test/CodeGen/AMDGPU/global-atomic-scan.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN:  opt -S -mtriple=amdgcn-- -passes=amdgpu-atomic-optimizer %s | FileCheck -check-prefix=IR %s

define amdgpu_kernel void @atomic_add_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_add_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_max_neg_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_add_i32_max_neg_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 -1024
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 -1024
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_soffset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_add_i32_soffset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 9000
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 9000
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_huge_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_add_i32_huge_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 47224239175595
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 47224239175595

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

define amdgpu_kernel void @atomic_add_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_add_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = add i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_add_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_add_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = add i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile add ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_add_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[OUT:%.*]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile add ptr addrspace(1) %out, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_add_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[OUT:%.*]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = add i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile add ptr addrspace(1) %out, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_add_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_add_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[PTR]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile add ptr addrspace(1) %ptr, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_add_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_add_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile add ptr addrspace(1) [[PTR]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = add i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile add ptr addrspace(1) %ptr, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_and_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_and_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -1, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = and i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    store i32 [[TMP13]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_and_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_and_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -1, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = and i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    store i32 [[TMP13]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile and ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_and_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile and ptr addrspace(1) %out, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_and_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -1, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = and i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    store i32 [[TMP13]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile and ptr addrspace(1) %out, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_and_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_and_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile and ptr addrspace(1) %ptr, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_and_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_and_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile and ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -1, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = and i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    store i32 [[TMP13]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile and ptr addrspace(1) %ptr, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_sub_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_sub_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = sub i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_sub_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_sub_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[GEP]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = sub i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile sub ptr addrspace(1) %gep, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_sub_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[OUT:%.*]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile sub ptr addrspace(1) %out, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_sub_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[OUT:%.*]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = sub i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile sub ptr addrspace(1) %out, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_sub_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[PTR]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile sub ptr addrspace(1) %ptr, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_sub_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_sub_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP0]])
; IR-NEXT:    [[TMP7:%.*]] = trunc i64 [[TMP6]] to i32
; IR-NEXT:    [[TMP8:%.*]] = mul i32 [[IN:%.*]], [[TMP7]]
; IR-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP9]], label [[TMP10:%.*]], label [[TMP12:%.*]]
; IR:       10:
; IR-NEXT:    [[TMP11:%.*]] = atomicrmw volatile sub ptr addrspace(1) [[PTR]], i32 [[TMP8]] seq_cst, align 4
; IR-NEXT:    br label [[TMP12]]
; IR:       12:
; IR-NEXT:    [[TMP13:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP11]], [[TMP10]] ]
; IR-NEXT:    [[TMP14:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP13]])
; IR-NEXT:    [[TMP15:%.*]] = mul i32 [[IN]], [[TMP5]]
; IR-NEXT:    [[TMP16:%.*]] = sub i32 [[TMP14]], [[TMP15]]
; IR-NEXT:    store i32 [[TMP16]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile sub ptr addrspace(1) %ptr, i32 %in seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_max_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_max_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -2147483648, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp sgt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_max_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_max_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -2147483648, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp sgt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile max ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_max_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile max ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_max_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -2147483648, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp sgt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile max ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_max_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_max_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile max ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_max_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_max_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile max ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 -2147483648, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp sgt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile max ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_umax_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_umax_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 0, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp ugt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_umax_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_umax_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 0, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp ugt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile umax ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_umax_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile umax ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_umax_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 0, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp ugt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile umax ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_umax_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umax ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_umax_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_umax_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile umax ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 0, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp ugt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile umax ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_offset(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_min_i32_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_min_i32_ret_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 2147483647, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp slt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %gep = getelementptr i32, ptr addrspace(1) %out, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64_offset(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_min_i32_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64_offset(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_min_i32_ret_addr64_offset(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr addrspace(1) [[PTR]], i64 4
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[GEP]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 2147483647, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp slt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %gep = getelementptr i32, ptr addrspace(1) %ptr, i64 4
  %val = atomicrmw volatile min ptr addrspace(1) %gep, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32(ptr addrspace(1) %out, i32 %in) {
; IR-LABEL: @atomic_min_i32(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile min ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in) {
; IR-LABEL: @atomic_min_i32_ret(
; IR-NEXT:  entry:
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[OUT:%.*]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 2147483647, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp slt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %val = atomicrmw volatile min ptr addrspace(1) %out, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}

define amdgpu_kernel void @atomic_min_i32_addr64(ptr addrspace(1) %out, i32 %in, i64 %index) {
; IR-LABEL: @atomic_min_i32_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile min ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  ret void
}

define amdgpu_kernel void @atomic_min_i32_ret_addr64(ptr addrspace(1) %out, ptr addrspace(1) %out2, i32 %in, i64 %index) {
; IR-LABEL: @atomic_min_i32_ret_addr64(
; IR-NEXT:  entry:
; IR-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(1) [[OUT:%.*]], i64 [[INDEX:%.*]]
; IR-NEXT:    [[TMP0:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
; IR-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x i32>
; IR-NEXT:    [[TMP2:%.*]] = extractelement <2 x i32> [[TMP1]], i32 0
; IR-NEXT:    [[TMP3:%.*]] = extractelement <2 x i32> [[TMP1]], i32 1
; IR-NEXT:    [[TMP4:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP2]], i32 0)
; IR-NEXT:    [[TMP5:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP3]], i32 [[TMP4]])
; IR-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[TMP5]], 0
; IR-NEXT:    br i1 [[TMP6]], label [[TMP7:%.*]], label [[TMP9:%.*]]
; IR:       7:
; IR-NEXT:    [[TMP8:%.*]] = atomicrmw volatile min ptr addrspace(1) [[PTR]], i32 [[IN:%.*]] syncscope("workgroup") seq_cst, align 4
; IR-NEXT:    br label [[TMP9]]
; IR:       9:
; IR-NEXT:    [[TMP10:%.*]] = phi i32 [ poison, [[ENTRY:%.*]] ], [ [[TMP8]], [[TMP7]] ]
; IR-NEXT:    [[TMP11:%.*]] = call i32 @llvm.amdgcn.readfirstlane(i32 [[TMP10]])
; IR-NEXT:    [[TMP12:%.*]] = select i1 [[TMP6]], i32 2147483647, i32 [[IN]]
; IR-NEXT:    [[TMP13:%.*]] = icmp slt i32 [[TMP11]], [[TMP12]]
; IR-NEXT:    [[TMP14:%.*]] = select i1 [[TMP13]], i32 [[TMP11]], i32 [[TMP12]]
; IR-NEXT:    store i32 [[TMP14]], ptr addrspace(1) [[OUT2:%.*]], align 4
; IR-NEXT:    ret void
;
entry:
  %ptr = getelementptr i32, ptr addrspace(1) %out, i64 %index
  %val = atomicrmw volatile min ptr addrspace(1) %ptr, i32 %in syncscope("workgroup") seq_cst
  store i32 %val, ptr addrspace(1) %out2
  ret void
}