Skip to content

Commit

Permalink
[AMDGPU] Disable atomic optimization of fadd/fsub with result (#96479)
Browse files Browse the repository at this point in the history 
An atomic fadd instruction like this should return %x:

  ; value at %ptr is %x
  %r = atomicrmw fadd ptr %ptr, float %y

After atomic optimization, if %y is uniform, the result is calculated
as %r = %x + * %y * +0.0. This has a couple of problems:

1. If %y is Inf or NaN, this will return NaN instead of %x.
2. If %x is -0.0 and %y is positive, this will return +0.0 instead of
   -0.0.

Avoid these problems by disabling the "%y is uniform" path if there are
any uses of the result.
  • Loading branch information
@jayfoad
jayfoad committed Jul 3, 2024
1 parent 2b3376f commit b76dd4e
Show file tree
Hide file tree
Showing 6 changed files with 1,394 additions and 1,010 deletions.
15 changes: 14 additions & 1 deletion llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -224,7 +224,14 @@ void AMDGPUAtomicOptimizerImpl::visitAtomicRMWInst(AtomicRMWInst &I) {
return;
}

const bool ValDivergent = UA->isDivergentUse(I.getOperandUse(ValIdx));
bool ValDivergent = UA->isDivergentUse(I.getOperandUse(ValIdx));

if ((Op == AtomicRMWInst::FAdd || Op == AtomicRMWInst::FSub) &&
!I.use_empty()) {
// Disable the uniform return value calculation using fmul because it
// mishandles infinities, NaNs and signed zeros. FIXME.
ValDivergent = true;
}

// If the value operand is divergent, each lane is contributing a different
// value to the atomic calculation. We can only optimize divergent values if
Expand Down Expand Up @@ -988,6 +995,12 @@ void AMDGPUAtomicOptimizerImpl::optimizeAtomic(Instruction &I,
break;
case AtomicRMWInst::FAdd:
case AtomicRMWInst::FSub: {
// FIXME: This path is currently disabled in visitAtomicRMWInst because
// of problems calculating the first active lane of the result (where
// Mbcnt is 0):
// - If V is infinity or NaN we will return NaN instead of BroadcastI.
// - If BroadcastI is -0.0 and V is positive we will return +0.0 instead
// of -0.0.
LaneOffset = B.CreateFMul(V, Mbcnt);
break;
}
Expand Down
188 changes: 186 additions & 2 deletions llvm/test/CodeGen/AMDGPU/GlobalISel/global-atomic-fadd.f32-rtn.ll
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefix=GFX90A_GFX940 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx90a -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefixes=GFX90A_GFX940,GFX90A %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx940 -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefixes=GFX90A_GFX940,GFX940 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefix=GFX11 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -stop-after=instruction-select -amdgpu-atomic-optimizer-strategy=DPP < %s | FileCheck -check-prefix=GFX11 %s

Expand Down Expand Up @@ -149,6 +149,190 @@ define amdgpu_ps float @global_atomic_fadd_f32_rtn_atomicrmw(ptr addrspace(1) %p
}

define amdgpu_ps float @global_atomic_fadd_f32_saddr_rtn_atomicrmw(ptr addrspace(1) inreg %ptr, float %data) #0 {
; GFX90A-LABEL: name: global_atomic_fadd_f32_saddr_rtn_atomicrmw
; GFX90A: bb.1 (%ir-block.0):
; GFX90A-NEXT: successors: %bb.2(0x40000000), %bb.4(0x40000000)
; GFX90A-NEXT: liveins: $sgpr0, $sgpr1, $vgpr0
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
; GFX90A-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
; GFX90A-NEXT: [[REG_SEQUENCE:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY]], %subreg.sub0, [[COPY1]], %subreg.sub1
; GFX90A-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX90A-NEXT: [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
; GFX90A-NEXT: [[SI_PS_LIVE:%[0-9]+]]:sreg_64_xexec = SI_PS_LIVE
; GFX90A-NEXT: [[SI_IF:%[0-9]+]]:sreg_64_xexec = SI_IF [[SI_PS_LIVE]], %bb.4, implicit-def $exec, implicit-def $scc, implicit $exec
; GFX90A-NEXT: S_BRANCH %bb.2
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: bb.2 (%ir-block.5):
; GFX90A-NEXT: successors: %bb.3(0x40000000), %bb.5(0x40000000)
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[COPY3:%[0-9]+]]:sreg_64 = COPY $exec
; GFX90A-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub0
; GFX90A-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub0
; GFX90A-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub1
; GFX90A-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
; GFX90A-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[S_MOV_B32_]], %subreg.sub1
; GFX90A-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE1]].sub0
; GFX90A-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 0
; GFX90A-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY4]]
; GFX90A-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
; GFX90A-NEXT: [[V_MBCNT_LO_U32_B32_e64_:%[0-9]+]]:vgpr_32 = V_MBCNT_LO_U32_B32_e64 [[COPY8]], [[COPY9]], implicit $exec
; GFX90A-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY7]]
; GFX90A-NEXT: [[V_MBCNT_HI_U32_B32_e64_:%[0-9]+]]:vgpr_32 = V_MBCNT_HI_U32_B32_e64 [[COPY10]], [[V_MBCNT_LO_U32_B32_e64_]], implicit $exec
; GFX90A-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 -2147483648
; GFX90A-NEXT: [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_2]]
; GFX90A-NEXT: [[V_SET_INACTIVE_B32_:%[0-9]+]]:vgpr_32 = V_SET_INACTIVE_B32 [[COPY2]], [[COPY11]], implicit-def dead $scc, implicit $exec
; GFX90A-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
; GFX90A-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY12]], [[V_SET_INACTIVE_B32_]], 273, 15, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_SET_INACTIVE_B32_]], 0, [[V_MOV_B32_dpp]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp1:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY13]], [[V_ADD_F32_e64_]], 274, 15, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_1:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_MOV_B32_dpp1]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY14:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp2:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY14]], [[V_ADD_F32_e64_1]], 276, 15, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_2:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_1]], 0, [[V_MOV_B32_dpp2]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY15:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp3:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY15]], [[V_ADD_F32_e64_2]], 280, 15, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_3:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_2]], 0, [[V_MOV_B32_dpp3]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp4:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY16]], [[V_ADD_F32_e64_3]], 322, 10, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_4:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_3]], 0, [[V_MOV_B32_dpp4]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY17:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY17]], [[V_ADD_F32_e64_4]], 323, 12, 15, 0, implicit $exec
; GFX90A-NEXT: [[V_ADD_F32_e64_5:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_4]], 0, [[V_MOV_B32_dpp5]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX90A-NEXT: [[V_MOV_B32_dpp6:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY18]], [[V_ADD_F32_e64_5]], 312, 15, 15, 0, implicit $exec
; GFX90A-NEXT: [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 63
; GFX90A-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_4]]
; GFX90A-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_]]
; GFX90A-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY19]], implicit $exec
; GFX90A-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
; GFX90A-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_MBCNT_HI_U32_B32_e64_]], [[COPY20]], implicit $exec
; GFX90A-NEXT: [[SI_IF1:%[0-9]+]]:sreg_64_xexec = SI_IF [[V_CMP_EQ_U32_e64_]], %bb.5, implicit-def $exec, implicit-def $scc, implicit $exec
; GFX90A-NEXT: S_BRANCH %bb.3
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: bb.3 (%ir-block.36):
; GFX90A-NEXT: successors: %bb.5(0x80000000)
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
; GFX90A-NEXT: [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN:%[0-9]+]]:vgpr_32 = GLOBAL_ATOMIC_ADD_F32_SADDR_RTN [[V_MOV_B32_e32_]], [[STRICT_WWM]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
; GFX90A-NEXT: S_BRANCH %bb.5
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: bb.4.Flow:
; GFX90A-NEXT: successors: %bb.6(0x80000000)
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI %43, %bb.5, [[DEF]], %bb.1
; GFX90A-NEXT: SI_END_CF [[SI_IF]], implicit-def $exec, implicit-def $scc, implicit $exec
; GFX90A-NEXT: S_BRANCH %bb.6
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: bb.5 (%ir-block.39):
; GFX90A-NEXT: successors: %bb.4(0x80000000)
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]], %bb.3, [[DEF]], %bb.2
; GFX90A-NEXT: SI_END_CF [[SI_IF1]], implicit-def $exec, implicit-def $scc, implicit $exec
; GFX90A-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[PHI1]], implicit $exec
; GFX90A-NEXT: [[STRICT_WWM1:%[0-9]+]]:vgpr_32 = STRICT_WWM [[V_MOV_B32_dpp6]], implicit $exec
; GFX90A-NEXT: [[COPY21:%[0-9]+]]:vgpr_32 = COPY [[V_READFIRSTLANE_B32_]]
; GFX90A-NEXT: [[V_ADD_F32_e64_6:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[COPY21]], 0, [[STRICT_WWM1]], 0, 0, implicit $mode, implicit $exec
; GFX90A-NEXT: S_BRANCH %bb.4
; GFX90A-NEXT: {{ $}}
; GFX90A-NEXT: bb.6 (%ir-block.46):
; GFX90A-NEXT: $vgpr0 = COPY [[PHI]]
; GFX90A-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
;
; GFX940-LABEL: name: global_atomic_fadd_f32_saddr_rtn_atomicrmw
; GFX940: bb.1 (%ir-block.0):
; GFX940-NEXT: successors: %bb.2(0x40000000), %bb.4(0x40000000)
; 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: [[DEF:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
; GFX940-NEXT: [[SI_PS_LIVE:%[0-9]+]]:sreg_64_xexec = SI_PS_LIVE
; GFX940-NEXT: [[SI_IF:%[0-9]+]]:sreg_64_xexec = SI_IF [[SI_PS_LIVE]], %bb.4, implicit-def $exec, implicit-def $scc, implicit $exec
; GFX940-NEXT: S_BRANCH %bb.2
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: bb.2 (%ir-block.5):
; GFX940-NEXT: successors: %bb.3(0x40000000), %bb.5(0x40000000)
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: [[COPY3:%[0-9]+]]:sreg_64 = COPY $exec
; GFX940-NEXT: [[COPY4:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub0
; GFX940-NEXT: [[COPY5:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub0
; GFX940-NEXT: [[COPY6:%[0-9]+]]:sreg_32 = COPY [[COPY3]].sub1
; GFX940-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 0
; GFX940-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sreg_64 = REG_SEQUENCE [[COPY6]], %subreg.sub0, [[S_MOV_B32_]], %subreg.sub1
; GFX940-NEXT: [[COPY7:%[0-9]+]]:sreg_32 = COPY [[REG_SEQUENCE1]].sub0
; GFX940-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sreg_32 = S_MOV_B32 0
; GFX940-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY4]]
; GFX940-NEXT: [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
; GFX940-NEXT: [[V_MBCNT_LO_U32_B32_e64_:%[0-9]+]]:vgpr_32 = V_MBCNT_LO_U32_B32_e64 [[COPY8]], [[COPY9]], implicit $exec
; GFX940-NEXT: [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[COPY7]]
; GFX940-NEXT: [[V_MBCNT_HI_U32_B32_e64_:%[0-9]+]]:vgpr_32 = V_MBCNT_HI_U32_B32_e64 [[COPY10]], [[V_MBCNT_LO_U32_B32_e64_]], implicit $exec
; GFX940-NEXT: [[S_MOV_B32_2:%[0-9]+]]:sreg_32 = S_MOV_B32 -2147483648
; GFX940-NEXT: [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_2]]
; GFX940-NEXT: [[V_SET_INACTIVE_B32_:%[0-9]+]]:vgpr_32 = V_SET_INACTIVE_B32 [[COPY2]], [[COPY11]], implicit-def dead $scc, implicit $exec
; GFX940-NEXT: [[S_MOV_B32_3:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
; GFX940-NEXT: [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY12]], [[V_SET_INACTIVE_B32_]], 273, 15, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_SET_INACTIVE_B32_]], 0, [[V_MOV_B32_dpp]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp1:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY13]], [[V_ADD_F32_e64_]], 274, 15, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_1:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_]], 0, [[V_MOV_B32_dpp1]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY14:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp2:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY14]], [[V_ADD_F32_e64_1]], 276, 15, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_2:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_1]], 0, [[V_MOV_B32_dpp2]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY15:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp3:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY15]], [[V_ADD_F32_e64_2]], 280, 15, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_3:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_2]], 0, [[V_MOV_B32_dpp3]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY16:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp4:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY16]], [[V_ADD_F32_e64_3]], 322, 10, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_4:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_3]], 0, [[V_MOV_B32_dpp4]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY17:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp5:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY17]], [[V_ADD_F32_e64_4]], 323, 12, 15, 0, implicit $exec
; GFX940-NEXT: [[V_ADD_F32_e64_5:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[V_ADD_F32_e64_4]], 0, [[V_MOV_B32_dpp5]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: [[COPY18:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_3]]
; GFX940-NEXT: [[V_MOV_B32_dpp6:%[0-9]+]]:vgpr_32 = V_MOV_B32_dpp [[COPY18]], [[V_ADD_F32_e64_5]], 312, 15, 15, 0, implicit $exec
; GFX940-NEXT: [[S_MOV_B32_4:%[0-9]+]]:sreg_32 = S_MOV_B32 63
; GFX940-NEXT: [[V_READLANE_B32_:%[0-9]+]]:sreg_32 = V_READLANE_B32 [[V_ADD_F32_e64_5]], [[S_MOV_B32_4]]
; GFX940-NEXT: [[COPY19:%[0-9]+]]:vgpr_32 = COPY [[V_READLANE_B32_]]
; GFX940-NEXT: [[STRICT_WWM:%[0-9]+]]:vgpr_32 = STRICT_WWM [[COPY19]], implicit $exec
; GFX940-NEXT: [[COPY20:%[0-9]+]]:vgpr_32 = COPY [[S_MOV_B32_1]]
; GFX940-NEXT: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_MBCNT_HI_U32_B32_e64_]], [[COPY20]], implicit $exec
; GFX940-NEXT: [[SI_IF1:%[0-9]+]]:sreg_64_xexec = SI_IF [[V_CMP_EQ_U32_e64_]], %bb.5, implicit-def $exec, implicit-def $scc, implicit $exec
; GFX940-NEXT: S_BRANCH %bb.3
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: bb.3 (%ir-block.36):
; GFX940-NEXT: successors: %bb.5(0x80000000)
; GFX940-NEXT: {{ $}}
; 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_]], [[STRICT_WWM]], [[REG_SEQUENCE]], 0, 1, implicit $exec :: (load store syncscope("wavefront") monotonic (s32) on %ir.ptr, addrspace 1)
; GFX940-NEXT: S_BRANCH %bb.5
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: bb.4.Flow:
; GFX940-NEXT: successors: %bb.6(0x80000000)
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: [[PHI:%[0-9]+]]:vgpr_32 = PHI %42, %bb.5, [[DEF]], %bb.1
; GFX940-NEXT: SI_END_CF [[SI_IF]], implicit-def $exec, implicit-def $scc, implicit $exec
; GFX940-NEXT: S_BRANCH %bb.6
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: bb.5 (%ir-block.39):
; GFX940-NEXT: successors: %bb.4(0x80000000)
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: [[PHI1:%[0-9]+]]:vgpr_32 = PHI [[GLOBAL_ATOMIC_ADD_F32_SADDR_RTN]], %bb.3, [[DEF]], %bb.2
; GFX940-NEXT: SI_END_CF [[SI_IF1]], implicit-def $exec, implicit-def $scc, implicit $exec
; GFX940-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[PHI1]], implicit $exec
; GFX940-NEXT: [[STRICT_WWM1:%[0-9]+]]:vgpr_32 = STRICT_WWM [[V_MOV_B32_dpp6]], implicit $exec
; GFX940-NEXT: [[COPY21:%[0-9]+]]:vgpr_32 = COPY [[V_READFIRSTLANE_B32_]]
; GFX940-NEXT: [[V_ADD_F32_e64_6:%[0-9]+]]:vgpr_32 = nofpexcept V_ADD_F32_e64 0, [[COPY21]], 0, [[STRICT_WWM1]], 0, 0, implicit $mode, implicit $exec
; GFX940-NEXT: S_BRANCH %bb.4
; GFX940-NEXT: {{ $}}
; GFX940-NEXT: bb.6 (%ir-block.46):
; GFX940-NEXT: $vgpr0 = COPY [[PHI]]
; GFX940-NEXT: SI_RETURN_TO_EPILOG implicit $vgpr0
;
; GFX11-LABEL: name: global_atomic_fadd_f32_saddr_rtn_atomicrmw
; GFX11: bb.1 (%ir-block.0):
; GFX11-NEXT: successors: %bb.2(0x40000000), %bb.4(0x40000000)
Expand Down
Loading

0 comments on commit b76dd4e

Please sign in to comment.