; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GCN,GFX10 %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX11 %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 -mattr=+wavefrontsize64 < %s | FileCheck -check-prefixes=GFX12 %s

; Test using saddr addressing mode of global_* flat atomic instructions.

define amdgpu_ps void @global_xchg_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw xchg ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

; Maximum positive offset on gfx10
define amdgpu_ps void @global_xchg_saddr_i32_nortn_offset_2047(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_nortn_offset_2047:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v1, s[2:3] offset:2047
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_i32_nortn_offset_2047:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap v0, v1, s[2:3] offset:2047
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_i32_nortn_offset_2047:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3] offset:2047
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_i32_nortn_offset_2047:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3] offset:2047 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 2047
  %unused = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

; Maximum negative offset on gfx10
define amdgpu_ps void @global_xchg_saddr_i32_nortn_offset_neg2048(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_nortn_offset_neg2048:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v1, s[2:3] offset:-2048
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_i32_nortn_offset_neg2048:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap v0, v1, s[2:3] offset:-2048
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_i32_nortn_offset_neg2048:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3] offset:-2048
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_i32_nortn_offset_neg2048:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v1, s[2:3] offset:-2048 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -2048
  %unused = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps float @global_xchg_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw xchg ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_xchg_saddr_i32_rtn_2048(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_rtn_2048:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v0, v1, s[2:3] offset:2048 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_i32_rtn_2048:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_add_co_u32 v0, s[0:1], s2, v0
; GFX10-NEXT:    v_add_co_ci_u32_e64 v3, s[0:1], s3, 0, s[0:1]
; GFX10-NEXT:    v_add_co_u32 v2, vcc, 0x800, v0
; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc, 0, v3, vcc
; GFX10-NEXT:    global_atomic_swap v0, v[2:3], v1, off glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_i32_rtn_2048:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] offset:2048 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_i32_rtn_2048:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] offset:2048 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 2048
  %rtn = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_xchg_saddr_i32_rtn_neg2048(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_i32_rtn_neg2048:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap v0, v0, v1, s[2:3] offset:-2048 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_i32_rtn_neg2048:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap v0, v0, v1, s[2:3] offset:-2048 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_i32_rtn_neg2048:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] offset:-2048 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_i32_rtn_neg2048:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[2:3] offset:-2048 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -2048
  %rtn = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

; --------------------------------------------------------------------------------
; Uniformity edge cases
; --------------------------------------------------------------------------------

@ptr.in.lds = internal addrspace(3) global ptr addrspace(1) poison

; Base pointer is uniform, but also in VGPRs
define amdgpu_ps float @global_xchg_saddr_uniform_ptr_in_vgprs_rtn(i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    ds_read_b64 v[2:3], v2
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s0, v2
; GFX9-NEXT:    v_readfirstlane_b32 s1, v3
; GFX9-NEXT:    s_nop 4
; GFX9-NEXT:    global_atomic_swap v0, v0, v1, s[0:1] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v2, 0
; GFX10-NEXT:    ds_read_b64 v[2:3], v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_readfirstlane_b32 s0, v2
; GFX10-NEXT:    v_readfirstlane_b32 s1, v3
; GFX10-NEXT:    global_atomic_swap v0, v0, v1, s[0:1] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v2, 0
; GFX11-NEXT:    ds_load_b64 v[2:3], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s0, v2
; GFX11-NEXT:    v_readfirstlane_b32 s1, v3
; GFX11-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[0:1] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v2, 0
; GFX12-NEXT:    ds_load_b64 v[2:3], v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v2
; GFX12-NEXT:    v_readfirstlane_b32 s1, v3
; GFX12-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[0:1] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %sbase = load ptr addrspace(1), ptr addrspace(3) @ptr.in.lds
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw xchg ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

; Base pointer is uniform, but also in VGPRs, with imm offset
define amdgpu_ps float @global_xchg_saddr_uniform_ptr_in_vgprs_rtn_immoffset(i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn_immoffset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    ds_read_b64 v[2:3], v2
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s0, v2
; GFX9-NEXT:    v_readfirstlane_b32 s1, v3
; GFX9-NEXT:    s_nop 4
; GFX9-NEXT:    global_atomic_swap v0, v0, v1, s[0:1] offset:42 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn_immoffset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v2, 0
; GFX10-NEXT:    ds_read_b64 v[2:3], v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_readfirstlane_b32 s0, v2
; GFX10-NEXT:    v_readfirstlane_b32 s1, v3
; GFX10-NEXT:    global_atomic_swap v0, v0, v1, s[0:1] offset:42 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn_immoffset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v2, 0
; GFX11-NEXT:    ds_load_b64 v[2:3], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s0, v2
; GFX11-NEXT:    v_readfirstlane_b32 s1, v3
; GFX11-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[0:1] offset:42 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_rtn_immoffset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v2, 0
; GFX12-NEXT:    ds_load_b64 v[2:3], v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v2
; GFX12-NEXT:    v_readfirstlane_b32 s1, v3
; GFX12-NEXT:    global_atomic_swap_b32 v0, v0, v1, s[0:1] offset:42 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %sbase = load ptr addrspace(1), ptr addrspace(3) @ptr.in.lds
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 42
  %rtn = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

; Base pointer is uniform, but also in VGPRs
define amdgpu_ps void @global_xchg_saddr_uniform_ptr_in_vgprs_nortn(i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    ds_read_b64 v[2:3], v2
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s0, v2
; GFX9-NEXT:    v_readfirstlane_b32 s1, v3
; GFX9-NEXT:    s_nop 4
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v2, 0
; GFX10-NEXT:    ds_read_b64 v[2:3], v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_readfirstlane_b32 s0, v2
; GFX10-NEXT:    v_readfirstlane_b32 s1, v3
; GFX10-NEXT:    global_atomic_swap v0, v1, s[0:1]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v2, 0
; GFX11-NEXT:    ds_load_b64 v[2:3], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s0, v2
; GFX11-NEXT:    v_readfirstlane_b32 s1, v3
; GFX11-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v2, 0
; GFX12-NEXT:    ds_load_b64 v[2:3], v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v2
; GFX12-NEXT:    v_readfirstlane_b32 s1, v3
; GFX12-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %sbase = load ptr addrspace(1), ptr addrspace(3) @ptr.in.lds
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw xchg ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

; Base pointer is uniform, but also in VGPRs, with imm offset
define amdgpu_ps void @global_xchg_saddr_uniform_ptr_in_vgprs_nortn_immoffset(i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn_immoffset:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v2, 0
; GFX9-NEXT:    ds_read_b64 v[2:3], v2
; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s0, v2
; GFX9-NEXT:    v_readfirstlane_b32 s1, v3
; GFX9-NEXT:    s_nop 4
; GFX9-NEXT:    global_atomic_swap v0, v1, s[0:1] offset:42
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn_immoffset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v2, 0
; GFX10-NEXT:    ds_read_b64 v[2:3], v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_readfirstlane_b32 s0, v2
; GFX10-NEXT:    v_readfirstlane_b32 s1, v3
; GFX10-NEXT:    global_atomic_swap v0, v1, s[0:1] offset:42
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn_immoffset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v2, 0
; GFX11-NEXT:    ds_load_b64 v[2:3], v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s0, v2
; GFX11-NEXT:    v_readfirstlane_b32 s1, v3
; GFX11-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1] offset:42
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_uniform_ptr_in_vgprs_nortn_immoffset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v2, 0
; GFX12-NEXT:    ds_load_b64 v[2:3], v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_readfirstlane_b32 s0, v2
; GFX12-NEXT:    v_readfirstlane_b32 s1, v3
; GFX12-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1] offset:42 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %sbase = load ptr addrspace(1), ptr addrspace(3) @ptr.in.lds
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 42
  %unused = atomicrmw xchg ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; All atomicrmw ops
; --------------------------------------------------------------------------------

; --------------------------------------------------------------------------------
; atomicrmw xchg
; --------------------------------------------------------------------------------

define amdgpu_ps <2 x float> @global_xchg_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xchg_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw xchg ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_xchg_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xchg_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xchg_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xchg_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xchg_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw xchg ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_xchg_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xchg_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw xchg ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_xchg_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xchg_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_swap_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xchg_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_swap_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xchg_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_swap_b64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xchg_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_swap_b64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw xchg ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw add
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_add_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_add_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_add_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_add_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_add_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw add ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_add_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_add_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_add_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_add_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_add_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw add ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_add_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_add_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_add_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_add_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_add_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw add ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_add_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_add_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_add_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_add_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_add_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw add ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_add_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_add_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_add_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_add_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_add_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw add ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_add_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_add_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_add_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_add_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_add_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw add ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_add_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_add_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_add_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_add_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_add_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw add ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_add_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_add_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_add_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_add_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_add_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_add_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_add_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_add_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_add_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw add ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw sub
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_sub_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_sub_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_sub_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_sub_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_sub_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw sub ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_sub_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_sub_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_sub_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_sub_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_sub_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw sub ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_sub_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_sub_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_sub_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_sub_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_sub_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw sub ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_sub_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_sub_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_sub_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_sub_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_sub_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw sub ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_sub_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_sub_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_sub_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_sub_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_sub_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw sub ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_sub_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_sub_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_sub_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_sub_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_sub_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw sub ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_sub_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_sub_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_sub_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_sub_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_sub_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw sub ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_sub_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_sub_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_sub_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_sub_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_sub_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_sub_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_sub_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_sub_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_sub_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw sub ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw and
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_and_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_and_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_and_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_and_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_and_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw and ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_and_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_and_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_and_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_and_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_and_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw and ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_and_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_and_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_and_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_and_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_and_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw and ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_and_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_and_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_and_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_and_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_and_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw and ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_and_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_and_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_and_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_and_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_and_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw and ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_and_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_and_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_and_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_and_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_and_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw and ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_and_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_and_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_and_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_and_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_and_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw and ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_and_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_and_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_and_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_and_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_and_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_and_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_and_b64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_and_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_and_b64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw and ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw or
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_or_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_or_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_or_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_or_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_or_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw or ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_or_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_or_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_or_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_or_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_or_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw or ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_or_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_or_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_or_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_or_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_or_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw or ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_or_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_or_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_or_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_or_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_or_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw or ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_or_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_or_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_or_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_or_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_or_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw or ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_or_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_or_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_or_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_or_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_or_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw or ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_or_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_or_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_or_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_or_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_or_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw or ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_or_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_or_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_or_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_or_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_or_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_or_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_or_b64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_or_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_or_b64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw or ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw xor
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_xor_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xor_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xor_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xor_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xor_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw xor ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_xor_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xor_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xor_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xor_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xor_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw xor ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_xor_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xor_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor v0, v1, s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xor_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xor_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xor_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw xor ptr addrspace(1) %gep0, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_xor_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_xor_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xor_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xor_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xor_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw xor ptr addrspace(1) %gep1, i32 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_xor_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xor_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xor_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xor_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xor_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw xor ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_xor_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xor_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_xor_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_xor_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_xor_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw xor ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_xor_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xor_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xor_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xor_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xor_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw xor ptr addrspace(1) %gep0, i64 %data syncscope("agent") seq_cst
  ret void
}

define amdgpu_ps void @global_xor_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_xor_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_xor_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_xor_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_xor_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_xor_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_xor_b64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_xor_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_xor_b64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw xor ptr addrspace(1) %gep1, i64 %data syncscope("agent") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw max
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_max_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_max_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_max_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_max_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_max_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw max ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_max_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_max_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_max_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_max_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_max_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw max ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_max_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_max_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_max_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_max_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_max_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i32 v0, v1, s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw max ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_max_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_max_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_max_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_max_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_max_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i32 v0, v1, s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw max ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_max_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_max_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_max_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_max_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw max ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_max_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_max_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_max_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_max_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_max_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw max ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_max_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_max_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_max_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_max_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_max_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw max ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_max_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_max_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_max_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smax_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_max_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_max_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_i64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw max ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw min
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_min_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_min_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_min_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_min_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_min_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw min ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_min_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_min_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_min_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_min_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_min_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw min ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_min_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_min_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_min_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_min_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_min_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i32 v0, v1, s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw min ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_min_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_min_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_min_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_min_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_min_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i32 v0, v1, s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw min ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_min_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_min_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_min_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_min_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw min ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_min_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_min_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_min_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_min_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_min_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw min ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_min_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_min_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_min_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_min_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_min_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw min ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_min_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_min_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_min_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_smin_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_min_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_min_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_i64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw min ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw umax
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_umax_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umax_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umax_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umax_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umax_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw umax ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_umax_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umax_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umax_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umax_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umax_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw umax ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_umax_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umax_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umax_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umax_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umax_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u32 v0, v1, s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw umax ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_umax_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umax_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umax_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umax_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umax_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw umax ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umax_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umax_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umax_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umax_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw umax ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_umax_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umax_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umax_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umax_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umax_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw umax ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_umax_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umax_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umax_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umax_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umax_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw umax ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_umax_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umax_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umax_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umax_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umax_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umax_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_max_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw umax ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; atomicrmw umin
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_umin_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umin_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umin_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umin_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umin_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw umin ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_umin_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umin_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umin_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin v0, v0, v1, s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umin_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umin_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw umin ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_umin_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umin_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umin_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umin_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umin_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u32 v0, v1, s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw umin ptr addrspace(1) %gep0, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_umin_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GFX9-LABEL: global_umin_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umin_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin v0, v1, s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umin_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umin_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw umin ptr addrspace(1) %gep1, i32 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umin_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umin_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umin_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umin_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw umin ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_umin_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umin_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_umin_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_umin_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_umin_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SE
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw umin ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_umin_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umin_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umin_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umin_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umin_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3] scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw umin ptr addrspace(1) %gep0, i64 %data syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_ps void @global_umin_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GFX9-LABEL: global_umin_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_umin_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    global_atomic_umin_x2 v0, v[1:2], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_umin_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_umin_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_min_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_SE
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw umin ptr addrspace(1) %gep1, i64 %data syncscope("workgroup") seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; cmpxchg
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_cmpxchg_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %cmp, i32 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i32_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v3, v1
; GFX9-NEXT:    global_atomic_cmpswap v0, v0, v[2:3], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_cmpxchg_saddr_i32_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v3, v1
; GFX10-NEXT:    global_atomic_cmpswap v0, v0, v[2:3], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_cmpxchg_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v3, v1
; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v0, v[2:3], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_cmpxchg_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v3, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b32 v0, v0, v[2:3], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %cmpxchg = cmpxchg ptr addrspace(1) %gep0, i32 %cmp, i32 %data seq_cst seq_cst
  %rtn = extractvalue { i32, i1 } %cmpxchg, 0
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_cmpxchg_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %cmp, i32 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i32_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v3, v1
; GFX9-NEXT:    global_atomic_cmpswap v0, v0, v[2:3], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_cmpxchg_saddr_i32_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v3, v1
; GFX10-NEXT:    global_atomic_cmpswap v0, v0, v[2:3], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_cmpxchg_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v3, v1
; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v0, v[2:3], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_cmpxchg_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v3, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b32 v0, v0, v[2:3], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %cmpxchg = cmpxchg ptr addrspace(1) %gep1, i32 %cmp, i32 %data seq_cst seq_cst
  %rtn = extractvalue { i32, i1 } %cmpxchg, 0
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_cmpxchg_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %cmp, i32 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i32_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v3, v1
; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_cmpxchg_saddr_i32_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v3, v1
; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_cmpxchg_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v3, v1
; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_cmpxchg_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v3, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], s[2:3] scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = cmpxchg ptr addrspace(1) %gep0, i32 %cmp, i32 %data seq_cst seq_cst
  ret void
}

define amdgpu_ps void @global_cmpxchg_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %cmp, i32 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i32_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v3, v1
; GFX9-NEXT:    global_atomic_cmpswap v0, v[2:3], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_cmpxchg_saddr_i32_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v3, v1
; GFX10-NEXT:    global_atomic_cmpswap v0, v[2:3], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_cmpxchg_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v3, v1
; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_cmpxchg_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v3, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b32 v0, v[2:3], s[2:3] offset:-128 scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = cmpxchg ptr addrspace(1) %gep1, i32 %cmp, i32 %data seq_cst seq_cst
  ret void
}

define amdgpu_ps <2 x float> @global_cmpxchg_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %cmp, i64 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i64_rtn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v6, v2
; GFX9-NEXT:    v_mov_b32_e32 v5, v1
; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v0, v[3:6], s[2:3] glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_cmpxchg_saddr_i64_rtn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v6, v2
; GFX10-NEXT:    v_mov_b32_e32 v5, v1
; GFX10-NEXT:    global_atomic_cmpswap_x2 v[0:1], v0, v[3:6], s[2:3] glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_cmpxchg_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v6, v2
; GFX11-NEXT:    v_mov_b32_e32 v5, v1
; GFX11-NEXT:    global_atomic_cmpswap_b64 v[0:1], v0, v[3:6], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_cmpxchg_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v6, v2
; GFX12-NEXT:    v_mov_b32_e32 v5, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b64 v[0:1], v0, v[3:6], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %cmpxchg = cmpxchg ptr addrspace(1) %gep0, i64 %cmp, i64 %data seq_cst seq_cst
  %rtn = extractvalue { i64, i1 } %cmpxchg, 0
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_cmpxchg_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %cmp, i64 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i64_rtn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v6, v2
; GFX9-NEXT:    v_mov_b32_e32 v5, v1
; GFX9-NEXT:    global_atomic_cmpswap_x2 v[0:1], v0, v[3:6], s[2:3] offset:-128 glc
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    ; return to shader part epilog
;
; GFX10-LABEL: global_cmpxchg_saddr_i64_rtn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v6, v2
; GFX10-NEXT:    v_mov_b32_e32 v5, v1
; GFX10-NEXT:    global_atomic_cmpswap_x2 v[0:1], v0, v[3:6], s[2:3] offset:-128 glc
; GFX10-NEXT:    s_waitcnt vmcnt(0)
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_cmpxchg_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v6, v2
; GFX11-NEXT:    v_mov_b32_e32 v5, v1
; GFX11-NEXT:    global_atomic_cmpswap_b64 v[0:1], v0, v[3:6], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_cmpxchg_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v6, v2
; GFX12-NEXT:    v_mov_b32_e32 v5, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b64 v[0:1], v0, v[3:6], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %cmpxchg = cmpxchg ptr addrspace(1) %gep1, i64 %cmp, i64 %data seq_cst seq_cst
  %rtn = extractvalue { i64, i1 } %cmpxchg, 0
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_cmpxchg_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %cmp, i64 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i64_nortn:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v6, v2
; GFX9-NEXT:    v_mov_b32_e32 v5, v1
; GFX9-NEXT:    global_atomic_cmpswap_x2 v0, v[3:6], s[2:3]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_cmpxchg_saddr_i64_nortn:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v6, v2
; GFX10-NEXT:    v_mov_b32_e32 v5, v1
; GFX10-NEXT:    global_atomic_cmpswap_x2 v0, v[3:6], s[2:3]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_cmpxchg_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v6, v2
; GFX11-NEXT:    v_mov_b32_e32 v5, v1
; GFX11-NEXT:    global_atomic_cmpswap_b64 v0, v[3:6], s[2:3]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_cmpxchg_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v6, v2
; GFX12-NEXT:    v_mov_b32_e32 v5, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b64 v0, v[3:6], s[2:3] scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = cmpxchg ptr addrspace(1) %gep0, i64 %cmp, i64 %data seq_cst seq_cst
  ret void
}

define amdgpu_ps void @global_cmpxchg_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %cmp, i64 %data) {
; GFX9-LABEL: global_cmpxchg_saddr_i64_nortn_neg128:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    v_mov_b32_e32 v6, v2
; GFX9-NEXT:    v_mov_b32_e32 v5, v1
; GFX9-NEXT:    global_atomic_cmpswap_x2 v0, v[3:6], s[2:3] offset:-128
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_wbinvl1
; GFX9-NEXT:    s_endpgm
;
; GFX10-LABEL: global_cmpxchg_saddr_i64_nortn_neg128:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    v_mov_b32_e32 v6, v2
; GFX10-NEXT:    v_mov_b32_e32 v5, v1
; GFX10-NEXT:    global_atomic_cmpswap_x2 v0, v[3:6], s[2:3] offset:-128
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    buffer_gl1_inv
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    s_endpgm
;
; GFX11-LABEL: global_cmpxchg_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    v_mov_b32_e32 v6, v2
; GFX11-NEXT:    v_mov_b32_e32 v5, v1
; GFX11-NEXT:    global_atomic_cmpswap_b64 v0, v[3:6], s[2:3] offset:-128
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    buffer_gl1_inv
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_cmpxchg_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    v_mov_b32_e32 v6, v2
; GFX12-NEXT:    v_mov_b32_e32 v5, v1
; GFX12-NEXT:    global_wb scope:SCOPE_SYS
; GFX12-NEXT:    global_atomic_cmpswap_b64 v0, v[3:6], s[2:3] offset:-128 scope:SCOPE_SYS
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SYS
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = cmpxchg ptr addrspace(1) %gep1, i64 %cmp, i64 %data seq_cst seq_cst
  ret void
}

; --------------------------------------------------------------------------------
; amdgcn atomic inc
; --------------------------------------------------------------------------------

define amdgpu_ps float @global_inc_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_inc_saddr_i32_rtn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc v0, v0, v1, s[2:3] glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_inc_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_inc_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw uinc_wrap ptr addrspace(1) %gep0, i32 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_inc_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_inc_saddr_i32_rtn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc v0, v0, v1, s[2:3] offset:-128 glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_inc_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_inc_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw uinc_wrap ptr addrspace(1) %gep1, i32 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_inc_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_inc_saddr_i32_nortn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc v0, v1, s[2:3]
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_inc_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_inc_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw uinc_wrap ptr addrspace(1) %gep0, i32 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps void @global_inc_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_inc_saddr_i32_nortn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc v0, v1, s[2:3] offset:-128
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_inc_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_inc_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw uinc_wrap ptr addrspace(1) %gep1, i32 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps <2 x float> @global_inc_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_inc_saddr_i64_rtn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_inc_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_inc_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw uinc_wrap ptr addrspace(1) %gep0, i64 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_inc_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_inc_saddr_i64_rtn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_inc_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_inc_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw uinc_wrap ptr addrspace(1) %gep1, i64 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_inc_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_inc_saddr_i64_nortn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc_x2 v0, v[1:2], s[2:3]
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_inc_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_inc_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw uinc_wrap ptr addrspace(1) %gep0, i64 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps void @global_inc_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_inc_saddr_i64_nortn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_inc_x2 v0, v[1:2], s[2:3] offset:-128
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_inc_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_inc_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_inc_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_inc_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw uinc_wrap ptr addrspace(1) %gep1, i64 %data syncscope("agent") monotonic
  ret void
}

; --------------------------------------------------------------------------------
; amdgcn atomic dec
; --------------------------------------------------------------------------------


define amdgpu_ps float @global_dec_saddr_i32_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_dec_saddr_i32_rtn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec v0, v0, v1, s[2:3] glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_dec_saddr_i32_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u32 v0, v0, v1, s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_dec_saddr_i32_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u32 v0, v0, v1, s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw udec_wrap ptr addrspace(1) %gep0, i32 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps float @global_dec_saddr_i32_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_dec_saddr_i32_rtn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec v0, v0, v1, s[2:3] offset:-128 glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_dec_saddr_i32_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u32 v0, v0, v1, s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_dec_saddr_i32_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u32 v0, v0, v1, s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw udec_wrap ptr addrspace(1) %gep1, i32 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i32 %rtn to float
  ret float %cast.rtn
}

define amdgpu_ps void @global_dec_saddr_i32_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_dec_saddr_i32_nortn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec v0, v1, s[2:3]
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_dec_saddr_i32_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u32 v0, v1, s[2:3]
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_dec_saddr_i32_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u32 v0, v1, s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw udec_wrap ptr addrspace(1) %gep0, i32 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps void @global_dec_saddr_i32_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i32 %data) {
; GCN-LABEL: global_dec_saddr_i32_nortn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec v0, v1, s[2:3] offset:-128
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_dec_saddr_i32_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u32 v0, v1, s[2:3] offset:-128
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_dec_saddr_i32_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u32 v0, v1, s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw udec_wrap ptr addrspace(1) %gep1, i32 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps <2 x float> @global_dec_saddr_i64_rtn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_dec_saddr_i64_rtn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec_x2 v[0:1], v0, v[1:2], s[2:3] glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_dec_saddr_i64_rtn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u64 v[0:1], v0, v[1:2], s[2:3] glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_dec_saddr_i64_rtn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u64 v[0:1], v0, v[1:2], s[2:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %rtn = atomicrmw udec_wrap ptr addrspace(1) %gep0, i64 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps <2 x float> @global_dec_saddr_i64_rtn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_dec_saddr_i64_rtn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec_x2 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GCN-NEXT:    s_waitcnt vmcnt(0)
; GCN-NEXT:    ; return to shader part epilog
;
; GFX11-LABEL: global_dec_saddr_i64_rtn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 glc
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    ; return to shader part epilog
;
; GFX12-LABEL: global_dec_saddr_i64_rtn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u64 v[0:1], v0, v[1:2], s[2:3] offset:-128 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
; GFX12-NEXT:    s_wait_loadcnt 0x0
; GFX12-NEXT:    ; return to shader part epilog
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %rtn = atomicrmw udec_wrap ptr addrspace(1) %gep1, i64 %data syncscope("agent") monotonic
  %cast.rtn = bitcast i64 %rtn to <2 x float>
  ret <2 x float> %cast.rtn
}

define amdgpu_ps void @global_dec_saddr_i64_nortn(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_dec_saddr_i64_nortn:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec_x2 v0, v[1:2], s[2:3]
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_dec_saddr_i64_nortn:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u64 v0, v[1:2], s[2:3]
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_dec_saddr_i64_nortn:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u64 v0, v[1:2], s[2:3] scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %unused = atomicrmw udec_wrap ptr addrspace(1) %gep0, i64 %data syncscope("agent") monotonic
  ret void
}

define amdgpu_ps void @global_dec_saddr_i64_nortn_neg128(ptr addrspace(1) inreg %sbase, i32 %voffset, i64 %data) {
; GCN-LABEL: global_dec_saddr_i64_nortn_neg128:
; GCN:       ; %bb.0:
; GCN-NEXT:    global_atomic_dec_x2 v0, v[1:2], s[2:3] offset:-128
; GCN-NEXT:    s_endpgm
;
; GFX11-LABEL: global_dec_saddr_i64_nortn_neg128:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    global_atomic_dec_u64 v0, v[1:2], s[2:3] offset:-128
; GFX11-NEXT:    s_endpgm
;
; GFX12-LABEL: global_dec_saddr_i64_nortn_neg128:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    global_atomic_dec_u64 v0, v[1:2], s[2:3] offset:-128 scope:SCOPE_DEV
; GFX12-NEXT:    s_endpgm
  %zext.offset = zext i32 %voffset to i64
  %gep0 = getelementptr inbounds i8, ptr addrspace(1) %sbase, i64 %zext.offset
  %gep1 = getelementptr inbounds i8, ptr addrspace(1) %gep0, i64 -128
  %unused = atomicrmw udec_wrap ptr addrspace(1) %gep1, i64 %data syncscope("agent") monotonic
  ret void
}

attributes #0 = { argmemonly nounwind willreturn }