AMDGPU: Improve atomicrmw fadd selection

[Petar.Avramovic]

Summary

Use same atomicrmw fadd expansion rules for gfx908, gfx940 and gfx11
as for gfx90a. Add missing globalisel legalizer support for flat
atomicrmw fadd f32 on gfx940 and gfx11.
Isel support for gfx11 will be added in D130579.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

1346

Using this predicate function instead of another ST method is inconsistent. Also can merge these two ifs into an or

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.global.atomic.fadd-with-ret.ll

7

Degrading the error is somewhat unfortunate, but we don't consistently do this anyway

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12783–12786

The flow of these options is getting too hard to follow. You're basically repeating all the same checks as above, so why not treat this as an independent case before the outer if?

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12787

I think both are still unsafe even when supported, on both targets.

[b-sumner]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12787

I agree, these are unsafe.

[Petar.Avramovic]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12787

In order to simplify all these conditions:
Obvious first requirement is that subtarget needs to have instruction.

• for gfx940 (MI300) having instruction is enough
• for all other subtargets fadd is unsafe and can be selected only when
  1. function has amdgpu-unsafe-fp-atomics=true
  2. and atomic rmw needs to be non-system scope and not "one-as"

Second bullet is already implemented for gfx90a (MI200).
If I understand this correctly, fadd atomic rmw for gfx908(MI100) and gfx11 is unsafe like for MI200 and can be selected under same conditions.

summary of logic

// mi300
if `subtarget == gfx940` and `subtarget has instruction`
  don't expand and return

// mi100, mi200, gfx11+
if 'amdgpu-unsafe-fp-atomics=true' and 'scope is non-system' and `subtarget has instruction`
  don't expand and return

expand

[b-sumner]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12787

There will be variants of gfx940 where atomics are still unsafe, so probably better to use the same approach as for gfx90a.

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12787

I am not completely sure about system scope though, is that OK for gfx908 and gfx11? It is certainly unsafe in terms of denorm handling.

[Petar.Avramovic]

Treat fadd rmw atomics on gfx908, gfx940 and gfx11+ the same way as for gfx90a.
Regression for gfx940 which now expands rmw in some cases that it did not used to.
Added tests to transforms for interesting cases that don't expand rmw.

[rampitec]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12786

The logic below seems correct to me, but why dropping hasAtomicFaddNoRtnInsts check? It is a fast way to skip the whole block for the targets which do not have it at all.

[rampitec]

LGTM for the expansion part.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

1344

Probably should factor this into a feature test

[Petar.Avramovic]

Use hasFlatAtomicFaddF32Inst feature check.

[tianshilei1992]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12804

If I read it correctly, Subtarget->hasAtomicFaddNoRtnInsts() is always true here right? Because it is in the block:

if ((AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) &&
    Subtarget->hasAtomicFaddNoRtnInsts()) {

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

if (ST.hasLDSFPAtomicAdd()) {
     Atomic.legalFor({{S32, LocalPtr}, {S32, RegionPtr}});
     if (ST.hasGFX90AInsts())
       Atomic.legalFor({{S64, LocalPtr}});
     if (ST.hasGFX940Insts())
       Atomic.legalFor({{V2S16, LocalPtr}});
   }
   if (ST.hasAtomicFaddInsts())
     Atomic.legalFor({{S32, GlobalPtr}});
+  if (ST.hasGFX940Insts() || ST.getGeneration() >= AMDGPUSubtarget::GFX11)
+    Atomic.legalFor({{S32, FlatPtr}});
 
   if (ST.hasGFX90AInsts()) {
     // These are legal with some caveats, and should have undergone expansion in
     // the IR in most situations
     // TODO: Move atomic expansion into legalizer
-    // TODO: Also supports <2 x f16>
     Atomic.legalFor({
         {S32, GlobalPtr},
         {S64, GlobalPtr},
         {S64, FlatPtr}
       });
   }
 
   // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling, and output

llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

case AMDGPU::G_INTRINSIC_W_SIDE_EFFECTS: {
     case Intrinsic::amdgcn_global_atomic_fmax:
     case Intrinsic::amdgcn_flat_atomic_fadd:
     case Intrinsic::amdgcn_flat_atomic_fmin:
     case Intrinsic::amdgcn_flat_atomic_fmax:
     case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
     case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16:
       return getDefaultMappingAllVGPR(MI);
     case Intrinsic::amdgcn_ds_ordered_add:
-    case Intrinsic::amdgcn_ds_ordered_swap: {
+    case Intrinsic::amdgcn_ds_ordered_swap:
+    case Intrinsic::amdgcn_ds_fadd_v2bf16: {
       unsigned DstSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
       OpdsMapping[0] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, DstSize);
       unsigned M0Bank = getRegBankID(MI.getOperand(2).getReg(), MRI,
                                  AMDGPU::SGPRRegBankID);
       OpdsMapping[2] = AMDGPU::getValueMapping(M0Bank, 32);
       OpdsMapping[3] = AMDGPU::getValueMapping(AMDGPU::VGPRRegBankID, 32);
       break;
     }

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   case AtomicRMWInst::FAdd: {
     Type *Ty = RMW->getType();
 
     // We don't have a way to support 16-bit atomics now, so just leave them
     // as-is.
     if (Ty->isHalfTy())
       return AtomicExpansionKind::None;
 
     if (!Ty->isFloatTy() && (!Subtarget->hasGFX90AInsts() || !Ty->isDoubleTy()))
       return AtomicExpansionKind::CmpXChg;
 
-    if ((AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) &&
-        Subtarget->hasAtomicFaddNoRtnInsts()) {
-      if (Subtarget->hasGFX940Insts())
-        return AtomicExpansionKind::None;
-
+    if (AS == AMDGPUAS::GLOBAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) {
       // The amdgpu-unsafe-fp-atomics attribute enables generation of unsafe
       // floating point atomic instructions. May generate more efficient code,
       // but may not respect rounding and denormal modes, and may give incorrect
       // results for certain memory destinations.
       if (RMW->getFunction()
               ->getFnAttribute("amdgpu-unsafe-fp-atomics")
               .getValueAsString() != "true")
         return AtomicExpansionKind::CmpXChg;
 
-      if (Subtarget->hasGFX90AInsts()) {
-        if (Ty->isFloatTy() && AS == AMDGPUAS::FLAT_ADDRESS)
-          return AtomicExpansionKind::CmpXChg;
-
-        auto SSID = RMW->getSyncScopeID();
-        if (SSID == SyncScope::System ||
-            SSID == RMW->getContext().getOrInsertSyncScopeID("one-as"))
-          return AtomicExpansionKind::CmpXChg;
+      // Always expand system scope fp atomics.
 
+      if (AS == AMDGPUAS::GLOBAL_ADDRESS && Ty->isFloatTy()) {
+        // global atomic fadd f32 no-rtn: gfx908, gfx90a, gfx940, gfx11+.
+        if (RMW->use_empty() && Subtarget->hasAtomicFaddNoRtnInsts())
+          return ReportUnsafeHWInst(AtomicExpansionKind::None);
+        // global atomic fadd f32 rtn: gfx90a, gfx940, gfx11+.
+        if (!RMW->use_empty() && Subtarget->hasAtomicFaddRtnInsts())
           return ReportUnsafeHWInst(AtomicExpansionKind::None);
       }
 
-      if (AS == AMDGPUAS::FLAT_ADDRESS)
-        return AtomicExpansionKind::CmpXChg;
+      // flat atomic fadd f32: gfx940, gfx11+.
+      if (AS == AMDGPUAS::FLAT_ADDRESS && Ty->isFloatTy() &&
+          (Subtarget->hasGFX940Insts() ||
+           Subtarget->getGeneration() >= AMDGPUSubtarget::GFX11))
+        return ReportUnsafeHWInst(AtomicExpansionKind::None);
 
-      return RMW->use_empty() ? ReportUnsafeHWInst(AtomicExpansionKind::None)
-                              : AtomicExpansionKind::CmpXChg;
+      // global and flat atomic fadd f64: gfx90a, gfx940.
+      if (Ty->isDoubleTy() && Subtarget->hasGFX90AInsts())
+        return ReportUnsafeHWInst(AtomicExpansionKind::None);
+
+      return AtomicExpansionKind::CmpXChg;
     }
 
     // DS FP atomics do respect the denormal mode, but the rounding mode is
     // fixed to round-to-nearest-even.
     // The only exception is DS_ADD_F64 which never flushes regardless of mode.
     if (AS == AMDGPUAS::LOCAL_ADDRESS && Subtarget->hasLDSFPAtomicAdd()) {
       if (!Ty->isDoubleTy())
         return AtomicExpansionKind::None;

llvm/test/CodeGen/AMDGPU/GlobalISel/flat-atomic-fadd.f32.ll

 ; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=MI300 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=MI300 %s
 ; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX11 %s
 
-; MI300: LLVM ERROR: unable to legalize instruction: %{{[0-9]+}}:_(s32) = G_ATOMICRMW_FADD
 ; GFX11: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.flat.atomic.fadd)
 
 define amdgpu_ps void @flat_atomic_fadd_f32_no_rtn_intrinsic(float* %ptr, float %data) {
+  ; MI300-LABEL: name: flat_atomic_fadd_f32_no_rtn_intrinsic
+  ; MI300: bb.1 (%ir-block.0):
+  ; MI300-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; MI300-NEXT: {{  $}}
+  ; MI300-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; MI300-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; MI300-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; MI300-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; MI300-NEXT:   FLAT_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; MI300-NEXT:   S_ENDPGM 0
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float* %ptr, float %data)
   ret void
 }
 
 define amdgpu_ps float @flat_atomic_fadd_f32_rtn_intrinsic(float* %ptr, float %data) {
+  ; MI300-LABEL: name: flat_atomic_fadd_f32_rtn_intrinsic
+  ; MI300: bb.1 (%ir-block.0):
+  ; MI300-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; MI300-NEXT: {{  $}}
+  ; MI300-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; MI300-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; MI300-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; MI300-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; MI300-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; MI300-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; MI300-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float* %ptr, float %data)
   ret float %ret
 }
 
 define amdgpu_ps void @flat_atomic_fadd_f32_no_rtn_atomicrmw(float* %ptr, float %data) #0 {
+  ; MI300-LABEL: name: flat_atomic_fadd_f32_no_rtn_atomicrmw
+  ; MI300: bb.1 (%ir-block.0):
+  ; MI300-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; MI300-NEXT: {{  $}}
+  ; MI300-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; MI300-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; MI300-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; MI300-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; MI300-NEXT:   FLAT_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr)
+  ; MI300-NEXT:   S_ENDPGM 0
   %ret = atomicrmw fadd float* %ptr, float %data syncscope("wavefront") monotonic
   ret void
 }
 
 define amdgpu_ps float @flat_atomic_fadd_f32_rtn_atomicrmw(float* %ptr, float %data) #0 {
+  ; MI300-LABEL: name: flat_atomic_fadd_f32_rtn_atomicrmw
+  ; MI300: bb.1 (%ir-block.0):
+  ; MI300-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; MI300-NEXT: {{  $}}
+  ; MI300-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; MI300-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; MI300-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; MI300-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; MI300-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr)
+  ; MI300-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; MI300-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
   %ret = atomicrmw fadd float* %ptr, float %data syncscope("wavefront") monotonic
   ret float %ret
 }
 
 declare float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float*, float)
 
 attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/GlobalISel/fp-atomics-gfx940.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -march=amdgcn -mcpu=gfx940 -verify-machineinstrs | FileCheck %s -check-prefix=GFX940
+; RUN: llc < %s -march=amdgcn -mcpu=gfx940 -global-isel -verify-machineinstrs | FileCheck %s -check-prefix=GFX940
 
 declare float @llvm.amdgcn.flat.atomic.fadd.f32.p0f32.f32(float* %ptr, float %data)
 declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p0v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
 
 ; bf16 atomics use v2i16 argument since there is no bf16 data type in the llvm.
 declare <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0v2i16(<2 x i16>* %ptr, <2 x i16> %data)
 declare <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1v2i16(<2 x i16> addrspace(1)* %ptr, <2 x i16> %data)
 declare <2 x half> @llvm.amdgcn.ds.fadd.v2f16(<2 x half> addrspace(3) * %ptr, <2 x half> %data, i32, i32, i1)
 declare <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(<2 x i16> addrspace(3) * %ptr, <2 x i16> %data)
 
 define amdgpu_kernel void @flat_atomic_fadd_f32_noret(float* %ptr, float %data) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_noret:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[2:3]
 ; GFX940-NEXT:    v_mov_b32_e32 v2, s4
 ; GFX940-NEXT:    flat_atomic_add_f32 v[0:1], v2
 ; GFX940-NEXT:    s_endpgm
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p0f32.f32(float* %ptr, float %data)
   ret void
 }
 
+define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat(float* %ptr) {
+; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; GFX940-NEXT:    s_mov_b64 s[2:3], 0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
+; GFX940-NEXT:    flat_load_dword v1, v[0:1]
+; GFX940-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_add_f32_e32 v0, 4.0, v1
+; GFX940-NEXT:    v_mov_b64_e32 v[2:3], s[0:1]
+; GFX940-NEXT:    buffer_wbl2 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT:    v_mov_b32_e32 v1, v0
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT:    s_endpgm
+  %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
+  ret void
+}
+
+define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat_ieee(float* %ptr) #0 {
+; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat_ieee:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; GFX940-NEXT:    s_mov_b64 s[2:3], 0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
+; GFX940-NEXT:    flat_load_dword v1, v[0:1]
+; GFX940-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_add_f32_e32 v0, 4.0, v1
+; GFX940-NEXT:    v_mov_b64_e32 v[2:3], s[0:1]
+; GFX940-NEXT:    buffer_wbl2 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT:    v_mov_b32_e32 v1, v0
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT:    s_endpgm
+  %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
+  ret void
+}
+
 define float @flat_atomic_fadd_f32_rtn(float* %ptr, float %data) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_rtn:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    flat_atomic_add_f32 v0, v[0:1], v2 sc0
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p0f32.f32(float* %ptr, float %data)
   ret float %ret
 }
 
+define float @flat_atomic_fadd_f32_rtn_pat(float* %ptr, float %data) {
+; GFX940-LABEL: flat_atomic_fadd_f32_rtn_pat:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    flat_load_dword v2, v[0:1]
+; GFX940-NEXT:    s_mov_b64 s[0:1], 0
+; GFX940-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b32_e32 v3, v2
+; GFX940-NEXT:    v_add_f32_e32 v2, 4.0, v3
+; GFX940-NEXT:    buffer_wbl2 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT:    v_mov_b32_e32 v0, v2
+; GFX940-NEXT:    s_setpc_b64 s[30:31]
+  %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
+  ret float %ret
+}
+
 define amdgpu_kernel void @flat_atomic_fadd_v2f16_noret(<2 x half>* %ptr, <2 x half> %data) {
 ; GFX940-LABEL: flat_atomic_fadd_v2f16_noret:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x2c
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[2:3]
 ; GFX940-NEXT:    v_mov_b32_e32 v2, s4

 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    flat_atomic_pk_add_f16 v0, v[0:1], v2 sc0
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p0v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
   ret <2 x half> %ret
 }
 
+define amdgpu_kernel void @flat_atomic_fadd_v2bf16_noret(<2 x i16>* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: flat_atomic_fadd_v2bf16_noret:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[2:3]
+; GFX940-NEXT:    v_mov_b32_e32 v2, s4
+; GFX940-NEXT:    flat_atomic_pk_add_bf16 v[0:1], v2
+; GFX940-NEXT:    s_endpgm
+  %ret = call <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0v2i16(<2 x i16>* %ptr, <2 x i16> %data)
+  ret void
+}
+
+define <2 x i16> @flat_atomic_fadd_v2bf16_rtn(<2 x i16>* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: flat_atomic_fadd_v2bf16_rtn:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    flat_atomic_pk_add_bf16 v0, v[0:1], v2 sc0
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call <2 x i16> @llvm.amdgcn.flat.atomic.fadd.v2bf16.p0v2i16(<2 x i16>* %ptr, <2 x i16> %data)
+  ret <2 x i16> %ret
+}
+
+define amdgpu_kernel void @global_atomic_fadd_v2bf16_noret(<2 x i16> addrspace(1)* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: global_atomic_fadd_v2bf16_noret:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_load_dword s4, s[0:1], 0x2c
+; GFX940-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
+; GFX940-NEXT:    v_mov_b32_e32 v1, 0
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b32_e32 v0, s4
+; GFX940-NEXT:    global_atomic_pk_add_bf16 v1, v0, s[2:3]
+; GFX940-NEXT:    s_endpgm
+  %ret = call <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1v2i16(<2 x i16> addrspace(1)* %ptr, <2 x i16> %data)
+  ret void
+}
+
+define <2 x i16> @global_atomic_fadd_v2bf16_rtn(<2 x i16> addrspace(1)* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: global_atomic_fadd_v2bf16_rtn:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    global_atomic_pk_add_bf16 v0, v[0:1], v2, off sc0
+; GFX940-NEXT:    s_waitcnt vmcnt(0)
+; GFX940-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call <2 x i16> @llvm.amdgcn.global.atomic.fadd.v2bf16.p1v2i16(<2 x i16> addrspace(1)* %ptr, <2 x i16> %data)
+  ret <2 x i16> %ret
+}
+
 define amdgpu_kernel void @local_atomic_fadd_v2f16_noret(<2 x half> addrspace(3)* %ptr, <2 x half> %data) {
 ; GFX940-LABEL: local_atomic_fadd_v2f16_noret:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX940-NEXT:    v_mov_b32_e32 v1, s1
 ; GFX940-NEXT:    ds_pk_add_f16 v0, v1
 ; GFX940-NEXT:    s_endpgm
   %ret = call <2 x half> @llvm.amdgcn.ds.fadd.v2f16(<2 x half> addrspace(3)* %ptr, <2 x half> %data, i32 0, i32 0, i1 0)
   ret void
 }
 
 define <2 x half> @local_atomic_fadd_v2f16_rtn(<2 x half> addrspace(3)* %ptr, <2 x half> %data) {
 ; GFX940-LABEL: local_atomic_fadd_v2f16_rtn:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    ds_pk_add_rtn_f16 v0, v0, v1
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %ret = call <2 x half> @llvm.amdgcn.ds.fadd.v2f16(<2 x half> addrspace(3)* %ptr, <2 x half> %data, i32 0, i32 0, i1 0)
   ret <2 x half> %ret
 }
+
+define amdgpu_kernel void @local_atomic_fadd_v2bf16_noret(<2 x i16> addrspace(3)* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: local_atomic_fadd_v2bf16_noret:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
+; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b32_e32 v0, s1
+; GFX940-NEXT:    v_mov_b32_e32 v1, s0
+; GFX940-NEXT:    buffer_wbl2 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    ds_pk_add_bf16 v1, v0
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    s_endpgm
+  %ret = call <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(<2 x i16> addrspace(3)* %ptr, <2 x i16> %data)
+  ret void
+}
+
+define <2 x i16> @local_atomic_fadd_v2bf16_rtn(<2 x i16> addrspace(3)* %ptr, <2 x i16> %data) {
+; GFX940-LABEL: local_atomic_fadd_v2bf16_rtn:
+; GFX940:       ; %bb.0:
+; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_wbl2 sc0 sc1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    ds_pk_add_rtn_bf16 v0, v0, v1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    s_setpc_b64 s[30:31]
+  %ret = call <2 x i16> @llvm.amdgcn.ds.fadd.v2bf16(<2 x i16> addrspace(3)* %ptr, <2 x i16> %data)
+  ret <2 x i16> %ret
+}
+
+attributes #0 = { "denormal-fp-math-f32"="ieee,ieee" }

llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.global.atomic.fadd-with-ret.ll

 ; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GFX90A %s
 ; RUN: not --crash llc -global-isel < %s -march=amdgcn -mcpu=gfx908 -verify-machineinstrs 2>&1 | FileCheck %s -check-prefix=GFX908
 
 declare float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* nocapture, float)
 declare <2 x half> @llvm.amdgcn.global.atomic.fadd.f32.p1v2f16.v2f16(<2 x half> addrspace(1)* nocapture, <2 x half>)
 
-; GFX908: error: {{.*}} return versions of fp atomics not supported
+; GFX908: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr_32(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.global.atomic.fadd)
 
 ; GFX90A-LABEL: {{^}}global_atomic_fadd_f32_rtn:
 ; GFX90A: global_atomic_add_f32 v0, v[0:1], v2, off glc
 define float @global_atomic_fadd_f32_rtn(float addrspace(1)* %ptr, float %data) {
   %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
   ret float %ret
 }
 
 ; GFX90A-LABEL: {{^}}global_atomic_fadd_v2f16_rtn:
 ; GFX90A: global_atomic_pk_add_f16 v0, v[0:1], v2, off glc
 define <2 x half> @global_atomic_fadd_v2f16_rtn(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
   %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.f32.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
   ret <2 x half> %ret
 }

llvm/test/CodeGen/AMDGPU/fp-atomics-gfx940.ll

; GFX940-NEXT:    s_endpgm
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p0f32.f32(float* %ptr, float %data)
   ret void
 }
 
 define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat(float* %ptr) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT:    v_mov_b32_e32 v2, 4.0
+; GFX940-NEXT:    s_mov_b64 s[2:3], 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
+; GFX940-NEXT:    flat_load_dword v1, v[0:1]
+; GFX940-NEXT:  .LBB1_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_add_f32_e32 v0, 4.0, v1
+; GFX940-NEXT:    v_mov_b64_e32 v[2:3], s[0:1]
 ; GFX940-NEXT:    buffer_wbl2 sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    flat_atomic_add_f32 v[0:1], v2 sc1
+; GFX940-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT:    v_mov_b32_e32 v1, v0
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT:    s_cbranch_execnz .LBB1_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX940-NEXT:    s_endpgm
   %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
   ret void
 }
 
 define amdgpu_kernel void @flat_atomic_fadd_f32_noret_pat_ieee(float* %ptr) #0 {
 ; GFX940-LABEL: flat_atomic_fadd_f32_noret_pat_ieee:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
-; GFX940-NEXT:    v_mov_b32_e32 v2, 4.0
+; GFX940-NEXT:    s_mov_b64 s[2:3], 0
 ; GFX940-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX940-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
+; GFX940-NEXT:    flat_load_dword v1, v[0:1]
+; GFX940-NEXT:  .LBB2_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_add_f32_e32 v0, 4.0, v1
+; GFX940-NEXT:    v_mov_b64_e32 v[2:3], s[0:1]
 ; GFX940-NEXT:    buffer_wbl2 sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    flat_atomic_add_f32 v[0:1], v2 sc1
+; GFX940-NEXT:    flat_atomic_cmpswap v0, v[2:3], v[0:1] sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v1
+; GFX940-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
+; GFX940-NEXT:    v_mov_b32_e32 v1, v0
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[2:3]
+; GFX940-NEXT:    s_cbranch_execnz .LBB2_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX940-NEXT:    s_endpgm
   %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
   ret void
 }
 
 define float @flat_atomic_fadd_f32_rtn(float* %ptr, float %data) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_rtn:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    flat_atomic_add_f32 v0, v[0:1], v2 sc0
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p0f32.f32(float* %ptr, float %data)
   ret float %ret
 }
 
 define float @flat_atomic_fadd_f32_rtn_pat(float* %ptr, float %data) {
 ; GFX940-LABEL: flat_atomic_fadd_f32_rtn_pat:
 ; GFX940:       ; %bb.0:
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    v_mov_b32_e32 v2, 4.0
+; GFX940-NEXT:    flat_load_dword v2, v[0:1]
+; GFX940-NEXT:    s_mov_b64 s[0:1], 0
+; GFX940-NEXT:  .LBB4_1: ; %atomicrmw.start
+; GFX940-NEXT:    ; =>This Inner Loop Header: Depth=1
+; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX940-NEXT:    v_mov_b32_e32 v3, v2
+; GFX940-NEXT:    v_add_f32_e32 v2, 4.0, v3
 ; GFX940-NEXT:    buffer_wbl2 sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX940-NEXT:    flat_atomic_add_f32 v0, v[0:1], v2 sc0 sc1
+; GFX940-NEXT:    flat_atomic_cmpswap v2, v[0:1], v[2:3] sc0 sc1
 ; GFX940-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX940-NEXT:    buffer_inv sc0 sc1
+; GFX940-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
+; GFX940-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
+; GFX940-NEXT:    s_andn2_b64 exec, exec, s[0:1]
+; GFX940-NEXT:    s_cbranch_execnz .LBB4_1
+; GFX940-NEXT:  ; %bb.2: ; %atomicrmw.end
+; GFX940-NEXT:    s_or_b64 exec, exec, s[0:1]
+; GFX940-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX940-NEXT:    s_setpc_b64 s[30:31]
   %ret = atomicrmw fadd float* %ptr, float 4.0 seq_cst
   ret float %ret
 }
 
 define amdgpu_kernel void @flat_atomic_fadd_v2f16_noret(<2 x half>* %ptr, <2 x half> %data) {
 ; GFX940-LABEL: flat_atomic_fadd_v2f16_noret:
 ; GFX940:       ; %bb.0:

llvm/test/CodeGen/AMDGPU/global-atomics-fp.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX900 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX908 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX90A %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX11 %s
+; RUN: not --crash llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s  2>&1 | FileCheck -enable-var-scope -check-prefixes=GFX11 %s
+
+; GFX11: LLVM ERROR: Cannot select: {{.+}}: f32,ch = AtomicLoadFAdd<(load store syncscope("agent") seq_cst
 
 define amdgpu_kernel void @global_atomic_fadd_ret_f32(float addrspace(1)* %ptr) #0 {
 ; GFX900-LABEL: global_atomic_fadd_ret_f32:
 ; GFX900:       ; %bb.0:
 ; GFX900-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; GFX900-NEXT:    s_mov_b64 s[2:3], 0
 ; GFX900-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX900-NEXT:    s_waitcnt lgkmcnt(0)

 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB0_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    global_store_dword v[0:1], v1, off
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:  .LBB0_1: ; %atomicrmw.start
-; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_add_f32_e32 v1, 4.0, v2
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT:    s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    s_cbranch_execnz .LBB0_1
-; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    global_store_b32 v[0:1], v1, off
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 seq_cst
   store float %result, float addrspace(1)* undef
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_ret_f32_ieee(float addrspace(1)* %ptr) #2 {
 ; GFX900-LABEL: global_atomic_fadd_ret_f32_ieee:
 ; GFX900:       ; %bb.0:

 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB1_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    global_store_dword v[0:1], v1, off
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_ieee:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:  .LBB1_1: ; %atomicrmw.start
-; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_add_f32_e32 v1, 4.0, v2
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT:    s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    s_cbranch_execnz .LBB1_1
-; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    global_store_b32 v[0:1], v1, off
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("agent") seq_cst
   store float %result, float addrspace(1)* undef
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_noret_f32(float addrspace(1)* %ptr) #0 {
 ; GFX900-LABEL: global_atomic_fadd_noret_f32:
 ; GFX900:       ; %bb.0:

 ; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v0, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB2_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_add_f32 v0, v1, s[0:1]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("agent") seq_cst
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_noret_f32_ieee(float addrspace(1)* %ptr) #2 {
 ; GFX900-LABEL: global_atomic_fadd_noret_f32_ieee:
 ; GFX900:       ; %bb.0:
 ; GFX900-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24

 ; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v0, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB3_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32_ieee:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 4.0
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_add_f32 v0, v1, s[0:1]
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("agent") seq_cst
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_ret_f32_agent(float addrspace(1)* %ptr) #0 {
 ; GFX900-LABEL: global_atomic_fadd_ret_f32_agent:
 ; GFX900:       ; %bb.0:
 ; GFX900-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24

 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB4_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    global_store_dword v[0:1], v1, off
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_agent:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:  .LBB4_1: ; %atomicrmw.start
-; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_add_f32_e32 v1, 4.0, v2
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT:    s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    s_cbranch_execnz .LBB4_1
-; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    global_store_b32 v[0:1], v1, off
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("agent") seq_cst
   store float %result, float addrspace(1)* undef
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_ret_f32_system(float addrspace(1)* %ptr) #0 {
 ; GFX900-LABEL: global_atomic_fadd_ret_f32_system:
 ; GFX900:       ; %bb.0:

 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB5_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    global_store_dword v[0:1], v1, off
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_ret_f32_system:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:  .LBB5_1: ; %atomicrmw.start
-; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_add_f32_e32 v1, 4.0, v2
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_cmpswap_b32 v1, v0, v[1:2], s[0:1] glc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
-; GFX11-NEXT:    s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    s_cbranch_execnz .LBB5_1
-; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    global_store_b32 v[0:1], v1, off
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("one-as") seq_cst
   store float %result, float addrspace(1)* undef
   ret void
 }
 
 define amdgpu_kernel void @global_atomic_fadd_ret_f32_wrong_subtarget(float addrspace(1)* %ptr) #1 {
 ; GCN-LABEL: global_atomic_fadd_ret_f32_wrong_subtarget:
 ; GCN:       ; %bb.0:

 ; GFX10-NEXT:    buffer_gl1_inv
 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v0, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB8_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: global_atomic_fadd_noret_f32_safe:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v1, s2
-; GFX11-NEXT:    s_mov_b32 s2, 0
-; GFX11-NEXT:  .LBB8_1: ; %atomicrmw.start
-; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
-; GFX11-NEXT:    v_add_f32_e32 v0, 4.0, v1
-; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] glc
-; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    buffer_gl0_inv
-; GFX11-NEXT:    buffer_gl1_inv
-; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v0, v1
-; GFX11-NEXT:    v_mov_b32_e32 v1, v0
-; GFX11-NEXT:    s_or_b32 s2, vcc_lo, s2
-; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s2
-; GFX11-NEXT:    s_cbranch_execnz .LBB8_1
-; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
-; GFX11-NEXT:    s_endpgm
   %result = atomicrmw fadd float addrspace(1)* %ptr, float 4.0 syncscope("agent") seq_cst
   ret void
 }
 
 define amdgpu_kernel void @infer_as_before_atomic(float* addrspace(4)* %arg) #0 {
 ; GFX900-LABEL: infer_as_before_atomic:
 ; GFX900:       ; %bb.0:
 ; GFX900-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24

 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v0, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v1, v0
 ; GFX10-NEXT:    s_or_b32 s2, vcc_lo, s2
 ; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s2
 ; GFX10-NEXT:    s_cbranch_execnz .LBB9_1
 ; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
 ; GFX10-NEXT:    s_endpgm
-;
-; GFX11-LABEL: infer_as_before_atomic:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, 1.0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    global_atomic_add_f32 v0, v1, s[0:1]
-; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
-; GFX11-NEXT:    s_endpgm
   %load = load float*, float* addrspace(4)* %arg
   %v = atomicrmw fadd float* %load, float 1.0 syncscope("agent-one-as") monotonic, align 4
   ret void
 }
 
 attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
 attributes #1 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "target-cpu"="gfx803" "target-features"="+atomic-fadd-no-rtn-insts" "amdgpu-unsafe-fp-atomics"="true" }
 attributes #2 = { "amdgpu-unsafe-fp-atomics"="true" }

llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll

 ; GFX90a-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
 ; GFX90a-NEXT:    ret float [[RES]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
 ; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
 ; GFX940-NEXT:    ret float [[RES]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
-; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
-; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
-; GFX11:       atomicrmw.start:
-; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
-; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
-; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
-; GFX11-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
-; GFX11-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
-; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
-; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
-; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
-; GFX11-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
-; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
-; GFX11:       atomicrmw.end:
-; GFX11-NEXT:    ret float [[TMP6]]
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX11-NEXT:    ret float [[RES]]
 ;
   %res = atomicrmw fadd float addrspace(1)* %ptr, float %value syncscope("wavefront") monotonic
   ret float %res
 }
 
 define double @test_atomicrmw_fadd_f64_global_unsafe(double addrspace(1)* %ptr, double %value) #0 {
 ; CI-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
 ; CI-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8

 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret float [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
 ; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
 ; GFX940-NEXT:    ret float [[RES]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
-; GFX11-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
-; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
-; GFX11:       atomicrmw.start:
-; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
-; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
-; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
-; GFX11-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
-; GFX11-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
-; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
-; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
-; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
-; GFX11-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
-; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
-; GFX11:       atomicrmw.end:
-; GFX11-NEXT:    ret float [[TMP6]]
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd float* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX11-NEXT:    ret float [[RES]]
 ;
   %res = atomicrmw fadd float* %ptr, float %value syncscope("wavefront") monotonic
   ret float %res
 }
 
 define double @test_atomicrmw_fadd_f64_flat_unsafe(double* %ptr, double %value) #0 {
 ; CI-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
 ; CI-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret float [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_flat(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
-; GFX940-NEXT:    ret float [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret float [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_flat(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret float [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
-; GFX940-NEXT:    ret float [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret float [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret void
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_no_use_ieee(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
 ; GFX940-NEXT:    ret void
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_no_use_ieee(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]

 ; GFX9-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX9-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX9-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX9-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX9:       atomicrmw.end:
 ; GFX9-NEXT:    ret void
 ;
 ; GFX908-LABEL: @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(
-; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX908-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX908-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX908:       atomicrmw.start:
+; GFX908-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX908-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX908-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX908-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX908-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX908-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX908:       atomicrmw.end:
 ; GFX908-NEXT:    ret void
 ;
 ; GFX90a-LABEL: @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(
 ; GFX90a-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; GFX90a-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX90a:       atomicrmw.start:
 ; GFX90a-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX90a-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
 ; GFX90a-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
 ; GFX90a-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
 ; GFX90a-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
 ; GFX90a-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret void
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
 ; GFX940-NEXT:    ret void
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(
-; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX11:       atomicrmw.start:
+; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
 ; GFX11-NEXT:    ret void
 ;
   %res = atomicrmw fadd float addrspace(1)* %ptr, float %value seq_cst
   ret void
 }
 
 define float @test_atomicrmw_fadd_f32_local(float addrspace(3)* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_local(

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret double [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f64_flat(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd double* [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
-; GFX940-NEXT:    ret double [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret double [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f64_flat(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret double [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f64_global(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd double addrspace(1)* [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
-; GFX940-NEXT:    ret double [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret double [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f64_global(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret float [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_agent(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("agent") monotonic, align 4
-; GFX940-NEXT:    ret float [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("agent") monotonic monotonic, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret float [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_agent(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*

 ; GFX90a-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
 ; GFX90a-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
 ; GFX90a-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GFX90a-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX90a:       atomicrmw.end:
 ; GFX90a-NEXT:    ret float [[TMP6]]
 ;
 ; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_one_as(
-; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("one-as") monotonic, align 4
-; GFX940-NEXT:    ret float [[RES]]
+; GFX940-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX940-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX940:       atomicrmw.start:
+; GFX940-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX940-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX940-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX940-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX940-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX940-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("one-as") monotonic monotonic, align 4
+; GFX940-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GFX940-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GFX940-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GFX940-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX940:       atomicrmw.end:
+; GFX940-NEXT:    ret float [[TMP6]]
 ;
 ; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_one_as(
 ; GFX11-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GFX11:       atomicrmw.start:
 ; GFX11-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GFX11-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
 ; GFX11-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*