    s_lshl_b64 s[8:9], s[8:9], 2
; SI-NEXT:    s_mov_b32 s6, 0
; SI-NEXT:    s_mov_b32 s7, s3
; SI-NEXT:    v_mov_b32_e32 v2, s2
; SI-NEXT:    v_mov_b32_e32 v0, s8
; SI-NEXT:    v_mov_b32_e32 v1, s9
; SI-NEXT:    buffer_atomic_dec v2, v[0:1], s[4:7], 0 addr64 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s2, -1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_dec_i32_ret_addr64_offset:
; VI:       ; %bb.0: ; %entry
; VI-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; VI-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; VI-NEXT:    s_load_dword s2, s[2:3], 0x34
; VI-NEXT:    s_mov_b32 s3, 0xf000
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; VI-NEXT:    s_add_u32 s0, s4, s0
; VI-NEXT:    s_addc_u32 s1, s5, s1
; VI-NEXT:    s_add_u32 s0, s0, 16
; VI-NEXT:    s_addc_u32 s1, s1, 0
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    v_mov_b32_e32 v2, s2
; VI-NEXT:    flat_atomic_dec v0, v[0:1], v2 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s2, -1
; VI-NEXT:    s_mov_b32 s0, s6
; VI-NEXT:    s_mov_b32 s1, s7
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_dec_i32_ret_addr64_offset:
; GFX9:       ; %bb.0: ; %entry
; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x3c
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    s_load_dword s8, s[2:3], 0x34
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX9-NEXT:    s_add_u32 s0, s4, s0
; GFX9-NEXT:    s_addc_u32 s1, s5, s1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    global_atomic_dec v1, v0, v1, s[0:1] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile udec_wrap ptr addrspace(1) %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_load_f16_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_f16_offset:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_f16_offset:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    s_mov_b32 s2, s6
; VI-NEXT:    s_mov_b32 s3, s7
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_f16_offset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
  %gep = getelementptr half, ptr addrspace(1) %in, i64 8
  %val = load atomic half, ptr addrspace(1) %gep  seq_cst, align 2
  store half %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_f16_negoffset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_f16_negoffset:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffffe00
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    buffer_load_ushort v0, v[0:1], s[0:3], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_f16_negoffset:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffffe00
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_ushort v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_f16_negoffset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:-512 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
  %gep = getelementptr half, ptr addrspace(1) %in, i64 -256
  %val = load atomic half, ptr addrspace(1) %gep  seq_cst, align 2
  store half %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_bf16_offset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_bf16_offset:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, s6
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_bf16_offset:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    s_mov_b32 s2, s6
; VI-NEXT:    s_mov_b32 s3, s7
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:16 glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_bf16_offset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:16 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
  %gep = getelementptr bfloat, ptr addrspace(1) %in, i64 8
  %val = load atomic bfloat, ptr addrspace(1) %gep  seq_cst, align 2
  store bfloat %val, ptr addrspace(1) %out
  ret void
}

define amdgpu_kernel void @atomic_load_bf16_negoffset(ptr addrspace(1) %in, ptr addrspace(1) %out) {
; SI-LABEL: atomic_load_bf16_negoffset:
; SI:       ; %bb.0:
; SI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x9
; SI-NEXT:    s_mov_b32 s7, 0xf000
; SI-NEXT:    s_waitcnt lgkmcnt(0)
; SI-NEXT:    s_mov_b32 s4, s2
; SI-NEXT:    s_mov_b32 s5, s3
; SI-NEXT:    s_mov_b32 s2, 0
; SI-NEXT:    s_mov_b32 s3, s7
; SI-NEXT:    v_mov_b32_e32 v0, 0xfffffe00
; SI-NEXT:    v_mov_b32_e32 v1, -1
; SI-NEXT:    buffer_load_ushort v0, v[0:1], s[0:3], 0 addr64 glc
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_wbinvl1
; SI-NEXT:    s_mov_b32 s6, -1
; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; SI-NEXT:    s_endpgm
;
; VI-LABEL: atomic_load_bf16_negoffset:
; VI:       ; %bb.0:
; VI-NEXT:    s_load_dwordx4 s[0:3], s[2:3], 0x24
; VI-NEXT:    s_mov_b32 s7, 0xf000
; VI-NEXT:    s_mov_b32 s6, -1
; VI-NEXT:    s_waitcnt lgkmcnt(0)
; VI-NEXT:    s_add_u32 s0, s0, 0xfffffe00
; VI-NEXT:    s_addc_u32 s1, s1, -1
; VI-NEXT:    v_mov_b32_e32 v0, s0
; VI-NEXT:    v_mov_b32_e32 v1, s1
; VI-NEXT:    flat_load_ushort v0, v[0:1] glc
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_wbinvl1_vol
; VI-NEXT:    s_mov_b32 s4, s2
; VI-NEXT:    s_mov_b32 s5, s3
; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
; VI-NEXT:    s_endpgm
;
; GFX9-LABEL: atomic_load_bf16_negoffset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
; GFX9-NEXT:    v_mov_b32_e32 v0, 0
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    global_load_ushort v1, v0, s[4:5] offset:-512 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1_vol
; GFX9-NEXT:    global_store_short v0, v1, s[6:7]
; GFX9-NEXT:    s_endpgm
  %gep = getelementptr bfloat, ptr addrspace(1) %in, i64 -256
  %val = load atomic bfloat, ptr addrspace(1) %gep  seq_cst, align 2
  store bfloat %val, ptr addrspace(1) %out
  ret void
}
llvm/llvm/test/CodeGen/AMDGPU/global_atomics.ll
