; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefixes=GCN1 %s
; RUN: llc -mtriple=amdgcn -mcpu=tonga < %s | FileCheck -check-prefixes=GCN2 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN3 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s

define amdgpu_kernel void @atomic_add_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_add_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_add v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw add ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_max_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_add_i32_max_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_max_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_max_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_add v[0:1], v2 offset:4092
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_max_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2 offset:4092
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1023
  %val = atomicrmw volatile add ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_max_offset_p1(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_add_i32_max_offset_p1:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_max_offset_p1:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_max_offset_p1:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
; GCN3-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_add v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_max_offset_p1:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_co_u32 v0, s0, 0x1000, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, s1, s0
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1024
  %val = atomicrmw volatile add ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_add_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_add v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_add_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile add ptr %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_add_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_add v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile add ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_add_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_add v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_add_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile add ptr %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_add_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_add v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile add ptr %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_add_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_add_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile add ptr %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_add_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_add v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_add v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_add v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_add_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile add ptr %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_add_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_add_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_add_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_add v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_add_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_add_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile add ptr %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_and_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_and v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_and v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_and v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_and_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile and ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_and_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_and v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_and_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile and ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_and_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_and v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_and v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_and v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_and_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile and ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_and_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_and v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_and_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile and ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_and_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_and v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_and v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_and v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_and_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile and ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_and_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_and_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile and ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_and_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_and v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_and v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_and v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_and_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile and ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_and_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_and_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_and_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_and v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_and_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_and_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile and ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_sub_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_sub v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_sub v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_sub v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_sub_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile sub ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_sub_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_sub v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_sub_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile sub ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_sub_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_sub v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_sub v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_sub v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_sub_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile sub ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_sub_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_sub v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_sub_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile sub ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_sub_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_sub v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_sub v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_sub v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_sub_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile sub ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_sub_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_sub_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile sub ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_sub_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_sub v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_sub v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_sub v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_sub_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile sub ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_sub_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_sub_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_sub_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_sub v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_sub_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_sub_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile sub ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_max_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_max_i32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile max ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_max_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_i32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile max ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_max_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smax v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_i32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile max ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_max_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_max_i32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile max ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_max_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_max_i32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile max ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_max_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_i32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile max ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_max_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smax v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_i32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile max ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_max_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_max_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_max_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_smax v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_max_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_max_i32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile max ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_umax_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_max_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile umax ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_umax_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile umax ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umax_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umax v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile umax ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umax_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_max_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile umax ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_umax_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_max_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umax ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_umax_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umax ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umax_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umax v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umax v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umax v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_max_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile umax ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umax_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umax_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umax_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_umax v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umax_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_max_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile umax ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_min_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_min_i32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile min ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_min_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_i32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile min ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_min_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smin v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_i32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile min ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_min_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_min_i32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile min ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_min_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_min_i32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile min ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_min_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_i32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile min ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_min_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_smin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_smin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_smin v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_i32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile min ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_min_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_min_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_min_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_smin v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_min_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_min_i32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile min ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_umin_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_min_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile umin ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_umin_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile umin ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umin_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umin v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile umin ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umin_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_min_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile umin ptr %gep, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_umin_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_min_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umin ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_umin_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile umin ptr %out, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umin_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_umin v[0:1], v2
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_umin v[0:1], v2
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_umin v[0:1], v2
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_min_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile umin ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_umin_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_umin_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    s_waitcnt vmcnt(0)
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_umin_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    s_waitcnt vmcnt(0)
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_umin_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_umin v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    s_waitcnt vmcnt(0)
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_umin_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_min_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile umin ptr %ptr, i32 %in syncscope("workgroup") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_or_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_or v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_or v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_or v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_or_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile or ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_or_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_or v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_or_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile or ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_or_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_or v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_or v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_or v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_or_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile or ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_or_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_or v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_or_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile or ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_or_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_or v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_or v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_or v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_or_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile or ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_or_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_or_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile or ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_or_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_or_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_or v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_or v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_or v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_or_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile or ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_or_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_or_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_or_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_or_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_or v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_or_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_or_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile or ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_xchg_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_swap v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_swap v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_swap v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_swap_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile xchg ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_f32_offset(ptr %out, float %in) {
; GCN1-LABEL: atomic_xchg_f32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_swap v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_f32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_swap v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_f32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_swap v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_f32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_swap_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds float, ptr %out, i32 4
  %val = atomicrmw volatile xchg ptr %gep, float %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_xchg_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_swap v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile xchg ptr %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xchg_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_swap v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_swap v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_swap v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_swap_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile xchg ptr %gep, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xchg_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_swap v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile xchg ptr %gep, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_xchg_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_swap v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_swap v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_swap v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_swap_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xchg ptr %out, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_xchg_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xchg ptr %out, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xchg_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_swap v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_swap v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_swap v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_swap_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile xchg ptr %ptr, i32 %in syncscope("agent") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_xchg_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xchg_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xchg_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xchg_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_swap v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xchg_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile xchg ptr %ptr, i32 %in syncscope("agent") seq_cst
  store i32 %val, ptr %out2
  ret void
}

; CMP_SWAP

define amdgpu_kernel void @atomic_cmpxchg_i32_offset(ptr %out, i32 %in, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v3, s3
; GCN1-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v3, s3
; GCN2-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s3
; GCN3-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3] offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = cmpxchg volatile ptr %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_offset(ptr %out, ptr %out2, i32 %in, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v3, s5
; GCN1-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v3, s5
; GCN2-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s7
; GCN3-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = cmpxchg volatile ptr %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %flag = extractvalue { i32, i1 } %val, 0
  store i32 %flag, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_addr64_offset(ptr %out, i32 %in, i64 %index, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dword s6, s[4:5], 0xb
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xf
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    v_mov_b32_e32 v0, s6
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v3, s1
; GCN1-NEXT:    v_mov_b32_e32 v1, s4
; GCN1-NEXT:    v_mov_b32_e32 v2, s0
; GCN1-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x3c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    v_mov_b32_e32 v0, s6
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v3, s1
; GCN2-NEXT:    v_mov_b32_e32 v1, s4
; GCN2-NEXT:    v_mov_b32_e32 v2, s0
; GCN2-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s7, s[4:5], 0x3c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    v_mov_b32_e32 v0, s6
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s1
; GCN3-NEXT:    v_mov_b32_e32 v1, s7
; GCN3-NEXT:    v_mov_b32_e32 v2, s0
; GCN3-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1] offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x3
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s6, s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x3c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s4
; GFX11-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v[2:3], v[0:1] offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val  = cmpxchg volatile ptr %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s9, s[4:5], 0x11
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    v_mov_b32_e32 v0, s8
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v3, s1
; GCN1-NEXT:    v_mov_b32_e32 v1, s9
; GCN1-NEXT:    v_mov_b32_e32 v2, s0
; GCN1-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s9, s[4:5], 0x44
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    v_mov_b32_e32 v0, s8
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v3, s1
; GCN2-NEXT:    v_mov_b32_e32 v1, s9
; GCN2-NEXT:    v_mov_b32_e32 v2, s0
; GCN2-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s9, s[4:5], 0x44
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    v_mov_b32_e32 v0, s8
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v3, s1
; GCN3-NEXT:    v_mov_b32_e32 v1, s9
; GCN3-NEXT:    v_mov_b32_e32 v2, s0
; GCN3-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x3
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_load_b32 s9, s[4:5], 0x44
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
; GFX11-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v2, v[2:3], v[0:1] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val  = cmpxchg volatile ptr %gep, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %flag = extractvalue { i32, i1 } %val, 0
  store i32 %flag, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32(ptr %out, i32 %in, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v3, s3
; GCN1-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v3, s3
; GCN2-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s3
; GCN3-NEXT:    flat_atomic_cmpswap v[0:1], v[2:3]
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3]
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = cmpxchg volatile ptr %out, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret(ptr %out, ptr %out2, i32 %in, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v3, s5
; GCN1-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v3, s5
; GCN2-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s7
; GCN3-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = cmpxchg volatile ptr %out, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %flag = extractvalue { i32, i1 } %val, 0
  store i32 %flag, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_addr64(ptr %out, i32 %in, i64 %index, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dword s6, s[4:5], 0xb
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xf
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    v_mov_b32_e32 v0, s6
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v3, s1
; GCN1-NEXT:    v_mov_b32_e32 v1, s4
; GCN1-NEXT:    v_mov_b32_e32 v2, s0
; GCN1-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x3c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    v_mov_b32_e32 v0, s6
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v3, s1
; GCN2-NEXT:    v_mov_b32_e32 v1, s4
; GCN2-NEXT:    v_mov_b32_e32 v2, s0
; GCN2-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s7, s[4:5], 0x3c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    v_mov_b32_e32 v0, s6
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v3, s1
; GCN3-NEXT:    v_mov_b32_e32 v1, s7
; GCN3-NEXT:    v_mov_b32_e32 v2, s0
; GCN3-NEXT:    flat_atomic_cmpswap v[2:3], v[0:1]
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x3
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s6, s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x3c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s4
; GFX11-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v[2:3], v[0:1]
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = cmpxchg volatile ptr %ptr, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  ret void
}

define amdgpu_kernel void @atomic_cmpxchg_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index, i32 %old) {
; GCN1-LABEL: atomic_cmpxchg_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s9, s[4:5], 0x11
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    v_mov_b32_e32 v0, s8
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v3, s1
; GCN1-NEXT:    v_mov_b32_e32 v1, s9
; GCN1-NEXT:    v_mov_b32_e32 v2, s0
; GCN1-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_cmpxchg_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s9, s[4:5], 0x44
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    v_mov_b32_e32 v0, s8
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v3, s1
; GCN2-NEXT:    v_mov_b32_e32 v1, s9
; GCN2-NEXT:    v_mov_b32_e32 v2, s0
; GCN2-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_cmpxchg_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s9, s[4:5], 0x44
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    v_mov_b32_e32 v0, s8
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v3, s1
; GCN3-NEXT:    v_mov_b32_e32 v1, s9
; GCN3-NEXT:    v_mov_b32_e32 v2, s0
; GCN3-NEXT:    flat_atomic_cmpswap v2, v[2:3], v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_cmpxchg_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x3
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_load_b32 s9, s[4:5], 0x44
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
; GFX11-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
; GFX11-NEXT:    flat_atomic_cmpswap_b32 v2, v[2:3], v[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = cmpxchg volatile ptr %ptr, i32 %old, i32 %in syncscope("agent") seq_cst seq_cst
  %flag = extractvalue { i32, i1 } %val, 0
  store i32 %flag, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_xor_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_xor v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_xor v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_xor v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_xor_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile xor ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_xor_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_xor v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_xor_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile xor ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_addr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xor_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_xor v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_xor v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_xor v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_xor_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile xor ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_addr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xor_i32_ret_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_ret_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_ret_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_xor v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_ret_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_xor_b32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile xor ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_xor_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_xor v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_xor v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_xor v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_xor_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xor ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_xor_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_xor_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile xor ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_addr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xor_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_xor v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_xor v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_xor v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_xor_b32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile xor ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_xor_i32_ret_addr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_xor_i32_ret_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_xor_i32_ret_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_xor_i32_ret_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_xor v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_xor_i32_ret_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_xor_b32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile xor ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_load_i32_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %in, i32 4
  %val = load atomic i32, ptr %gep  seq_cst, align 4
  store i32 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = load atomic i32, ptr %in seq_cst, align 4
  store i32 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32_addr64_offset(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %in, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = load atomic i32, ptr %gep seq_cst, align 4
  store i32 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i32_addr64(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %in, i64 %index
  %val = load atomic i32, ptr %ptr seq_cst, align 4
  store i32 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_store_i32_offset(i32 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_dword v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_store_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  store atomic i32 %in, ptr %gep  seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32(i32 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  store atomic i32 %in, ptr %out seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32_addr64_offset(i32 %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_i32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_dword v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_i32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    s_add_u32 s0, s0, s2
; GFX11-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_store_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  store atomic i32 %in, ptr %gep seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_i32_addr64(i32 %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_i32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_i32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    s_add_u32 s0, s0, s2
; GFX11-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  store atomic i32 %in, ptr %ptr seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_load_f32_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_f32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_f32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds float, ptr %in, i32 4
  %val = load atomic float, ptr %gep  seq_cst, align 4
  store float %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_f32(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_f32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_f32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = load atomic float, ptr %in seq_cst, align 4
  store float %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_f32_addr64_offset(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_f32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_f32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds float, ptr %in, i64 %index
  %gep = getelementptr inbounds float, ptr %ptr, i32 4
  %val = load atomic float, ptr %gep seq_cst, align 4
  store float %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_f32_addr64(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_f32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_dword v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_load_f32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_load_b32 v2, v[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds float, ptr %in, i64 %index
  %val = load atomic float, ptr %ptr seq_cst, align 4
  store float %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_store_f32_offset(float %in, ptr %out) {
; GCN1-LABEL: atomic_store_f32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_dword v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_f32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_store_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds float, ptr %out, i32 4
  store atomic float %in, ptr %gep  seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32(float %in, ptr %out) {
; GCN1-LABEL: atomic_store_f32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_f32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  store atomic float %in, ptr %out seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32_addr64_offset(float %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_f32_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f32_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f32_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_dword v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_f32_addr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    s_add_u32 s0, s0, s2
; GFX11-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_store_b32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds float, ptr %out, i64 %index
  %gep = getelementptr inbounds float, ptr %ptr, i32 4
  store atomic float %in, ptr %gep seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_store_f32_addr64(float %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_f32_addr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f32_addr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f32_addr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_store_f32_addr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[2:3], s[2:3], 2
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    s_add_u32 s0, s0, s2
; GFX11-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds float, ptr %out, i64 %index
  store atomic float %in, ptr %ptr seq_cst, align 4
  ret void
}

define amdgpu_kernel void @atomic_load_i8_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i8_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i8_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i8_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ubyte v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_byte v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i8_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_u8 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i8_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u8 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i8, ptr %in, i64 16
  %val = load atomic i8, ptr %gep  seq_cst, align 1
  store i8 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i8(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i8:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i8:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i8:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_byte v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i8:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_u8 v0, v[0:1] glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i8:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u8 v2, v[0:1] glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %val = load atomic i8, ptr %in seq_cst, align 1
  store i8 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i8_addr64_offset(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_i8_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i8_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ubyte v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i8_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_add_u32 s0, s0, s6
; GCN3-NEXT:    s_addc_u32 s1, s1, s7
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ubyte v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_byte v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i8_addr64_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, s4
; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    flat_load_d16_u8 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i8_addr64_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_load_b64 s[4:5], s[4:5], 0x34
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, s4
; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u8 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i8, ptr %in, i64 %index
  %gep = getelementptr inbounds i8, ptr %ptr, i64 16
  %val = load atomic i8, ptr %gep seq_cst, align 1
  store i8 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_store_i8_offset(i8 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i8_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i8_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i8_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_byte v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i8_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0 offset:16
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i8_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2 offset:16
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i8, ptr %out, i64 16
  store atomic i8 %in, ptr %gep  seq_cst, align 1
  ret void
}

define amdgpu_kernel void @atomic_store_i8(i8 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i8:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i8:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i8:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_byte v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i8:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i8:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  store atomic i8 %in, ptr %out seq_cst, align 1
  ret void
}

define amdgpu_kernel void @atomic_store_i8_addr64_offset(i8 %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_i8_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_byte v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i8_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_byte v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i8_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_byte v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i8_addr64_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, s2
; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    flat_store_b8 v[1:2], v0 offset:16
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i8_addr64_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, s2
; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-FAKE16-NEXT:    flat_store_b8 v[0:1], v2 offset:16
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i8, ptr %out, i64 %index
  %gep = getelementptr inbounds i8, ptr %ptr, i64 16
  store atomic i8 %in, ptr %gep seq_cst, align 1
  ret void
}

define amdgpu_kernel void @atomic_load_i16_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i16_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i16_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i16_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i16_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i16_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i16, ptr %in, i64 8
  %val = load atomic i16, ptr %gep  seq_cst, align 2
  store i16 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i16(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_i16:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i16:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i16:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i16:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i16:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %val = load atomic i16, ptr %in seq_cst, align 2
  store i16 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_i16_addr64_offset(ptr %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_load_i16_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 1
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_i16_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 1
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_i16_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 1
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_i16_addr64_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_lshl_b64 s[4:5], s[6:7], 1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, s4
; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_i16_addr64_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[6:7], s[4:5], 0x34
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_lshl_b64 s[4:5], s[6:7], 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, s4
; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i16, ptr %in, i64 %index
  %gep = getelementptr inbounds i16, ptr %ptr, i64 8
  %val = load atomic i16, ptr %gep seq_cst, align 2
  store i16 %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_store_i16_offset(i16 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i16_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i16_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i16_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i16_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0 offset:16
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i16_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2 offset:16
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i16, ptr %out, i64 8
  store atomic i16 %in, ptr %gep  seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_i16(i16 %in, ptr %out) {
; GCN1-LABEL: atomic_store_i16:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i16:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i16:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i16:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i16:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  store atomic i16 %in, ptr %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_i16_addr64_offset(i16 %in, ptr %out, i64 %index) {
; GCN1-LABEL: atomic_store_i16_addr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[2:3], s[2:3], 1
; GCN1-NEXT:    s_add_u32 s0, s0, s2
; GCN1-NEXT:    s_addc_u32 s1, s1, s3
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_i16_addr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[2:3], s[2:3], 1
; GCN2-NEXT:    s_add_u32 s0, s0, s2
; GCN2-NEXT:    s_addc_u32 s1, s1, s3
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_i16_addr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[2:3], s[2:3], 1
; GCN3-NEXT:    s_add_u32 s0, s0, s2
; GCN3-NEXT:    s_addc_u32 s1, s1, s3
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_store_short v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_i16_addr64_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_lshl_b64 s[2:3], s[2:3], 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s4
; GFX11-TRUE16-NEXT:    s_add_u32 s0, s0, s2
; GFX11-TRUE16-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0 offset:16
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_i16_addr64_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s4, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_lshl_b64 s[2:3], s[2:3], 1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-FAKE16-NEXT:    s_add_u32 s0, s0, s2
; GFX11-FAKE16-NEXT:    s_addc_u32 s1, s1, s3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2 offset:16
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i16, ptr %out, i64 %index
  %gep = getelementptr inbounds i16, ptr %ptr, i64 8
  store atomic i16 %in, ptr %gep seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_f16_offset(half %in, ptr %out) {
; GCN1-LABEL: atomic_store_f16_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f16_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f16_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2 offset:16
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_f16_offset:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0 offset:16
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_f16_offset:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2 offset:16
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds half, ptr %out, i64 8
  store atomic half %in, ptr %gep  seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_f16(half %in, ptr %out) {
; GCN1-LABEL: atomic_store_f16:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_f16:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_f16:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_f16:
; GFX11-TRUE16:       ; %bb.0: ; %entry
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_f16:
; GFX11-FAKE16:       ; %bb.0: ; %entry
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
entry:
  store atomic half %in, ptr %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_bf16_offset(bfloat %in, ptr %out) {
; GCN1-LABEL: atomic_store_bf16_offset:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_bf16_offset:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_bf16_offset:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_bf16_offset:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_bf16_offset:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  %gep = getelementptr inbounds bfloat, ptr %out, i64 8
  store atomic bfloat %in, ptr %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_store_bf16(bfloat %in, ptr %out) {
; GCN1-LABEL: atomic_store_bf16:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xb
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_store_bf16:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_store_bf16:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x2c
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_store_bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-TRUE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s1 :: v_dual_mov_b32 v1, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_store_bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    s_load_b64 s[0:1], s[4:5], 0x2c
; GFX11-FAKE16-NEXT:    s_load_b32 s2, s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  store atomic bfloat %in, ptr %out seq_cst, align 2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_inc_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_max_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_inc_i32_max_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_max_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_max_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2 offset:4092
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_max_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2 offset:4092
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1023
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_max_offset_p1(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_inc_i32_max_offset_p1:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_max_offset_p1:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_max_offset_p1:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
; GCN3-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_max_offset_p1:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_co_u32 v0, s0, 0x1000, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, s1, s0
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1024
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_inc_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_inc v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_inc_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_incr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_inc_i32_incr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_incr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_incr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_incr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret_incr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_inc_i32_ret_incr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_ret_incr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_ret_incr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_inc v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_ret_incr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_inc_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile uinc_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_inc_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile uinc_wrap ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_inc_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_inc_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile uinc_wrap ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_incr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_inc_i32_incr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_inc v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_incr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_inc v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_incr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_inc v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_incr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_inc_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile uinc_wrap ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_inc_i32_ret_incr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_inc_i32_ret_incr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_inc_i32_ret_incr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_inc_i32_ret_incr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_inc v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_inc_i32_ret_incr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_inc_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile uinc_wrap ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_dec_i32_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_max_offset(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_dec_i32_max_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_max_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0xffc
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_max_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2 offset:4092
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_max_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2 offset:4092
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1023
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_max_offset_p1(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_dec_i32_max_offset_p1:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_max_offset_p1:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 0x1000
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_max_offset_p1:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_add_co_u32_e32 v0, vcc, 0x1000, v0
; GCN3-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_max_offset_p1:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_co_u32 v0, s0, 0x1000, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, s1, s0
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 1024
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret_offset(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_dec_i32_ret_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_ret_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_ret_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_dec v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_ret_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_dec_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %gep = getelementptr inbounds i32, ptr %out, i32 4
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_decr64_offset(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_dec_i32_decr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_decr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_decr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2 offset:16
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_decr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2 offset:16
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret_decr64_offset(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_dec_i32_ret_decr64_offset:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_ret_decr64_offset:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_ret_decr64_offset:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_dec v2, v[0:1], v2 offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_ret_decr64_offset:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_dec_u32 v2, v[0:1], v2 offset:16 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %gep = getelementptr inbounds i32, ptr %ptr, i32 4
  %val = atomicrmw volatile udec_wrap ptr %gep, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32(ptr %out, i32 %in) {
; GCN1-LABEL: atomic_dec_i32:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s2, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s2
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s2
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s2, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s2
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s2, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s2
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile udec_wrap ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret(ptr %out, ptr %out2, i32 %in) {
; GCN1-LABEL: atomic_dec_i32_ret:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_ret:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_ret:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_ret:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_dec_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %val = atomicrmw volatile udec_wrap ptr %out, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_decr64(ptr %out, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_dec_i32_decr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
; GCN1-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s4, s[4:5], 0xb
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN1-NEXT:    s_add_u32 s0, s2, s0
; GCN1-NEXT:    s_addc_u32 s1, s3, s1
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s4
; GCN1-NEXT:    flat_atomic_dec v[0:1], v2
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_decr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN2-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s4, s[4:5], 0x2c
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN2-NEXT:    s_add_u32 s0, s2, s0
; GCN2-NEXT:    s_addc_u32 s1, s3, s1
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s4
; GCN2-NEXT:    flat_atomic_dec v[0:1], v2
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_decr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x34
; GCN3-NEXT:    s_load_dwordx2 s[2:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s6, s[4:5], 0x2c
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GCN3-NEXT:    s_add_u32 s0, s2, s0
; GCN3-NEXT:    s_addc_u32 s1, s3, s1
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s6
; GCN3-NEXT:    flat_atomic_dec v[0:1], v2
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_decr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[0:1], s[4:5], 0x34
; GFX11-NEXT:    s_load_b64 s[2:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s4, s[4:5], 0x2c
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[0:1], s[0:1], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s2, s0
; GFX11-NEXT:    s_addc_u32 s1, s3, s1
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s4
; GFX11-NEXT:    flat_atomic_dec_u32 v[0:1], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile udec_wrap ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  ret void
}

define amdgpu_kernel void @atomic_dec_i32_ret_decr64(ptr %out, ptr %out2, i32 %in, i64 %index) {
; GCN1-LABEL: atomic_dec_i32_ret_decr64:
; GCN1:       ; %bb.0: ; %entry
; GCN1-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0xf
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_load_dword s8, s[4:5], 0xd
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN1-NEXT:    s_add_u32 s0, s0, s4
; GCN1-NEXT:    s_addc_u32 s1, s1, s5
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    v_mov_b32_e32 v2, s8
; GCN1-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_dword v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_dec_i32_ret_decr64:
; GCN2:       ; %bb.0: ; %entry
; GCN2-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN2-NEXT:    s_add_u32 s0, s0, s4
; GCN2-NEXT:    s_addc_u32 s1, s1, s5
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    v_mov_b32_e32 v2, s8
; GCN2-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_dword v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_dec_i32_ret_decr64:
; GCN3:       ; %bb.0: ; %entry
; GCN3-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x3c
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_load_dword s8, s[4:5], 0x34
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GCN3-NEXT:    s_add_u32 s0, s0, s4
; GCN3-NEXT:    s_addc_u32 s1, s1, s5
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    v_mov_b32_e32 v2, s8
; GCN3-NEXT:    flat_atomic_dec v2, v[0:1], v2 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_dword v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-LABEL: atomic_dec_i32_ret_decr64:
; GFX11:       ; %bb.0: ; %entry
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    s_load_b64 s[6:7], s[4:5], 0x3c
; GFX11-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-NEXT:    s_load_b32 s8, s[4:5], 0x34
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_lshl_b64 s[4:5], s[6:7], 2
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    s_add_u32 s0, s0, s4
; GFX11-NEXT:    s_addc_u32 s1, s1, s5
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_mov_b32_e32 v2, s8
; GFX11-NEXT:    flat_atomic_dec_u32 v2, v[0:1], v2 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-NEXT:    flat_store_b32 v[0:1], v2
; GFX11-NEXT:    s_endpgm
entry:
  %ptr = getelementptr inbounds i32, ptr %out, i64 %index
  %val = atomicrmw volatile udec_wrap ptr %ptr, i32 %in syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
  store i32 %val, ptr %out2
  ret void
}

define amdgpu_kernel void @atomic_load_f16_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_f16_offset:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f16_offset:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f16_offset:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_f16_offset:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_f16_offset:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  %gep = getelementptr inbounds half, ptr %in, i64 8
  %val = load atomic half, ptr %gep  seq_cst, align 2
  store half %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_f16(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_f16:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_f16:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_f16:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_f16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_f16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  %val = load atomic half, ptr %in seq_cst, align 2
  store half %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_bf16_offset(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_bf16_offset:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    s_add_u32 s0, s0, 16
; GCN1-NEXT:    s_addc_u32 s1, s1, 0
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_bf16_offset:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    s_add_u32 s0, s0, 16
; GCN2-NEXT:    s_addc_u32 s1, s1, 0
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_bf16_offset:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] offset:16 glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_bf16_offset:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] offset:16 glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_bf16_offset:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] offset:16 glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  %gep = getelementptr inbounds bfloat, ptr %in, i64 8
  %val = load atomic bfloat, ptr %gep  seq_cst, align 2
  store bfloat %val, ptr %out
  ret void
}

define amdgpu_kernel void @atomic_load_bf16(ptr %in, ptr %out) {
; GCN1-LABEL: atomic_load_bf16:
; GCN1:       ; %bb.0:
; GCN1-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
; GCN1-NEXT:    v_mov_b32_e32 v0, s0
; GCN1-NEXT:    v_mov_b32_e32 v1, s1
; GCN1-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN1-NEXT:    buffer_wbinvl1_vol
; GCN1-NEXT:    v_mov_b32_e32 v0, s2
; GCN1-NEXT:    v_mov_b32_e32 v1, s3
; GCN1-NEXT:    flat_store_short v[0:1], v2
; GCN1-NEXT:    s_endpgm
;
; GCN2-LABEL: atomic_load_bf16:
; GCN2:       ; %bb.0:
; GCN2-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
; GCN2-NEXT:    v_mov_b32_e32 v0, s0
; GCN2-NEXT:    v_mov_b32_e32 v1, s1
; GCN2-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN2-NEXT:    buffer_wbinvl1_vol
; GCN2-NEXT:    v_mov_b32_e32 v0, s2
; GCN2-NEXT:    v_mov_b32_e32 v1, s3
; GCN2-NEXT:    flat_store_short v[0:1], v2
; GCN2-NEXT:    s_endpgm
;
; GCN3-LABEL: atomic_load_bf16:
; GCN3:       ; %bb.0:
; GCN3-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
; GCN3-NEXT:    s_waitcnt lgkmcnt(0)
; GCN3-NEXT:    v_mov_b32_e32 v0, s0
; GCN3-NEXT:    v_mov_b32_e32 v1, s1
; GCN3-NEXT:    flat_load_ushort v2, v[0:1] glc
; GCN3-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GCN3-NEXT:    buffer_wbinvl1_vol
; GCN3-NEXT:    v_mov_b32_e32 v0, s2
; GCN3-NEXT:    v_mov_b32_e32 v1, s3
; GCN3-NEXT:    flat_store_short v[0:1], v2
; GCN3-NEXT:    s_endpgm
;
; GFX11-TRUE16-LABEL: atomic_load_bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-TRUE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v2, s3
; GFX11-TRUE16-NEXT:    flat_load_d16_b16 v0, v[0:1] glc
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    buffer_gl1_inv
; GFX11-TRUE16-NEXT:    buffer_gl0_inv
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, s2
; GFX11-TRUE16-NEXT:    flat_store_b16 v[1:2], v0
; GFX11-TRUE16-NEXT:    s_endpgm
;
; GFX11-FAKE16-LABEL: atomic_load_bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_load_b128 s[0:3], s[4:5], 0x24
; GFX11-FAKE16-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-FAKE16-NEXT:    flat_load_u16 v2, v[0:1] glc
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    buffer_gl1_inv
; GFX11-FAKE16-NEXT:    buffer_gl0_inv
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
; GFX11-FAKE16-NEXT:    flat_store_b16 v[0:1], v2
; GFX11-FAKE16-NEXT:    s_endpgm
  %val = load atomic bfloat, ptr %in seq_cst, align 2
  store bfloat %val, ptr %out
  ret void
}

!0 = !{}