AMDGPU: Add detailed buffer, global and flat atomic fadd tests

[Petar.Avramovic]

Summary

AMDGPU: Add detailed buffer, global and flat atomic fadd tests

Precommit for D130579 that will remove manual selection and use
patterns from td files. Tests are grouped based on target features.

All patterns have rtn and no-rtn versions.

buffer atomics patterns are selected based on the intrinsic used
(raw or struct) and the offset operand (imm or vgpr):
_offset raw with imm offset
_offen raw with vgpr offset (or large imm offset)
_idxen struct with imm offset
_bothen struct with vgpr offset (or large imm offset)

global and flat atomics are selected via intrinsic or the atomicrmw fadd.
atomicrmw tests have amdgpu-unsafe-fp-atomics=true and non-system scope
since they get expanded otherwise. atomicrmw fadd does not support vector
type, test float and double.

global atomics patterns are selected based on address type via (global or
flat) intrinsic or atomicrmw fadd with global address(addrspace(1)*).
'no suffix' vgpr addrspace(1)* address
_saddr sgpr addrspace(1)* address

flat atomics patterns are selected via (flat)intrinsic or atomicrmw fadd
with flat address (* - address space 0).

[arsenm]

LGTM. The buffer tests could use some cases that need waterfall loops

[Petar.Avramovic]

The buffer tests could use some cases that need waterfall loops

These already exist in GlobalISel/llvm.amdgcn.raw.buffer.atomic.fadd.ll and GlobalISel/llvm.amdgcn.struct.buffer.atomic.fadd.ll. I will add gfx11 run lines in another patch.
gfx11 was missing some tests and patterns, I wanted to have exactly one test for each fadd pattern that exists in td file since some patterns were missing regression test.

[arsenm]

Usually the type suffix is .type, not -type

[Petar.Avramovic]

Renamed tests to .type.
Added tests for flat intrinsic with global address (selects into global atomic).
Added test to Transforms/AtomicExpand for cases that don't expand atomic rmw fadd.

[arsenm]

LGTM

llvm/test/CodeGen/AMDGPU/GlobalISel/buffer-atomic-fadd.f32-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_offset_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_offset_no_rtn
+  ; GFX908_GFX11: bb.1 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offset_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_offen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_offen_no_rtn
+  ; GFX908_GFX11: bb.1 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_idxen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_idxen_no_rtn
+  ; GFX908_GFX11: bb.1 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_IDXEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_idxen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_IDXEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_bothen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_bothen_no_rtn
+  ; GFX908_GFX11: bb.1 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_bothen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+declare float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32 immarg)
+declare float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32, i32 immarg)

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

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX11 %s
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_offset_rtn(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offset_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFSET_RTN [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_offset_rtn
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFSET_RTN [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_offen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_offen_rtn
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_idxen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_idxen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_IDXEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_idxen_rtn
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_IDXEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_bothen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_bothen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_bothen_rtn
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32 immarg)
+declare float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/GlobalISel/buffer-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_offset_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offset_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_OFFSET [[REG_SEQUENCE]], [[REG_SEQUENCE1]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_offen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_OFFEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_idxen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_idxen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_IDXEN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_bothen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_bothen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_BOTHEN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_offset_rtn(double %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offset_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_OFFSET_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY7]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY8]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_offen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_OFFEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY8]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY9]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_idxen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_idxen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_IDXEN_RTN [[REG_SEQUENCE]], [[COPY6]], [[REG_SEQUENCE1]], [[COPY7]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY8]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY9]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_bothen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_bothen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY5]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY7]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], [[COPY8]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY9]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY10]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double, <4 x i32>, i32, i32, i32 immarg)
+declare double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/GlobalISel/buffer-atomic-fadd.v2f16-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_offset_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_offset_no_rtn
+  ; GFX908: bb.1 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offset_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_offen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_offen_no_rtn
+  ; GFX908: bb.1 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_idxen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_idxen_no_rtn
+  ; GFX908: bb.1 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_IDXEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_idxen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_IDXEN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_bothen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_bothen_no_rtn
+  ; GFX908: bb.1 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 2, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_bothen_no_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_BOTHEN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 2, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+declare <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32 immarg)
+declare <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/GlobalISel/buffer-atomic-fadd.v2f16-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_offset_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offset_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN [[COPY]], [[REG_SEQUENCE]], [[COPY5]], 0, 1, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_offen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_idxen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_idxen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN [[COPY]], [[COPY5]], [[REG_SEQUENCE]], [[COPY6]], 0, 1, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_bothen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_bothen_rtn
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4, $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sreg_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sreg_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY2]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY4]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:sreg_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN [[COPY]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], [[COPY7]], 0, 3, implicit $exec :: (volatile dereferenceable load store (<2 x s16>), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32 immarg)
+declare <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32 immarg)

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=GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX11 %s
+
+; GFX940: LLVM ERROR: unable to legalize instruction: %{{[0-9]+}}:_(s32) = G_ATOMICRMW_FADD
+
+define amdgpu_ps void @flat_atomic_fadd_f32_no_rtn_intrinsic(float* %ptr, float %data) {
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   FLAT_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX11-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) {
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_rtn_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-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)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX11-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 {
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   successors: %bb.2(0x80000000)
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+  ; GFX11-NEXT:   [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[REG_SEQUENCE]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.2.atomicrmw.start:
+  ; GFX11-NEXT:   successors: %bb.3(0x04000000), %bb.2(0x7c000000)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[PHI:%[0-9]+]]:sreg_32_xm0_xexec = PHI %10, %bb.2, [[S_MOV_B32_]], %bb.1
+  ; GFX11-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.1, %8, %bb.2
+  ; GFX11-NEXT:   %7:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[PHI1]], 0, [[COPY2]], 0, 0, implicit $mode, implicit $exec
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE %7, %subreg.sub0, [[PHI1]], %subreg.sub1
+  ; GFX11-NEXT:   [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic monotonic (s32) on %ir.2)
+  ; GFX11-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
+  ; GFX11-NEXT:   [[SI_IF_BREAK:%[0-9]+]]:sreg_32_xm0_xexec = SI_IF_BREAK [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def $scc
+  ; GFX11-NEXT:   SI_LOOP [[SI_IF_BREAK]], %bb.2, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; GFX11-NEXT:   S_BRANCH %bb.3
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.3.atomicrmw.end:
+  ; GFX11-NEXT:   [[PHI2:%[0-9]+]]:sreg_32_xm0_xexec = PHI [[SI_IF_BREAK]], %bb.2
+  ; GFX11-NEXT:   SI_END_CF [[PHI2]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; GFX11-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 {
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_rtn_atomicrmw
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   successors: %bb.2(0x80000000)
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+  ; GFX11-NEXT:   [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[REG_SEQUENCE]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.2.atomicrmw.start:
+  ; GFX11-NEXT:   successors: %bb.3(0x04000000), %bb.2(0x7c000000)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[PHI:%[0-9]+]]:sreg_32_xm0_xexec = PHI %10, %bb.2, [[S_MOV_B32_]], %bb.1
+  ; GFX11-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.1, %8, %bb.2
+  ; GFX11-NEXT:   %7:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[PHI1]], 0, [[COPY2]], 0, 0, implicit $mode, implicit $exec
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE %7, %subreg.sub0, [[PHI1]], %subreg.sub1
+  ; GFX11-NEXT:   [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic monotonic (s32) on %ir.2)
+  ; GFX11-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32_xm0_xexec = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
+  ; GFX11-NEXT:   [[SI_IF_BREAK:%[0-9]+]]:sreg_32_xm0_xexec = SI_IF_BREAK [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def $scc
+  ; GFX11-NEXT:   SI_LOOP [[SI_IF_BREAK]], %bb.2, implicit-def $exec, implicit-def $scc, implicit $exec
+  ; GFX11-NEXT:   S_BRANCH %bb.3
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.3.atomicrmw.end:
+  ; GFX11-NEXT:   [[PHI2:%[0-9]+]]:vgpr_32 = PHI [[FLAT_ATOMIC_CMPSWAP_RTN]], %bb.2
+  ; GFX11-NEXT:   [[PHI3:%[0-9]+]]:sreg_32_xm0_xexec = PHI [[SI_IF_BREAK]], %bb.2
+  ; GFX11-NEXT:   SI_END_CF [[PHI3]], implicit-def $exec, implicit-def $scc, implicit $exec
+  ; GFX11-NEXT:   $vgpr0 = COPY [[PHI2]]
+  ; GFX11-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/flat-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @flat_atomic_fadd_f64_no_rtn_intrinsic(double* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @flat_atomic_fadd_f64_rtn_intrinsic(double* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @flat_atomic_fadd_f64_no_rtn_atomicrmw(double* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 0, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @flat_atomic_fadd_f64_rtn_atomicrmw(double* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = atomicrmw fadd double* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double*, double)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/GlobalISel/flat-atomic-fadd.v2f16.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX940 %s
+
+define amdgpu_ps void @flat_atomic_fadd_v2f16_no_rtn_intrinsic(<2 x half>* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_v2f16_no_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   FLAT_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps <2 x half> @flat_atomic_fadd_v2f16_rtn_intrinsic(<2 x half>* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_v2f16_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_PK_ADD_F16_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>*, <2 x half>)

llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f32-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX908_GFX90A %s
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX908_GFX90A %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX11 %s
+
+; GFX908_GFX90A: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.flat.atomic.fadd)
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[COPY3]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_flat_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   FLAT_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_flat_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   FLAT_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   FLAT_ATOMIC_ADD_F32 [[COPY3]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   FLAT_ATOMIC_ADD_F32 [[COPY3]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_atomicrmw(float addrspace(1)* %ptr, float %data) #0 {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_atomicrmw(float addrspace(1)* inreg %ptr, float %data) #0 {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_atomicrmw
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_atomicrmw
+  ; GFX11: bb.1 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 [[COPY3]], [[COPY2]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX11-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret void
+}
+
+declare float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+declare float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

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

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX940 %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
+
+; GFX90A: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr_32(s32) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.flat.atomic.fadd)
+; GFX11: error: {{.*}} return versions of fp atomics not supported
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_flat_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-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, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_flat_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN [[COPY3]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_atomicrmw(float addrspace(1)* %ptr, float %data) #0 {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_rtn_atomicrmw
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(float addrspace(1)* inreg %ptr, float %data) #0 {
+  ; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_atomicrmw
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+declare float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64 [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_flat_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_flat_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_flat_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 [[COPY4]], [[REG_SEQUENCE1]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_flat_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN [[COPY4]], [[REG_SEQUENCE1]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY6]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_atomicrmw(double addrspace(1)* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64 [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_atomicrmw(double addrspace(1)* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN [[REG_SEQUENCE]], [[REG_SEQUENCE1]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(double addrspace(1)* inreg %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(double addrspace(1)* inreg %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.1 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY3]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN [[V_MOV_B32_e32_]], [[REG_SEQUENCE1]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY4]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+  ; GFX90A_GFX940-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[COPY5]], implicit $exec
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[V_READFIRSTLANE_B32_1]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $sgpr0, implicit $sgpr1
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)*, double)
+declare double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)*, double)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.v2f16-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX908_GFX90A %s
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX908_GFX90A %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX940 %s
+
+; GFX908_GFX90A: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr(<2 x s16>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.flat.atomic.fadd)
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16_SADDR [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_flat_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   FLAT_ATOMIC_PK_ADD_F16 [[REG_SEQUENCE]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   FLAT_ATOMIC_PK_ADD_F16 [[COPY3]], [[COPY2]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)

llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.v2f16-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: not --crash llc -global-isel -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select < %s 2>&1 | FileCheck -check-prefix=GFX90A %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select < %s | FileCheck -check-prefix=GFX940 %s
+
+; GFX90A: LLVM ERROR: cannot select: %{{[0-9]+}}:vgpr_32(<2 x s16>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.amdgcn.flat.atomic.fadd)
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_rtn_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_saddr_rtn_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_rtn_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN [[V_MOV_B32_e32_]], [[COPY2]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_rtn_flat_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_PK_ADD_F16_RTN [[REG_SEQUENCE]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_saddr_rtn_flat_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_rtn_flat_intrinsic
+  ; GFX940: bb.1 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_PK_ADD_F16_RTN [[COPY3]], [[COPY2]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (<2 x s16>) on %ir.ptr, addrspace 1)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG implicit $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)

llvm/test/CodeGen/AMDGPU/buffer-atomic-fadd.f32-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_offset_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_offset_no_rtn
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offset_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFSET [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_offen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_offen_no_rtn
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_OFFEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_idxen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_idxen_no_rtn
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_IDXEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_idxen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_IDXEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f32_bothen_no_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908_GFX11-LABEL: name: buffer_atomic_fadd_f32_bothen_no_rtn
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908_GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   BUFFER_ATOMIC_ADD_F32_BOTHEN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_bothen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F32_BOTHEN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+declare float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32 immarg)
+declare float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/buffer-atomic-fadd.f32-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX11 %s
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_offset_rtn(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offset_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFSET_RTN [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_offset_rtn
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFSET_RTN [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFSET_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_offen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_offen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_offen_rtn
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_OFFEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_OFFEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_idxen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_idxen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_IDXEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_idxen_rtn
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_IDXEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_IDXEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+define amdgpu_ps float @buffer_atomic_fadd_f32_bothen_rtn(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f32_bothen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: buffer_atomic_fadd_f32_bothen_rtn
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX11-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX11-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_ADD_F32_BOTHEN_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0)
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32 immarg)
+declare float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/buffer-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_offset_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offset_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_OFFSET killed [[COPY7]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_offen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY7]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_OFFEN killed [[COPY8]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_idxen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_idxen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY7]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_IDXEN killed [[COPY8]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_f64_bothen_no_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_bothen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $vgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY8]], %subreg.sub0, [[COPY7]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_ADD_F64_BOTHEN killed [[COPY9]], killed [[REG_SEQUENCE2]], killed [[REG_SEQUENCE]], [[COPY]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_offset_rtn(double %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offset_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_OFFSET_RTN [[COPY7]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFSET_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY8]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY9]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_offen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_offen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY7]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_OFFEN_RTN [[COPY8]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_OFFEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY9]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY10]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_idxen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_idxen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY7]], %subreg.sub0, [[COPY6]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_IDXEN_RTN [[COPY8]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_IDXEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY9]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY10]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+define amdgpu_ps double @buffer_atomic_fadd_f64_bothen_rtn(double %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_f64_bothen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr2, $vgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY8]], %subreg.sub0, [[COPY7]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY9:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN:%[0-9]+]]:vreg_64_align2 = BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN [[COPY9]], killed [[REG_SEQUENCE2]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_ATOMIC_ADD_F64_BOTHEN_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY10]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY11]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 0)
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.raw.buffer.atomic.fadd.f64(double, <4 x i32>, i32, i32, i32 immarg)
+declare double @llvm.amdgcn.struct.buffer.atomic.fadd.f64(double, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/buffer-atomic-fadd.v2f16-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_offset_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_offset_no_rtn
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offset_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFSET [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 4095, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_offen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_offen_no_rtn
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_OFFEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_idxen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_idxen_no_rtn
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_IDXEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_idxen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_IDXEN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret void
+}
+
+define amdgpu_ps void @buffer_atomic_fadd_v2f16_bothen_no_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX908-LABEL: name: buffer_atomic_fadd_v2f16_bothen_no_rtn
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_BOTHEN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_bothen_no_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   BUFFER_ATOMIC_PK_ADD_F16_BOTHEN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 2, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret void
+}
+
+declare <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32 immarg)
+declare <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32 immarg)

llvm/test/CodeGen/AMDGPU/buffer-atomic-fadd.v2f16-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_offset_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offset_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY4]], %subreg.sub0, [[COPY3]], %subreg.sub1, [[COPY2]], %subreg.sub2, [[COPY1]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN [[COPY5]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_OFFSET_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 0, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_offen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_offen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_OFFEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_idxen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_idxen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1, [[COPY3]], %subreg.sub2, [[COPY2]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN [[COPY6]], [[COPY1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_IDXEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 0)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @buffer_atomic_fadd_v2f16_bothen_rtn(<2 x half> %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
+  ; GFX90A_GFX940-LABEL: name: buffer_atomic_fadd_v2f16_bothen_rtn
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr1, $vgpr2, $sgpr4
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr4
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:sgpr_32 = COPY $sgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[COPY5]], %subreg.sub1, [[COPY4]], %subreg.sub2, [[COPY3]], %subreg.sub3
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64_align2 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN:%[0-9]+]]:vgpr_32 = BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN [[COPY7]], killed [[REG_SEQUENCE1]], killed [[REG_SEQUENCE]], [[COPY]], 0, 3, implicit $exec :: (volatile dereferenceable load store (s32), align 1, addrspace 4)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[BUFFER_ATOMIC_PK_ADD_F16_BOTHEN_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half> %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.raw.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32 immarg)
+declare <2 x half> @llvm.amdgcn.struct.buffer.atomic.fadd.v2f16(<2 x half>, <4 x i32>, i32, i32, i32, i32 immarg)

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

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX11 %s
+
+define amdgpu_ps void @flat_atomic_fadd_f32_no_rtn_intrinsic(float* %ptr, float %data) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   FLAT_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   FLAT_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX11-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) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_f32_rtn_intrinsic
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_rtn_intrinsic
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX11-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $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 {
+  ; GFX940-LABEL: name: flat_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   FLAT_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   successors: %bb.1(0x80000000)
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[COPY4]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
+  ; GFX11-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.1.atomicrmw.start:
+  ; GFX11-NEXT:   successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[PHI:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %4, %bb.1
+  ; GFX11-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.0, %3, %bb.1
+  ; GFX11-NEXT:   %13:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[PHI1]], 0, [[COPY]], 0, 0, implicit $mode, implicit $exec
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE killed %13, %subreg.sub0, [[PHI1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]]
+  ; GFX11-NEXT:   [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[COPY3]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic monotonic (s32) on %ir.2)
+  ; GFX11-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32 = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
+  ; GFX11-NEXT:   [[SI_IF_BREAK:%[0-9]+]]:sreg_32 = SI_IF_BREAK killed [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
+  ; GFX11-NEXT:   SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX11-NEXT:   S_BRANCH %bb.2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.2.atomicrmw.end:
+  ; GFX11-NEXT:   [[PHI2:%[0-9]+]]:sreg_32 = PHI [[SI_IF_BREAK]], %bb.1
+  ; GFX11-NEXT:   SI_END_CF [[PHI2]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX11-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 {
+  ; GFX940-LABEL: name: flat_atomic_fadd_f32_rtn_atomicrmw
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_ADD_F32_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  ; GFX11-LABEL: name: flat_atomic_fadd_f32_rtn_atomicrmw
+  ; GFX11: bb.0 (%ir-block.0):
+  ; GFX11-NEXT:   successors: %bb.1(0x80000000)
+  ; GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   [[COPY4:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX11-NEXT:   [[FLAT_LOAD_DWORD:%[0-9]+]]:vgpr_32 = FLAT_LOAD_DWORD [[COPY4]], 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.ptr)
+  ; GFX11-NEXT:   [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.1.atomicrmw.start:
+  ; GFX11-NEXT:   successors: %bb.2(0x04000000), %bb.1(0x7c000000)
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT:   [[PHI:%[0-9]+]]:sreg_32 = PHI [[S_MOV_B32_]], %bb.0, %4, %bb.1
+  ; GFX11-NEXT:   [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[FLAT_LOAD_DWORD]], %bb.0, %3, %bb.1
+  ; GFX11-NEXT:   %14:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[PHI1]], 0, [[COPY]], 0, 0, implicit $mode, implicit $exec
+  ; GFX11-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE killed %14, %subreg.sub0, [[PHI1]], %subreg.sub1
+  ; GFX11-NEXT:   [[COPY5:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE1]]
+  ; GFX11-NEXT:   [[FLAT_ATOMIC_CMPSWAP_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_CMPSWAP_RTN [[COPY3]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic monotonic (s32) on %ir.2)
+  ; GFX11-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_32 = V_CMP_EQ_U32_e64 [[FLAT_ATOMIC_CMPSWAP_RTN]], [[PHI1]], implicit $exec
+  ; GFX11-NEXT:   [[SI_IF_BREAK:%[0-9]+]]:sreg_32 = SI_IF_BREAK killed [[V_CMP_EQ_U32_e64_]], [[PHI]], implicit-def dead $scc
+  ; GFX11-NEXT:   SI_LOOP [[SI_IF_BREAK]], %bb.1, implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX11-NEXT:   S_BRANCH %bb.2
+  ; GFX11-NEXT: {{  $}}
+  ; GFX11-NEXT: bb.2.atomicrmw.end:
+  ; GFX11-NEXT:   [[PHI2:%[0-9]+]]:vgpr_32 = PHI [[FLAT_ATOMIC_CMPSWAP_RTN]], %bb.1
+  ; GFX11-NEXT:   [[PHI3:%[0-9]+]]:sreg_32 = PHI [[SI_IF_BREAK]], %bb.1
+  ; GFX11-NEXT:   SI_END_CF [[PHI3]], implicit-def dead $exec, implicit-def dead $scc, implicit $exec
+  ; GFX11-NEXT:   $vgpr0 = COPY [[PHI2]]
+  ; GFX11-NEXT:   SI_RETURN_TO_EPILOG $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/flat-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @flat_atomic_fadd_f64_no_rtn_intrinsic(double* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @flat_atomic_fadd_f64_rtn_intrinsic(double* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY7]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @flat_atomic_fadd_f64_no_rtn_atomicrmw(double* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   FLAT_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @flat_atomic_fadd_f64_rtn_atomicrmw(double* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: flat_atomic_fadd_f64_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[FLAT_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = FLAT_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec, implicit $flat_scr :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr)
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[FLAT_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY7]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = atomicrmw fadd double* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double*, double)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/flat-atomic-fadd.v2f16.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX940 %s
+
+define amdgpu_ps void @flat_atomic_fadd_v2f16_no_rtn_intrinsic(<2 x half>* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_v2f16_no_rtn_intrinsic
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   FLAT_ATOMIC_PK_ADD_F16 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps <2 x half> @flat_atomic_fadd_v2f16_rtn_intrinsic(<2 x half>* %ptr, <2 x half> %data) {
+  ; GFX940-LABEL: name: flat_atomic_fadd_v2f16_rtn_intrinsic
+  ; GFX940: bb.0 (%ir-block.0):
+  ; GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX940-NEXT: {{  $}}
+  ; GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX940-NEXT:   [[FLAT_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = FLAT_ATOMIC_PK_ADD_F16_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec, implicit $flat_scr :: (volatile dereferenceable load store (s32) on %ir.ptr)
+  ; GFX940-NEXT:   $vgpr0 = COPY [[FLAT_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half>*, <2 x half>)

llvm/test/CodeGen/AMDGPU/global-atomic-fadd.f32-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908_GFX11 %s
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_intrinsic
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_flat_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_flat_intrinsic
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_no_rtn_atomicrmw(float addrspace(1)* %ptr, float %data) #0 {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_f32_saddr_no_rtn_atomicrmw(float addrspace(1)* inreg %ptr, float %data) #0 {
+  ; GFX908_GFX11-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_atomicrmw
+  ; GFX908_GFX11: bb.0 (%ir-block.0):
+  ; GFX908_GFX11-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX908_GFX11-NEXT: {{  $}}
+  ; GFX908_GFX11-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908_GFX11-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908_GFX11-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908_GFX11-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908_GFX11-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX908_GFX11-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX908_GFX11-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F32_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret void
+}
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }
+
+declare float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+declare float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)

llvm/test/CodeGen/AMDGPU/global-atomic-fadd.f32-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: not --crash llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s 2>&1 | FileCheck -check-prefix=GFX11 %s
+
+; GFX11: error: {{.*}} return versions of fp atomics not supported
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_flat_intrinsic(float addrspace(1)* %ptr, float %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_flat_intrinsic(float addrspace(1)* inreg %ptr, float %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)* inreg %ptr, float %data)
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_rtn_atomicrmw(float addrspace(1)* %ptr, float %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret float %ret
+}
+
+define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(float addrspace(1)* inreg %ptr, float %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = atomicrmw fadd float addrspace(1)* %ptr, float %data syncscope("wavefront") monotonic
+  ret float %ret
+}
+
+declare float @llvm.amdgcn.global.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+declare float @llvm.amdgcn.flat.atomic.fadd.f32.p1f32.f32(float addrspace(1)*, float)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/global-atomic-fadd.f64.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY7]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY5]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_flat_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_flat_intrinsic(double addrspace(1)* %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY7]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_flat_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_flat_intrinsic(double addrspace(1)* inreg %ptr, double %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY5]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = call double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)* %ptr, double %data)
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_no_rtn_atomicrmw(double addrspace(1)* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64 killed [[COPY4]], killed [[COPY5]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_rtn_atomicrmw(double addrspace(1)* %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_RTN killed [[COPY4]], killed [[COPY5]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY7]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+define amdgpu_ps void @global_atomic_fadd_f64_saddr_no_rtn_atomicrmw(double addrspace(1)* inreg %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_no_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_ADD_F64_SADDR killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret void
+}
+
+define amdgpu_ps double @global_atomic_fadd_f64_saddr_rtn_atomicrmw(double addrspace(1)* inreg %ptr, double %data) #0 {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_f64_saddr_rtn_atomicrmw
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY1]], %subreg.sub0, [[COPY]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[COPY4:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE1]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN:%[0-9]+]]:vreg_64_align2 = GLOBAL_ATOMIC_ADD_F64_SADDR_RTN killed [[V_MOV_B32_e32_]], killed [[COPY4]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s64) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub0
+  ; GFX90A_GFX940-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[GLOBAL_ATOMIC_ADD_F64_SADDR_RTN]].sub1
+  ; GFX90A_GFX940-NEXT:   $sgpr0 = COPY [[COPY5]]
+  ; GFX90A_GFX940-NEXT:   $sgpr1 = COPY [[COPY6]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $sgpr0, $sgpr1
+  %ret = atomicrmw fadd double addrspace(1)* %ptr, double %data syncscope("wavefront") monotonic
+  ret double %ret
+}
+
+declare double @llvm.amdgcn.global.atomic.fadd.f64.p1f64.f64(double addrspace(1)*, double)
+declare double @llvm.amdgcn.flat.atomic.fadd.f64.p1f64.f64(double addrspace(1)*, double)
+
+attributes #0 = {"amdgpu-unsafe-fp-atomics"="true" }

llvm/test/CodeGen/AMDGPU/global-atomic-fadd.v2f16-no-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX908 %s
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX908-LABEL: name: global_atomic_fadd_v2f16_no_rtn_intrinsic
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX908-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX908-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX908-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_no_rtn_flat_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX908-LABEL: name: global_atomic_fadd_v2f16_no_rtn_flat_intrinsic
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]]
+  ; GFX908-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16 killed [[COPY3]], [[COPY]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+define amdgpu_ps void @global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX908-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic
+  ; GFX908: bb.0 (%ir-block.0):
+  ; GFX908-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX908-NEXT: {{  $}}
+  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX908-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX908-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX908-NEXT:   S_ENDPGM 0
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_no_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   GLOBAL_ATOMIC_PK_ADD_F16_SADDR killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 0, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   S_ENDPGM 0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret void
+}
+
+declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)

llvm/test/CodeGen/AMDGPU/global-atomic-fadd.v2f16-rtn.ll

+; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+; RUN: llc -march=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+; RUN: llc -march=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_rtn_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_saddr_rtn_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_rtn_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_rtn_flat_intrinsic(<2 x half> addrspace(1)* %ptr, <2 x half> %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[COPY3:%[0-9]+]]:vreg_64_align2 = COPY [[REG_SEQUENCE]]
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_RTN killed [[COPY3]], [[COPY]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+define amdgpu_ps <2 x half> @global_atomic_fadd_v2f16_saddr_rtn_flat_intrinsic(<2 x half> addrspace(1)* inreg %ptr, <2 x half> %data) {
+  ; GFX90A_GFX940-LABEL: name: global_atomic_fadd_v2f16_saddr_rtn_flat_intrinsic
+  ; GFX90A_GFX940: bb.0 (%ir-block.0):
+  ; GFX90A_GFX940-NEXT:   liveins: $sgpr0, $sgpr1, $vgpr0
+  ; GFX90A_GFX940-NEXT: {{  $}}
+  ; GFX90A_GFX940-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GFX90A_GFX940-NEXT:   [[COPY1:%[0-9]+]]:sgpr_32 = COPY $sgpr1
+  ; GFX90A_GFX940-NEXT:   [[COPY2:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+  ; GFX90A_GFX940-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[COPY2]], %subreg.sub0, [[COPY1]], %subreg.sub1
+  ; GFX90A_GFX940-NEXT:   [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+  ; GFX90A_GFX940-NEXT:   [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN killed [[V_MOV_B32_e32_]], [[COPY]], killed [[REG_SEQUENCE]], 0, 1, implicit $exec :: (volatile dereferenceable load store (s32) on %ir.ptr, addrspace 1)
+  ; GFX90A_GFX940-NEXT:   $vgpr0 = COPY [[GLOBAL_ATOMIC_PK_ADD_F16_SADDR_RTN]]
+  ; GFX90A_GFX940-NEXT:   SI_RETURN_TO_EPILOG $vgpr0
+  %ret = call <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)* %ptr, <2 x half> %data)
+  ret <2 x half> %ret
+}
+
+declare <2 x half> @llvm.amdgcn.global.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)
+declare <2 x half> @llvm.amdgcn.flat.atomic.fadd.v2f16.p1v2f16.v2f16(<2 x half> addrspace(1)*, <2 x half>)

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.atomic.fadd.rtn_no-rtn.ll

-; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s
-; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=amdgpu-isel < %s | FileCheck -enable-var-scope -check-prefixes=GFX11 %s
-
-; no-rtn
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFEN
-define amdgpu_ps void @raw_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
-  %voffset.add = add i32 %voffset, 4095
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFEN
-define amdgpu_ps void @raw_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 2)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFSET
-define amdgpu_ps void @raw_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_IDXEN
-define amdgpu_ps void @struct_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__4095_voffset__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 4095, i32 %soffset, i32 0)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_IDXEN
-define amdgpu_ps void @struct_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 2)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_BOTHEN
-define amdgpu_ps void @struct_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
-  %voffset.add = add i32 %voffset, 4095
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset.add, i32 %soffset, i32 0)
-  ret void
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_BOTHEN
-define amdgpu_ps void @xstruct_buffer_atomic_add_f32_noret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
-  ret void
-}
-
-
-; rtn
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFEN
-define amdgpu_ps float @raw_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
-  %voffset.add = add i32 %voffset, 4095
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset.add, i32 %soffset, i32 0)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFEN
-define amdgpu_ps float @raw_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %voffset, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %voffset, i32 %soffset, i32 2)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_OFFSET
-define amdgpu_ps float @raw_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__vgpr_voffset_4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 4095, i32 %soffset, i32 0)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_IDXEN
-define amdgpu_ps float @struct_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__4095_voffset__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 4095, i32 %soffset, i32 0)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_IDXEN
-define amdgpu_ps float @struct_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__0_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 0, i32 %soffset, i32 2)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_BOTHEN
-define amdgpu_ps float @struct_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__vgpr_voffset_plus4095__sgpr_soffset(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
-  %voffset.add = add i32 %voffset, 4095
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset.add, i32 %soffset, i32 0)
-  ret float %ret
-}
-
-; GFX11: BUFFER_ATOMIC_ADD_F32_BOTHEN
-define amdgpu_ps float @xstruct_buffer_atomic_add_f32_ret__vgpr_val__sgpr_rsrc__vgpr_voffset__sgpr_soffset_slc(float %val, <4 x i32> inreg %rsrc, i32 %vindex, i32 %voffset, i32 inreg %soffset) {
-  %ret = call float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float %val, <4 x i32> %rsrc, i32 %vindex, i32 %voffset, i32 %soffset, i32 2)
-  ret float %ret
-}
-
-declare float @llvm.amdgcn.raw.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32 immarg) #0
-declare float @llvm.amdgcn.struct.buffer.atomic.fadd.f32(float, <4 x i32>, i32, i32, i32, i32 immarg) #0
-attributes #0 = { nounwind }

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

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -atomic-expand %s | FileCheck -check-prefix=CI %s
 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -atomic-expand %s | FileCheck -check-prefix=GFX9 %s
 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -atomic-expand %s | FileCheck -check-prefix=GFX908 %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx90a -atomic-expand %s | FileCheck -check-prefix=GFX90A %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx940 -atomic-expand %s | FileCheck -check-prefix=GFX940 %s
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -atomic-expand %s | FileCheck -check-prefix=GFX11 %s
+
+define void @test_atomicrmw_fadd_f32_global_no_use_unsafe(float addrspace(1)* %ptr, float %value) #0 {
+; CI-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe(
+; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; CI-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CI-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CI-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CI-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret void
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe(
+; GFX9-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX9-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX9-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; 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_unsafe(
+; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX908-NEXT:    ret void
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX90A-NEXT:    ret void
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX940-NEXT:    ret void
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX11-NEXT:    ret void
+;
+  %res = atomicrmw fadd float addrspace(1)* %ptr, float %value syncscope("wavefront") monotonic
+  ret void
+}
+
+define float @test_atomicrmw_fadd_f32_global_unsafe(float addrspace(1)* %ptr, float %value) #0 {
+; CI-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
+; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; CI-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CI-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CI-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CI-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret float [[TMP6]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
+; GFX9-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX9-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX9-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; 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 float [[TMP6]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
+; 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]] syncscope("wavefront") monotonic monotonic, 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 float [[TMP6]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global_unsafe(
+; 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]]
+;
+  %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
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; CI-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CI-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CI-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CI-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret double [[TMP6]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
+; GFX9-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX9-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX9-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX9-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX9-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX9-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX9-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX9:       atomicrmw.end:
+; GFX9-NEXT:    ret double [[TMP6]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
+; GFX908-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; GFX908-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX908:       atomicrmw.start:
+; GFX908-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX908-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX908-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; GFX908-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX908-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX908-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX908:       atomicrmw.end:
+; GFX908-NEXT:    ret double [[TMP6]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd double addrspace(1)* [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX90A-NEXT:    ret double [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd double addrspace(1)* [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX940-NEXT:    ret double [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f64_global_unsafe(
+; 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)*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fadd double addrspace(1)* %ptr, double %value syncscope("wavefront") monotonic
+  ret double %res
+}
+
+define float @test_atomicrmw_fadd_f32_flat_unsafe(float* %ptr, float %value) #0 {
+; CI-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
+; CI-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
+; CI-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CI-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CI-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CI-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret float [[TMP6]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
+; GFX9-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX9-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX9-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 4
+; 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 float [[TMP6]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
+; GFX908-NEXT:    [[TMP1:%.*]] = load float, float* [[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* [[PTR]] to i32*
+; GFX908-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX908-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, 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 float [[TMP6]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_flat_unsafe(
+; GFX90A-NEXT:    [[TMP1:%.*]] = load float, float* [[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* [[PTR]] to i32*
+; GFX90A-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX90A-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX90A-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("wavefront") monotonic monotonic, 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 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]]
+;
+  %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
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; CI-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CI-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CI-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CI-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret double [[TMP6]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
+; GFX9-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX9-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX9-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX9-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX9-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX9-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX9-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX9:       atomicrmw.end:
+; GFX9-NEXT:    ret double [[TMP6]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
+; GFX908-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; GFX908-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX908:       atomicrmw.start:
+; GFX908-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX908-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX908-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; GFX908-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX908-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX908-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX908:       atomicrmw.end:
+; GFX908-NEXT:    ret double [[TMP6]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd double* [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX90A-NEXT:    ret double [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd double* [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX940-NEXT:    ret double [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f64_flat_unsafe(
+; 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*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fadd double* %ptr, double %value syncscope("wavefront") monotonic
+  ret double %res
+}
 
 define float @test_atomicrmw_fadd_f32_flat(float* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_flat(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; CI:       atomicrmw.start:
 ; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; CI-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[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 float [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_flat(
+; GFX90A-NEXT:    [[TMP1:%.*]] = load float, float* [[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* [[PTR]] to i32*
+; GFX90A-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; GFX90A-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GFX90A-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[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 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]]
+;
+; 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*
+; 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]] 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 float [[TMP6]]
+;
   %res = atomicrmw fadd float* %ptr, float %value seq_cst
   ret float %res
 }
 
 define float @test_atomicrmw_fadd_f32_global(float addrspace(1)* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_global(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; 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 float [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global(
+; 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 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]]
+;
+; 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)*
+; 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 float [[TMP6]]
+;
   %res = atomicrmw fadd float addrspace(1)* %ptr, float %value seq_cst
   ret float %res
 }
 
 define void @test_atomicrmw_fadd_f32_global_no_use_ieee(float addrspace(1)* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_global_no_use_ieee(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; 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_ieee(
+; 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_ieee(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(1)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; 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:%.*]]
+; 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 void @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(float addrspace(1)* %ptr, float %value) #0 {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_global_no_use_denorm_flush(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; 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:    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:    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:    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(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(3)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX9-LABEL: @test_atomicrmw_fadd_f32_local(
 ; GFX9-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(3)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
 ; GFX9-NEXT:    ret float [[RES]]
 ;
 ; GFX908-LABEL: @test_atomicrmw_fadd_f32_local(
 ; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(3)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
 ; GFX908-NEXT:    ret float [[RES]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_local(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(3)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX90A-NEXT:    ret float [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f32_local(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(3)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX940-NEXT:    ret float [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f32_local(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd float addrspace(3)* [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX11-NEXT:    ret float [[RES]]
+;
   %res = atomicrmw fadd float addrspace(3)* %ptr, float %value seq_cst
   ret float %res
 }
 
 define half @test_atomicrmw_fadd_f16_flat(half* %ptr, half %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f16_flat(
 ; CI-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; CI-NEXT:    ret half [[RES]]
 ;
 ; GFX9-LABEL: @test_atomicrmw_fadd_f16_flat(
 ; GFX9-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX9-NEXT:    ret half [[RES]]
 ;
 ; GFX908-LABEL: @test_atomicrmw_fadd_f16_flat(
 ; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX908-NEXT:    ret half [[RES]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f16_flat(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX90A-NEXT:    ret half [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f16_flat(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX940-NEXT:    ret half [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f16_flat(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX11-NEXT:    ret half [[RES]]
+;
   %res = atomicrmw fadd half* %ptr, half %value seq_cst
   ret half %res
 }
 
 define half @test_atomicrmw_fadd_f16_global(half addrspace(1)* %ptr, half %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f16_global(
 ; CI-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; CI-NEXT:    ret half [[RES]]
 ;
 ; GFX9-LABEL: @test_atomicrmw_fadd_f16_global(
 ; GFX9-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX9-NEXT:    ret half [[RES]]
 ;
 ; GFX908-LABEL: @test_atomicrmw_fadd_f16_global(
 ; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX908-NEXT:    ret half [[RES]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f16_global(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX90A-NEXT:    ret half [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f16_global(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX940-NEXT:    ret half [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f16_global(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX11-NEXT:    ret half [[RES]]
+;
   %res = atomicrmw fadd half addrspace(1)* %ptr, half %value seq_cst
   ret half %res
 }
 
 define half @test_atomicrmw_fadd_f16_local(half addrspace(3)* %ptr, half %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f16_local(
 ; CI-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; CI-NEXT:    ret half [[RES]]
 ;
 ; GFX9-LABEL: @test_atomicrmw_fadd_f16_local(
 ; GFX9-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX9-NEXT:    ret half [[RES]]
 ;
 ; GFX908-LABEL: @test_atomicrmw_fadd_f16_local(
 ; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
 ; GFX908-NEXT:    ret half [[RES]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f16_local(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX90A-NEXT:    ret half [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f16_local(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX940-NEXT:    ret half [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f16_local(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst, align 2
+; GFX11-NEXT:    ret half [[RES]]
+;
   %res = atomicrmw fadd half addrspace(3)* %ptr, half %value seq_cst
   ret half %res
 }
 
 define double @test_atomicrmw_fadd_f64_flat(double* %ptr, double %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f64_flat(
 ; CI-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
 ; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
 ; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
 ; GFX908-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX908:       atomicrmw.end:
 ; GFX908-NEXT:    ret double [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_flat(
+; GFX90A-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; GFX90A-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX90A:       atomicrmw.start:
+; GFX90A-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX90A-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX90A-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; GFX90A-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX90A-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX90A-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; 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]]
+;
+; 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*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP6]]
+;
   %res = atomicrmw fadd double* %ptr, double %value seq_cst
   ret double %res
 }
 
 define double @test_atomicrmw_fadd_f64_global(double addrspace(1)* %ptr, double %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f64_global(
 ; CI-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
 ; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
 ; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
 ; GFX908-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX908:       atomicrmw.end:
 ; GFX908-NEXT:    ret double [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_global(
+; GFX90A-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; GFX90A-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX90A:       atomicrmw.start:
+; GFX90A-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GFX90A-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; GFX90A-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; GFX90A-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX90A-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX90A-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; 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]]
+;
+; 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)*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP6]]
+;
   %res = atomicrmw fadd double addrspace(1)* %ptr, double %value seq_cst
   ret double %res
 }
 
 define double @test_atomicrmw_fadd_f64_local(double addrspace(3)* %ptr, double %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f64_local(
 ; CI-NEXT:    [[TMP1:%.*]] = load double, double addrspace(3)* [[PTR:%.*]], align 8
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(3)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
 ; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
 ; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
 ; GFX908-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GFX908:       atomicrmw.end:
 ; GFX908-NEXT:    ret double [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_local(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd double addrspace(3)* [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
+; GFX90A-NEXT:    ret double [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f64_local(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd double addrspace(3)* [[PTR:%.*]], double [[VALUE:%.*]] seq_cst, align 8
+; GFX940-NEXT:    ret double [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f64_local(
+; GFX11-NEXT:    [[TMP1:%.*]] = load double, double addrspace(3)* [[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(3)* [[PTR]] to i64 addrspace(3)*
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(3)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GFX11-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP6]]
+;
   %res = atomicrmw fadd double addrspace(3)* %ptr, double %value seq_cst
   ret double %res
 }
 
 define float @test_atomicrmw_fadd_f32_global_agent(float addrspace(1)* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_global_agent(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("agent") monotonic monotonic, 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 float [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global_agent(
+; 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]] syncscope("agent") monotonic monotonic, 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 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]]
+;
+; 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)*
+; 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("agent") 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]]
+;
   %res = atomicrmw fadd float addrspace(1)* %ptr, float %value syncscope("agent") monotonic
   ret float %res
 }
 
 define float @test_atomicrmw_fadd_f32_global_one_as(float addrspace(1)* %ptr, float %value) {
 ; CI-LABEL: @test_atomicrmw_fadd_f32_global_one_as(
 ; CI-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
 ; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]

 ; GFX908-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] syncscope("one-as") monotonic monotonic, 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 float [[TMP6]]
 ;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global_one_as(
+; 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]] syncscope("one-as") monotonic monotonic, 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 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]]
+;
+; 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)*
+; 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("one-as") 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]]
+;
   %res = atomicrmw fadd float addrspace(1)* %ptr, float %value syncscope("one-as") monotonic
   ret float %res
 }
 
 attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }