From b1557d6e745d4ebd07c15dc13e229642da30e895 Mon Sep 17 00:00:00 2001 From: pvanhout Date: Tue, 24 Jun 2025 11:35:58 +0200 Subject: [PATCH 1/4] [AMDGPU] Use reverse iteration in CodeGenPrepare In order to make this easier, I also removed all "removeFromParent" calls from the visitors, instead adding instructions to a set of instructions to delete once the function has been visited. This avoids crashes due to functions deleting their operands. In theory we could allow functions to delete the instruction they visited (and only that one) but I think having one idiom for everything is less error-prone. Fixes #140219 --- .../Target/AMDGPU/AMDGPUCodeGenPrepare.cpp | 82 ++++------- ...egenprepare-break-large-phis-heuristics.ll | 42 +++--- .../AMDGPU/amdgpu-codegenprepare-fdiv.ll | 110 +++++++++----- llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll | 138 +++++------------- llvm/test/CodeGen/AMDGPU/uniform-select.ll | 64 ++++---- 5 files changed, 185 insertions(+), 251 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp index 24bef82464495..ee8f94e5d7be6 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp @@ -15,6 +15,7 @@ #include "AMDGPU.h" #include "AMDGPUTargetMachine.h" #include "SIModeRegisterDefaults.h" +#include "llvm/ADT/SetVector.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/TargetLibraryInfo.h" @@ -106,6 +107,7 @@ class AMDGPUCodeGenPrepareImpl bool FlowChanged = false; mutable Function *SqrtF32 = nullptr; mutable Function *LdexpF32 = nullptr; + mutable SetVector DeadVals; DenseMap BreakPhiNodesCache; @@ -281,28 +283,19 @@ bool AMDGPUCodeGenPrepareImpl::run() { BreakPhiNodesCache.clear(); bool MadeChange = false; - Function::iterator NextBB; - for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; FI = NextBB) { - BasicBlock *BB = &*FI; - NextBB = std::next(FI); - - BasicBlock::iterator Next; - for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; - I = Next) { - Next = std::next(I); - - MadeChange |= visit(*I); - - if (Next != E) { // Control flow changed - BasicBlock *NextInstBB = Next->getParent(); - if (NextInstBB != BB) { - BB = NextInstBB; - E = BB->end(); - FE = F.end(); - } - } + for (BasicBlock &BB : reverse(F)) { + for (Instruction &I : make_early_inc_range(reverse(BB))) { + if (!DeadVals.contains(&I)) + MadeChange |= visit(I); } } + + while (!DeadVals.empty()) { + RecursivelyDeleteTriviallyDeadInstructions( + DeadVals.pop_back_val(), TLI, /*MSSAU*/ nullptr, + [&](Value *V) { DeadVals.remove(V); }); + } + return MadeChange; } @@ -422,7 +415,7 @@ bool AMDGPUCodeGenPrepareImpl::replaceMulWithMul24(BinaryOperator &I) const { Value *NewVal = insertValues(Builder, Ty, ResultVals); NewVal->takeName(&I); I.replaceAllUsesWith(NewVal); - I.eraseFromParent(); + DeadVals.insert(&I); return true; } @@ -496,10 +489,10 @@ bool AMDGPUCodeGenPrepareImpl::foldBinOpIntoSelect(BinaryOperator &BO) const { FoldedT, FoldedF); NewSelect->takeName(&BO); BO.replaceAllUsesWith(NewSelect); - BO.eraseFromParent(); + DeadVals.insert(&BO); if (CastOp) - CastOp->eraseFromParent(); - Sel->eraseFromParent(); + DeadVals.insert(CastOp); + DeadVals.insert(Sel); return true; } @@ -895,7 +888,7 @@ bool AMDGPUCodeGenPrepareImpl::visitFDiv(BinaryOperator &FDiv) { if (NewVal) { FDiv.replaceAllUsesWith(NewVal); NewVal->takeName(&FDiv); - RecursivelyDeleteTriviallyDeadInstructions(&FDiv, TLI); + DeadVals.insert(&FDiv); } return true; @@ -1305,7 +1298,8 @@ within the byte are all 0. static bool tryNarrowMathIfNoOverflow(Instruction *I, const SITargetLowering *TLI, const TargetTransformInfo &TTI, - const DataLayout &DL) { + const DataLayout &DL, + SetVector &DeadVals) { unsigned Opc = I->getOpcode(); Type *OldType = I->getType(); @@ -1360,7 +1354,7 @@ static bool tryNarrowMathIfNoOverflow(Instruction *I, Value *Zext = Builder.CreateZExt(Arith, OldType); I->replaceAllUsesWith(Zext); - I->eraseFromParent(); + DeadVals.insert(I); return true; } @@ -1371,7 +1365,7 @@ bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) { if (UseMul24Intrin && replaceMulWithMul24(I)) return true; if (tryNarrowMathIfNoOverflow(&I, ST.getTargetLowering(), - TM.getTargetTransformInfo(F), DL)) + TM.getTargetTransformInfo(F), DL, DeadVals)) return true; bool Changed = false; @@ -1436,7 +1430,7 @@ bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) { if (NewDiv) { I.replaceAllUsesWith(NewDiv); - I.eraseFromParent(); + DeadVals.insert(&I); Changed = true; } } @@ -1492,7 +1486,7 @@ bool AMDGPUCodeGenPrepareImpl::visitLoadInst(LoadInst &I) { Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy); Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType()); I.replaceAllUsesWith(ValOrig); - I.eraseFromParent(); + DeadVals.insert(&I); return true; } @@ -1534,7 +1528,7 @@ bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) { Fract->takeName(&I); I.replaceAllUsesWith(Fract); - RecursivelyDeleteTriviallyDeadInstructions(&I, TLI); + DeadVals.insert(&I); return true; } @@ -1822,7 +1816,7 @@ bool AMDGPUCodeGenPrepareImpl::visitPHINode(PHINode &I) { } I.replaceAllUsesWith(Vec); - I.eraseFromParent(); + DeadVals.insert(&I); return true; } @@ -1903,7 +1897,7 @@ bool AMDGPUCodeGenPrepareImpl::visitAddrSpaceCastInst(AddrSpaceCastInst &I) { auto *Intrin = B.CreateIntrinsic( I.getType(), Intrinsic::amdgcn_addrspacecast_nonnull, {I.getOperand(0)}); I.replaceAllUsesWith(Intrin); - I.eraseFromParent(); + DeadVals.insert(&I); return true; } @@ -2000,16 +1994,10 @@ bool AMDGPUCodeGenPrepareImpl::visitFMinLike(IntrinsicInst &I) { Value *Fract = applyFractPat(Builder, FractArg); Fract->takeName(&I); I.replaceAllUsesWith(Fract); - - RecursivelyDeleteTriviallyDeadInstructions(&I, TLI); + DeadVals.insert(&I); return true; } -static bool isOneOrNegOne(const Value *Val) { - const APFloat *C; - return match(Val, m_APFloat(C)) && C->getExactLog2Abs() == 0; -} - // Expand llvm.sqrt.f32 calls with !fpmath metadata in a semi-fast way. bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) { Type *Ty = Sqrt.getType()->getScalarType(); @@ -2030,18 +2018,6 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) { if (ReqdAccuracy < 1.0f) return false; - // FIXME: This is an ugly hack for this pass using forward iteration instead - // of reverse. If it worked like a normal combiner, the rsq would form before - // we saw a sqrt call. - auto *FDiv = - dyn_cast_or_null(Sqrt.getUniqueUndroppableUser()); - if (FDiv && FDiv->getOpcode() == Instruction::FDiv && - FDiv->getFPAccuracy() >= 1.0f && - canOptimizeWithRsq(FPOp, FDiv->getFastMathFlags(), SqrtFMF) && - // TODO: We should also handle the arcp case for the fdiv with non-1 value - isOneOrNegOne(FDiv->getOperand(0))) - return false; - Value *SrcVal = Sqrt.getOperand(0); bool CanTreatAsDAZ = canIgnoreDenormalInput(SrcVal, &Sqrt); @@ -2065,7 +2041,7 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) { Value *NewSqrt = insertValues(Builder, Sqrt.getType(), ResultVals); NewSqrt->takeName(&Sqrt); Sqrt.replaceAllUsesWith(NewSqrt); - Sqrt.eraseFromParent(); + DeadVals.insert(&Sqrt); return true; } diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll index c94b33334646d..1f36902762f0b 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll @@ -726,16 +726,16 @@ define amdgpu_kernel void @used_by_unbreakable_and_breakable_phi(<5 x double> %i ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE815:%.*]] = extractelement <5 x double> [[LARGEPHI_INSERTSLICE4]], i64 4 ; CHECK-NEXT: br label [[END]] ; CHECK: end: -; CHECK-NEXT: [[TMP5:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE01]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP6:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE22]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP7:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE43]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP8:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE64]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP9:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE85]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP10:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE011]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP11:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE212]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP12:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE413]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP13:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE614]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP14:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE815]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP5:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE01]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP6:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE22]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP7:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE43]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP8:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE64]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP9:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE85]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP10:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE011]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP11:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE212]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP12:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE413]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP13:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE614]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP14:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE815]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[FINALLY]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE016:%.*]] = insertelement <5 x double> poison, double [[TMP10]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE117:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE016]], double [[TMP11]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE218:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE117]], double [[TMP12]], i64 2 @@ -746,8 +746,8 @@ define amdgpu_kernel void @used_by_unbreakable_and_breakable_phi(<5 x double> %i ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE28:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE17]], double [[TMP7]], i64 2 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE39:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE28]], double [[TMP8]], i64 3 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE410:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE39]], double [[TMP9]], i64 4 -; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE410]], ptr [[OUT]], align 1 ; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE420]], ptr [[OUT]], align 1 +; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE410]], ptr [[OUT]], align 1 ; CHECK-NEXT: ret void ; entry: @@ -1187,11 +1187,11 @@ define amdgpu_kernel void @test_breakable_chain_5_out_of_7(<5 x double> %in, ptr ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE960:%.*]] = extractelement <5 x double> [[IN]], i64 4 ; CHECK-NEXT: br i1 [[COND]], label [[END:%.*]], label [[COND5_END]] ; CHECK: cond5.end: -; CHECK-NEXT: [[TMP25:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE041]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP26:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE242]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP27:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE443]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP28:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE644]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP29:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE845]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP25:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE041]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE152]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP26:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE242]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE354]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP27:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE443]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE556]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP28:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE644]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE758]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP29:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE845]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE960]], [[COND5_TRUE]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE046:%.*]] = insertelement <5 x double> poison, double [[TMP25]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE147:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE046]], double [[TMP26]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE248:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE147]], double [[TMP27]], i64 2 @@ -1204,11 +1204,11 @@ define amdgpu_kernel void @test_breakable_chain_5_out_of_7(<5 x double> %in, ptr ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE859:%.*]] = extractelement <5 x double> [[LARGEPHI_INSERTSLICE450]], i64 4 ; CHECK-NEXT: br label [[END]] ; CHECK: end: -; CHECK-NEXT: [[TMP30:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE051]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE152]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP31:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE253]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE354]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP32:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE455]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE556]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP33:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE657]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE758]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP34:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE859]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE960]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP30:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE051]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP31:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE253]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP32:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE455]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP33:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE657]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP34:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE859]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[COND5_TRUE]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE061:%.*]] = insertelement <5 x double> poison, double [[TMP30]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE162:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE061]], double [[TMP31]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE263:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE162]], double [[TMP32]], i64 2 diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll index a4f9ce3e7350a..7ff86ac152feb 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll @@ -2160,7 +2160,22 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; IEEE-GOODFREXP-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP27]], i64 0 ; IEEE-GOODFREXP-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; IEEE-GOODFREXP-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; IEEE-GOODFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; IEEE-GOODFREXP-NEXT: [[TMP56:%.*]] = extractelement <2 x float> [[X]], i64 0 +; IEEE-GOODFREXP-NEXT: [[TMP57:%.*]] = extractelement <2 x float> [[X]], i64 1 +; IEEE-GOODFREXP-NEXT: [[TMP58:%.*]] = fcmp olt float [[TMP56]], 0x3810000000000000 +; IEEE-GOODFREXP-NEXT: [[TMP59:%.*]] = select i1 [[TMP58]], i32 32, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP60:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP56]], i32 [[TMP59]]) +; IEEE-GOODFREXP-NEXT: [[TMP61:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP60]]) +; IEEE-GOODFREXP-NEXT: [[TMP62:%.*]] = select i1 [[TMP58]], i32 -16, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP63:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP61]], i32 [[TMP62]]) +; IEEE-GOODFREXP-NEXT: [[TMP64:%.*]] = fcmp olt float [[TMP57]], 0x3810000000000000 +; IEEE-GOODFREXP-NEXT: [[TMP65:%.*]] = select i1 [[TMP64]], i32 32, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP66:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP57]], i32 [[TMP65]]) +; IEEE-GOODFREXP-NEXT: [[TMP67:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP66]]) +; IEEE-GOODFREXP-NEXT: [[TMP68:%.*]] = select i1 [[TMP64]], i32 -16, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP69:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP67]], i32 [[TMP68]]) +; IEEE-GOODFREXP-NEXT: [[TMP70:%.*]] = insertelement <2 x float> poison, float [[TMP63]], i64 0 +; IEEE-GOODFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP70]], float [[TMP69]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; IEEE-GOODFREXP-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP41:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2231,7 +2246,22 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; IEEE-BADFREXP-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP27]], i64 0 ; IEEE-BADFREXP-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; IEEE-BADFREXP-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; IEEE-BADFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; IEEE-BADFREXP-NEXT: [[TMP56:%.*]] = extractelement <2 x float> [[X]], i64 0 +; IEEE-BADFREXP-NEXT: [[TMP57:%.*]] = extractelement <2 x float> [[X]], i64 1 +; IEEE-BADFREXP-NEXT: [[TMP58:%.*]] = fcmp olt float [[TMP56]], 0x3810000000000000 +; IEEE-BADFREXP-NEXT: [[TMP59:%.*]] = select i1 [[TMP58]], i32 32, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP60:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP56]], i32 [[TMP59]]) +; IEEE-BADFREXP-NEXT: [[TMP61:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP60]]) +; IEEE-BADFREXP-NEXT: [[TMP62:%.*]] = select i1 [[TMP58]], i32 -16, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP63:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP61]], i32 [[TMP62]]) +; IEEE-BADFREXP-NEXT: [[TMP64:%.*]] = fcmp olt float [[TMP57]], 0x3810000000000000 +; IEEE-BADFREXP-NEXT: [[TMP65:%.*]] = select i1 [[TMP64]], i32 32, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP66:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP57]], i32 [[TMP65]]) +; IEEE-BADFREXP-NEXT: [[TMP67:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP66]]) +; IEEE-BADFREXP-NEXT: [[TMP68:%.*]] = select i1 [[TMP64]], i32 -16, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP69:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP67]], i32 [[TMP68]]) +; IEEE-BADFREXP-NEXT: [[TMP70:%.*]] = insertelement <2 x float> poison, float [[TMP63]], i64 0 +; IEEE-BADFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP70]], float [[TMP69]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; IEEE-BADFREXP-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP41:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2258,7 +2288,12 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[SQRT_X_NO_MD:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]) ; DAZ-NEXT: [[NO_MD:%.*]] = fdiv contract <2 x float> splat (float 1.000000e+00), [[SQRT_X_NO_MD]] ; DAZ-NEXT: store volatile <2 x float> [[NO_MD]], ptr addrspace(1) [[OUT]], align 4 -; DAZ-NEXT: [[SQRT_MD_1ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META2:![0-9]+]] +; DAZ-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP41:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP39]]) +; DAZ-NEXT: [[TMP42:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP40]]) +; DAZ-NEXT: [[TMP43:%.*]] = insertelement <2 x float> poison, float [[TMP41]], i64 0 +; DAZ-NEXT: [[SQRT_MD_1ULP:%.*]] = insertelement <2 x float> [[TMP43]], float [[TMP42]], i64 1 ; DAZ-NEXT: [[TMP1:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP]], i64 0 ; DAZ-NEXT: [[TMP2:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP]], i64 1 ; DAZ-NEXT: [[TMP3:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2276,7 +2311,9 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[SQRT_MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP12]], float [[TMP11]], i64 1 ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP_UNDEF]], i64 0 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP_UNDEF]], i64 1 -; DAZ-NEXT: [[TMP15:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) +; DAZ-NEXT: [[TMP44:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP45:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP15:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP44]]) ; DAZ-NEXT: [[TMP16:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP14]]) ; DAZ-NEXT: [[TMP17:%.*]] = extractvalue { float, i32 } [[TMP16]], 0 ; DAZ-NEXT: [[TMP18:%.*]] = extractvalue { float, i32 } [[TMP16]], 1 @@ -2290,7 +2327,12 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[TMP26:%.*]] = insertelement <2 x float> poison, float [[TMP15]], i64 0 ; DAZ-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP26]], float [[TMP25]], i64 1 ; DAZ-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; DAZ-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; DAZ-NEXT: [[TMP34:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP35:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP36:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP34]]) +; DAZ-NEXT: [[TMP37:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP35]]) +; DAZ-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP36]], i64 0 +; DAZ-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; DAZ-NEXT: [[TMP27:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; DAZ-NEXT: [[TMP28:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; DAZ-NEXT: [[TMP29:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -3200,9 +3242,13 @@ define <4 x float> @rsq_f32_vector_mixed_constant_numerator(<4 x float> %arg) { ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <4 x float> [[DENOM]], i64 1 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <4 x float> [[DENOM]], i64 2 ; DAZ-NEXT: [[TMP15:%.*]] = extractelement <4 x float> [[DENOM]], i64 3 -; DAZ-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP12]]) -; DAZ-NEXT: [[TMP17:%.*]] = fneg contract float [[TMP13]] -; DAZ-NEXT: [[TMP18:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP42:%.*]] = extractelement <4 x float> [[ARG]], i64 0 +; DAZ-NEXT: [[TMP17:%.*]] = extractelement <4 x float> [[ARG]], i64 1 +; DAZ-NEXT: [[TMP43:%.*]] = extractelement <4 x float> [[ARG]], i64 2 +; DAZ-NEXT: [[TMP44:%.*]] = extractelement <4 x float> [[ARG]], i64 3 +; DAZ-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP42]]) +; DAZ-NEXT: [[TMP45:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP45]] ; DAZ-NEXT: [[TMP19:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP14]]) ; DAZ-NEXT: [[TMP20:%.*]] = extractvalue { float, i32 } [[TMP19]], 0 ; DAZ-NEXT: [[TMP21:%.*]] = extractvalue { float, i32 } [[TMP19]], 1 @@ -3675,9 +3721,13 @@ define <4 x float> @rsq_f32_vector_mixed_constant_numerator_arcp(<4 x float> %ar ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <4 x float> [[DENOM]], i64 1 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <4 x float> [[DENOM]], i64 2 ; DAZ-NEXT: [[TMP15:%.*]] = extractelement <4 x float> [[DENOM]], i64 3 -; DAZ-NEXT: [[TMP16:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP12]]) -; DAZ-NEXT: [[TMP17:%.*]] = fneg arcp contract float [[TMP13]] -; DAZ-NEXT: [[TMP18:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP26:%.*]] = extractelement <4 x float> [[ARG]], i64 0 +; DAZ-NEXT: [[TMP17:%.*]] = extractelement <4 x float> [[ARG]], i64 1 +; DAZ-NEXT: [[TMP27:%.*]] = extractelement <4 x float> [[ARG]], i64 2 +; DAZ-NEXT: [[TMP28:%.*]] = extractelement <4 x float> [[ARG]], i64 3 +; DAZ-NEXT: [[TMP16:%.*]] = call arcp contract float @llvm.amdgcn.rsq.f32(float [[TMP26]]) +; DAZ-NEXT: [[TMP29:%.*]] = call arcp contract float @llvm.amdgcn.rsq.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP18:%.*]] = fneg arcp contract float [[TMP29]] ; DAZ-NEXT: [[TMP19:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP14]]) ; DAZ-NEXT: [[TMP20:%.*]] = fmul arcp contract float 4.000000e+00, [[TMP19]] ; DAZ-NEXT: [[TMP21:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP15]]) @@ -3850,19 +3900,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; IEEE-GOODFREXP-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; IEEE-GOODFREXP-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; IEEE-GOODFREXP-NEXT: [[TMP12:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP8]]) -; IEEE-GOODFREXP-NEXT: [[TMP13:%.*]] = extractvalue { float, i32 } [[TMP12]], 0 -; IEEE-GOODFREXP-NEXT: [[TMP14:%.*]] = extractvalue { float, i32 } [[TMP12]], 1 -; IEEE-GOODFREXP-NEXT: [[TMP15:%.*]] = sub i32 0, [[TMP14]] -; IEEE-GOODFREXP-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) -; IEEE-GOODFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP16]], i32 [[TMP15]]) -; IEEE-GOODFREXP-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP9]] -; IEEE-GOODFREXP-NEXT: [[TMP48:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP18]]) -; IEEE-GOODFREXP-NEXT: [[TMP49:%.*]] = extractvalue { float, i32 } [[TMP48]], 0 -; IEEE-GOODFREXP-NEXT: [[TMP50:%.*]] = extractvalue { float, i32 } [[TMP48]], 1 -; IEEE-GOODFREXP-NEXT: [[TMP22:%.*]] = sub i32 0, [[TMP50]] -; IEEE-GOODFREXP-NEXT: [[TMP51:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP49]]) -; IEEE-GOODFREXP-NEXT: [[TMP24:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP51]], i32 [[TMP22]]) +; IEEE-GOODFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; IEEE-GOODFREXP-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; IEEE-GOODFREXP-NEXT: [[TMP24:%.*]] = fneg contract float [[TMP13]] ; IEEE-GOODFREXP-NEXT: [[TMP29:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; IEEE-GOODFREXP-NEXT: [[TMP30:%.*]] = extractvalue { float, i32 } [[TMP29]], 0 ; IEEE-GOODFREXP-NEXT: [[TMP31:%.*]] = extractvalue { float, i32 } [[TMP29]], 1 @@ -3903,19 +3943,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; IEEE-BADFREXP-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; IEEE-BADFREXP-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; IEEE-BADFREXP-NEXT: [[TMP12:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP8]]) -; IEEE-BADFREXP-NEXT: [[TMP13:%.*]] = extractvalue { float, i32 } [[TMP12]], 0 -; IEEE-BADFREXP-NEXT: [[TMP14:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP8]]) -; IEEE-BADFREXP-NEXT: [[TMP15:%.*]] = sub i32 0, [[TMP14]] -; IEEE-BADFREXP-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) -; IEEE-BADFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP16]], i32 [[TMP15]]) -; IEEE-BADFREXP-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP9]] -; IEEE-BADFREXP-NEXT: [[TMP48:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP18]]) -; IEEE-BADFREXP-NEXT: [[TMP49:%.*]] = extractvalue { float, i32 } [[TMP48]], 0 -; IEEE-BADFREXP-NEXT: [[TMP21:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP18]]) -; IEEE-BADFREXP-NEXT: [[TMP22:%.*]] = sub i32 0, [[TMP21]] -; IEEE-BADFREXP-NEXT: [[TMP50:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP49]]) -; IEEE-BADFREXP-NEXT: [[TMP24:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP50]], i32 [[TMP22]]) +; IEEE-BADFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; IEEE-BADFREXP-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; IEEE-BADFREXP-NEXT: [[TMP24:%.*]] = fneg contract float [[TMP13]] ; IEEE-BADFREXP-NEXT: [[TMP29:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; IEEE-BADFREXP-NEXT: [[TMP30:%.*]] = extractvalue { float, i32 } [[TMP29]], 0 ; IEEE-BADFREXP-NEXT: [[TMP27:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP10]]) @@ -3956,9 +3986,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; DAZ-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; DAZ-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; DAZ-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; DAZ-NEXT: [[TMP12:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP8]]) -; DAZ-NEXT: [[TMP13:%.*]] = fneg contract float [[TMP9]] -; DAZ-NEXT: [[TMP14:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) +; DAZ-NEXT: [[TMP12:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; DAZ-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; DAZ-NEXT: [[TMP14:%.*]] = fneg contract float [[TMP13]] ; DAZ-NEXT: [[TMP15:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; DAZ-NEXT: [[TMP16:%.*]] = extractvalue { float, i32 } [[TMP15]], 0 ; DAZ-NEXT: [[TMP17:%.*]] = extractvalue { float, i32 } [[TMP15]], 1 diff --git a/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll b/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll index 3983655285e57..38239c5509318 100644 --- a/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll +++ b/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll @@ -1634,29 +1634,18 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; IR-IEEE-SDAG-LABEL: v_recip_sqrt_f32_ulp25_contract: ; IR-IEEE-SDAG: ; %bb.0: ; IR-IEEE-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-IEEE-SDAG-NEXT: s_mov_b32 s4, 0xf800000 -; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0 +; IR-IEEE-SDAG-NEXT: s_mov_b32 s4, 0x800000 ; IR-IEEE-SDAG-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc -; IR-IEEE-SDAG-NEXT: v_sqrt_f32_e32 v1, v0 -; IR-IEEE-SDAG-NEXT: v_add_i32_e64 v2, s[4:5], -1, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v3, -v2, v1, v0 -; IR-IEEE-SDAG-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5] -; IR-IEEE-SDAG-NEXT: v_add_i32_e64 v3, s[4:5], 1, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v1, -v3, v1, v0 -; IR-IEEE-SDAG-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v1 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[4:5] -; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-IEEE-SDAG-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-IEEE-SDAG-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc +; IR-IEEE-SDAG-NEXT: v_ldexp_f32_e32 v0, v0, v1 +; IR-IEEE-SDAG-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, -16, vcc +; IR-IEEE-SDAG-NEXT: v_ldexp_f32_e32 v0, v0, v1 ; IR-IEEE-SDAG-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 ; IR-IEEE-SDAG-NEXT: v_rcp_f32_e32 v2, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v3, -v1, v2, 1.0 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v2, v3, v2, v2 ; IR-IEEE-SDAG-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; IR-IEEE-SDAG-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; IR-IEEE-SDAG-NEXT: v_fma_f32 v2, v4, v2, v2 ; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v4, v3, v2 ; IR-IEEE-SDAG-NEXT: v_fma_f32 v5, -v1, v4, v3 ; IR-IEEE-SDAG-NEXT: v_fma_f32 v4, v5, v2, v4 @@ -1668,24 +1657,14 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; IR-IEEE-GISEL-LABEL: v_recip_sqrt_f32_ulp25_contract: ; IR-IEEE-GISEL: ; %bb.0: ; IR-IEEE-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v1, 0xf800000 -; IR-IEEE-GISEL-NEXT: v_mul_f32_e32 v2, 0x4f800000, v0 +; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v1, 0x800000 ; IR-IEEE-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc -; IR-IEEE-GISEL-NEXT: v_sqrt_f32_e32 v1, v0 -; IR-IEEE-GISEL-NEXT: v_add_i32_e64 v2, s[4:5], -1, v1 -; IR-IEEE-GISEL-NEXT: v_fma_f32 v3, -v2, v1, v0 -; IR-IEEE-GISEL-NEXT: v_add_i32_e64 v4, s[4:5], 1, v1 -; IR-IEEE-GISEL-NEXT: v_fma_f32 v5, -v4, v1, v0 -; IR-IEEE-GISEL-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[4:5] -; IR-IEEE-GISEL-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v5 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[4:5] -; IR-IEEE-GISEL-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-IEEE-GISEL-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc +; IR-IEEE-GISEL-NEXT: v_lshlrev_b32_e32 v1, 5, v1 +; IR-IEEE-GISEL-NEXT: v_ldexp_f32_e32 v0, v0, v1 +; IR-IEEE-GISEL-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, -16, vcc +; IR-IEEE-GISEL-NEXT: v_ldexp_f32_e32 v0, v0, v1 ; IR-IEEE-GISEL-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 ; IR-IEEE-GISEL-NEXT: v_rcp_f32_e32 v2, v1 ; IR-IEEE-GISEL-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 @@ -1705,75 +1684,24 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; CODEGEN-DAZ-NEXT: v_rsq_f32_e32 v0, v0 ; CODEGEN-DAZ-NEXT: s_setpc_b64 s[30:31] ; -; IR-DAZ-SDAG-LABEL: v_recip_sqrt_f32_ulp25_contract: -; IR-DAZ-SDAG: ; %bb.0: -; IR-DAZ-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-DAZ-SDAG-NEXT: s_mov_b32 s4, 0xf800000 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0 -; IR-DAZ-SDAG-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc -; IR-DAZ-SDAG-NEXT: v_rsq_f32_e32 v1, v0 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v2, v0, v1 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v1, 0.5, v1 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v3, -v1, v2, 0.5 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v2, v2, v3, v2 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, -v2, v2, v0 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, v1, v3, v1 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, v4, v1, v2 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-DAZ-SDAG-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-DAZ-SDAG-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc -; IR-DAZ-SDAG-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 -; IR-DAZ-SDAG-NEXT: v_rcp_f32_e32 v2, v1 -; IR-DAZ-SDAG-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 -; IR-DAZ-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, -v1, v2, 1.0 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v2, v4, v2, v2 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v4, v3, v2 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v5, -v1, v4, v3 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, v5, v2, v4 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, -v1, v4, v3 -; IR-DAZ-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; IR-DAZ-SDAG-NEXT: v_div_fmas_f32 v1, v1, v2, v4 -; IR-DAZ-SDAG-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 -; IR-DAZ-SDAG-NEXT: s_setpc_b64 s[30:31] -; -; IR-DAZ-GISEL-LABEL: v_recip_sqrt_f32_ulp25_contract: -; IR-DAZ-GISEL: ; %bb.0: -; IR-DAZ-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-DAZ-GISEL-NEXT: v_mov_b32_e32 v1, 0xf800000 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, 0x4f800000, v0 -; IR-DAZ-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc -; IR-DAZ-GISEL-NEXT: v_rsq_f32_e32 v1, v0 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, v0, v1 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v1, 0.5, v1 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v3, -v1, v2, 0.5 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v2, v2, v3, v2 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, v1, v3, v1 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v3, -v2, v2, v0 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, v3, v1, v2 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-DAZ-GISEL-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-DAZ-GISEL-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc -; IR-DAZ-GISEL-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 -; IR-DAZ-GISEL-NEXT: v_rcp_f32_e32 v2, v1 -; IR-DAZ-GISEL-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 -; IR-DAZ-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v4, -v1, v2, 1.0 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v2, v4, v2, v2 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v4, v3, v2 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v5, -v1, v4, v3 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v4, v5, v2, v4 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, -v1, v4, v3 -; IR-DAZ-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; IR-DAZ-GISEL-NEXT: v_div_fmas_f32 v1, v1, v2, v4 -; IR-DAZ-GISEL-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 -; IR-DAZ-GISEL-NEXT: s_setpc_b64 s[30:31] +; IR-DAZ-LABEL: v_recip_sqrt_f32_ulp25_contract: +; IR-DAZ: ; %bb.0: +; IR-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; IR-DAZ-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-DAZ-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; IR-DAZ-NEXT: v_rcp_f32_e32 v2, v1 +; IR-DAZ-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; IR-DAZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 +; IR-DAZ-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; IR-DAZ-NEXT: v_fma_f32 v2, v4, v2, v2 +; IR-DAZ-NEXT: v_mul_f32_e32 v4, v3, v2 +; IR-DAZ-NEXT: v_fma_f32 v5, -v1, v4, v3 +; IR-DAZ-NEXT: v_fma_f32 v4, v5, v2, v4 +; IR-DAZ-NEXT: v_fma_f32 v1, -v1, v4, v3 +; IR-DAZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 +; IR-DAZ-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; IR-DAZ-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 +; IR-DAZ-NEXT: s_setpc_b64 s[30:31] %sqrt = call contract float @llvm.sqrt.f32(float %x), !fpmath !0 %fdiv = fdiv contract float 1.0, %sqrt, !fpmath !0 ret float %fdiv diff --git a/llvm/test/CodeGen/AMDGPU/uniform-select.ll b/llvm/test/CodeGen/AMDGPU/uniform-select.ll index f001bf0d5e498..b52913fbb969f 100644 --- a/llvm/test/CodeGen/AMDGPU/uniform-select.ll +++ b/llvm/test/CodeGen/AMDGPU/uniform-select.ll @@ -20,34 +20,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX90A-NEXT: s_cmp_eq_u32 s1, 1 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s4, s3 +; GFX90A-NEXT: s_cselect_b32 s7, s3, s2 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 2 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s5, s7 +; GFX90A-NEXT: s_cselect_b32 s7, s4, s7 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 3 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s6, s7 +; GFX90A-NEXT: s_cselect_b32 s7, s5, s7 ; GFX90A-NEXT: s_or_b32 s7, s7, s0 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 1 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[10:11], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s4, s7, s4 +; GFX90A-NEXT: s_cselect_b32 s3, s7, s3 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 3 ; GFX90A-NEXT: s_cselect_b64 s[10:11], -1, 0 ; GFX90A-NEXT: s_and_b64 s[12:13], s[10:11], exec -; GFX90A-NEXT: s_cselect_b32 s6, s7, s6 +; GFX90A-NEXT: s_cselect_b32 s5, s7, s5 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 2 ; GFX90A-NEXT: s_cselect_b64 s[12:13], -1, 0 ; GFX90A-NEXT: s_and_b64 s[14:15], s[12:13], exec -; GFX90A-NEXT: s_cselect_b32 s5, s7, s5 +; GFX90A-NEXT: s_cselect_b32 s4, s7, s4 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 0 -; GFX90A-NEXT: s_cselect_b32 s3, s7, s3 +; GFX90A-NEXT: s_cselect_b32 s2, s7, s2 ; GFX90A-NEXT: s_or_b64 s[8:9], s[12:13], s[8:9] ; GFX90A-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9] ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s2, 0, s2 +; GFX90A-NEXT: s_cselect_b32 s6, 0, s6 ; GFX90A-NEXT: s_mov_b64 vcc, vcc ; GFX90A-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX90A-NEXT: ; %bb.2: ; %DummyReturnBlock @@ -68,34 +68,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX942-NEXT: s_cmp_eq_u32 s1, 1 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s4, s3 +; GFX942-NEXT: s_cselect_b32 s7, s3, s2 ; GFX942-NEXT: s_cmp_eq_u32 s1, 2 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s5, s7 +; GFX942-NEXT: s_cselect_b32 s7, s4, s7 ; GFX942-NEXT: s_cmp_eq_u32 s1, 3 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s6, s7 +; GFX942-NEXT: s_cselect_b32 s7, s5, s7 ; GFX942-NEXT: s_or_b32 s7, s7, s0 ; GFX942-NEXT: s_cmp_eq_u32 s1, 1 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[10:11], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s4, s7, s4 +; GFX942-NEXT: s_cselect_b32 s3, s7, s3 ; GFX942-NEXT: s_cmp_eq_u32 s1, 3 ; GFX942-NEXT: s_cselect_b64 s[10:11], -1, 0 ; GFX942-NEXT: s_and_b64 s[12:13], s[10:11], exec -; GFX942-NEXT: s_cselect_b32 s6, s7, s6 +; GFX942-NEXT: s_cselect_b32 s5, s7, s5 ; GFX942-NEXT: s_cmp_eq_u32 s1, 2 ; GFX942-NEXT: s_cselect_b64 s[12:13], -1, 0 ; GFX942-NEXT: s_and_b64 s[14:15], s[12:13], exec -; GFX942-NEXT: s_cselect_b32 s5, s7, s5 +; GFX942-NEXT: s_cselect_b32 s4, s7, s4 ; GFX942-NEXT: s_cmp_eq_u32 s1, 0 -; GFX942-NEXT: s_cselect_b32 s3, s7, s3 +; GFX942-NEXT: s_cselect_b32 s2, s7, s2 ; GFX942-NEXT: s_or_b64 s[8:9], s[12:13], s[8:9] ; GFX942-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9] ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s2, 0, s2 +; GFX942-NEXT: s_cselect_b32 s6, 0, s6 ; GFX942-NEXT: s_mov_b64 vcc, vcc ; GFX942-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX942-NEXT: ; %bb.2: ; %DummyReturnBlock @@ -117,34 +117,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX1030-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1030-NEXT: s_cselect_b32 s7, -1, 0 ; GFX1030-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s4, s3 +; GFX1030-NEXT: s_cselect_b32 s7, s3, s2 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s5, s7 +; GFX1030-NEXT: s_cselect_b32 s7, s4, s7 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s6, s7 +; GFX1030-NEXT: s_cselect_b32 s7, s5, s7 ; GFX1030-NEXT: s_or_b32 s7, s7, s0 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s9, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s4, s7, s4 +; GFX1030-NEXT: s_cselect_b32 s3, s7, s3 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1030-NEXT: s_cselect_b32 s9, -1, 0 ; GFX1030-NEXT: s_and_b32 s10, s9, exec_lo -; GFX1030-NEXT: s_cselect_b32 s6, s7, s6 +; GFX1030-NEXT: s_cselect_b32 s5, s7, s5 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1030-NEXT: s_cselect_b32 s10, -1, 0 ; GFX1030-NEXT: s_and_b32 s11, s10, exec_lo -; GFX1030-NEXT: s_cselect_b32 s5, s7, s5 +; GFX1030-NEXT: s_cselect_b32 s4, s7, s4 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 0 -; GFX1030-NEXT: s_cselect_b32 s3, s7, s3 +; GFX1030-NEXT: s_cselect_b32 s2, s7, s2 ; GFX1030-NEXT: s_or_b32 s7, s10, s8 ; GFX1030-NEXT: s_or_b32 s7, s9, s7 ; GFX1030-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1030-NEXT: s_cselect_b32 s2, 0, s2 +; GFX1030-NEXT: s_cselect_b32 s6, 0, s6 ; GFX1030-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX1030-NEXT: ; %bb.2: ; %DummyReturnBlock ; GFX1030-NEXT: s_endpgm @@ -166,38 +166,38 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX1100-NEXT: s_cselect_b32 s7, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s4, s3 +; GFX1100-NEXT: s_cselect_b32 s7, s3, s2 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s5, s7 +; GFX1100-NEXT: s_cselect_b32 s7, s4, s7 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s6, s7 +; GFX1100-NEXT: s_cselect_b32 s7, s5, s7 ; GFX1100-NEXT: s_or_b32 s7, s7, s0 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s9, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s4, s7, s4 +; GFX1100-NEXT: s_cselect_b32 s3, s7, s3 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1100-NEXT: s_cselect_b32 s9, -1, 0 ; GFX1100-NEXT: s_and_b32 s10, s9, exec_lo -; GFX1100-NEXT: s_cselect_b32 s6, s7, s6 +; GFX1100-NEXT: s_cselect_b32 s5, s7, s5 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1100-NEXT: s_cselect_b32 s10, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s11, s10, exec_lo -; GFX1100-NEXT: s_cselect_b32 s5, s7, s5 +; GFX1100-NEXT: s_cselect_b32 s4, s7, s4 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 0 -; GFX1100-NEXT: s_cselect_b32 s3, s7, s3 +; GFX1100-NEXT: s_cselect_b32 s2, s7, s2 ; GFX1100-NEXT: s_or_b32 s7, s10, s8 ; GFX1100-NEXT: s_or_b32 s7, s9, s7 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1100-NEXT: s_cselect_b32 s2, 0, s2 +; GFX1100-NEXT: s_cselect_b32 s6, 0, s6 ; GFX1100-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX1100-NEXT: ; %bb.2: ; %DummyReturnBlock ; GFX1100-NEXT: s_endpgm From 12ec26d1c1bd44292132d3ba3e21cc137fdf4557 Mon Sep 17 00:00:00 2001 From: pvanhout Date: Mon, 30 Jun 2025 11:45:21 +0200 Subject: [PATCH 2/4] Use vector + weakvh --- .../Target/AMDGPU/AMDGPUCodeGenPrepare.cpp | 38 +++++++++---------- 1 file changed, 19 insertions(+), 19 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp index ee8f94e5d7be6..dd14c8801fe71 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp @@ -29,6 +29,7 @@ #include "llvm/IR/IntrinsicsAMDGPU.h" #include "llvm/IR/PatternMatch.h" #include "llvm/InitializePasses.h" +#include "llvm/IR/ValueHandle.h" #include "llvm/Pass.h" #include "llvm/Support/KnownBits.h" #include "llvm/Support/KnownFPClass.h" @@ -107,7 +108,7 @@ class AMDGPUCodeGenPrepareImpl bool FlowChanged = false; mutable Function *SqrtF32 = nullptr; mutable Function *LdexpF32 = nullptr; - mutable SetVector DeadVals; + mutable SmallVector DeadVals; DenseMap BreakPhiNodesCache; @@ -285,15 +286,14 @@ bool AMDGPUCodeGenPrepareImpl::run() { for (BasicBlock &BB : reverse(F)) { for (Instruction &I : make_early_inc_range(reverse(BB))) { - if (!DeadVals.contains(&I)) + if (!isInstructionTriviallyDead(&I, TLI)) MadeChange |= visit(I); } } while (!DeadVals.empty()) { - RecursivelyDeleteTriviallyDeadInstructions( - DeadVals.pop_back_val(), TLI, /*MSSAU*/ nullptr, - [&](Value *V) { DeadVals.remove(V); }); + if (auto *I = dyn_cast_or_null(DeadVals.pop_back_val())) + RecursivelyDeleteTriviallyDeadInstructions(I, TLI); } return MadeChange; @@ -415,7 +415,7 @@ bool AMDGPUCodeGenPrepareImpl::replaceMulWithMul24(BinaryOperator &I) const { Value *NewVal = insertValues(Builder, Ty, ResultVals); NewVal->takeName(&I); I.replaceAllUsesWith(NewVal); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -489,10 +489,10 @@ bool AMDGPUCodeGenPrepareImpl::foldBinOpIntoSelect(BinaryOperator &BO) const { FoldedT, FoldedF); NewSelect->takeName(&BO); BO.replaceAllUsesWith(NewSelect); - DeadVals.insert(&BO); + DeadVals.push_back(&BO); if (CastOp) - DeadVals.insert(CastOp); - DeadVals.insert(Sel); + DeadVals.push_back(CastOp); + DeadVals.push_back(Sel); return true; } @@ -888,7 +888,7 @@ bool AMDGPUCodeGenPrepareImpl::visitFDiv(BinaryOperator &FDiv) { if (NewVal) { FDiv.replaceAllUsesWith(NewVal); NewVal->takeName(&FDiv); - DeadVals.insert(&FDiv); + DeadVals.push_back(&FDiv); } return true; @@ -1299,7 +1299,7 @@ static bool tryNarrowMathIfNoOverflow(Instruction *I, const SITargetLowering *TLI, const TargetTransformInfo &TTI, const DataLayout &DL, - SetVector &DeadVals) { + SmallVector &DeadVals) { unsigned Opc = I->getOpcode(); Type *OldType = I->getType(); @@ -1354,7 +1354,7 @@ static bool tryNarrowMathIfNoOverflow(Instruction *I, Value *Zext = Builder.CreateZExt(Arith, OldType); I->replaceAllUsesWith(Zext); - DeadVals.insert(I); + DeadVals.push_back(I); return true; } @@ -1430,7 +1430,7 @@ bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) { if (NewDiv) { I.replaceAllUsesWith(NewDiv); - DeadVals.insert(&I); + DeadVals.push_back(&I); Changed = true; } } @@ -1486,7 +1486,7 @@ bool AMDGPUCodeGenPrepareImpl::visitLoadInst(LoadInst &I) { Value *ValTrunc = Builder.CreateTrunc(WidenLoad, IntNTy); Value *ValOrig = Builder.CreateBitCast(ValTrunc, I.getType()); I.replaceAllUsesWith(ValOrig); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -1528,7 +1528,7 @@ bool AMDGPUCodeGenPrepareImpl::visitSelectInst(SelectInst &I) { Fract->takeName(&I); I.replaceAllUsesWith(Fract); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -1816,7 +1816,7 @@ bool AMDGPUCodeGenPrepareImpl::visitPHINode(PHINode &I) { } I.replaceAllUsesWith(Vec); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -1897,7 +1897,7 @@ bool AMDGPUCodeGenPrepareImpl::visitAddrSpaceCastInst(AddrSpaceCastInst &I) { auto *Intrin = B.CreateIntrinsic( I.getType(), Intrinsic::amdgcn_addrspacecast_nonnull, {I.getOperand(0)}); I.replaceAllUsesWith(Intrin); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -1994,7 +1994,7 @@ bool AMDGPUCodeGenPrepareImpl::visitFMinLike(IntrinsicInst &I) { Value *Fract = applyFractPat(Builder, FractArg); Fract->takeName(&I); I.replaceAllUsesWith(Fract); - DeadVals.insert(&I); + DeadVals.push_back(&I); return true; } @@ -2041,7 +2041,7 @@ bool AMDGPUCodeGenPrepareImpl::visitSqrt(IntrinsicInst &Sqrt) { Value *NewSqrt = insertValues(Builder, Sqrt.getType(), ResultVals); NewSqrt->takeName(&Sqrt); Sqrt.replaceAllUsesWith(NewSqrt); - DeadVals.insert(&Sqrt); + DeadVals.push_back(&Sqrt); return true; } From de7e49b0f7ba291507c558a33780942193941951 Mon Sep 17 00:00:00 2001 From: pvanhout Date: Tue, 2 Sep 2025 15:18:28 +0200 Subject: [PATCH 3/4] clang-format --- llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp index dd14c8801fe71..74ea182270e1f 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp @@ -28,8 +28,8 @@ #include "llvm/IR/InstVisitor.h" #include "llvm/IR/IntrinsicsAMDGPU.h" #include "llvm/IR/PatternMatch.h" -#include "llvm/InitializePasses.h" #include "llvm/IR/ValueHandle.h" +#include "llvm/InitializePasses.h" #include "llvm/Pass.h" #include "llvm/Support/KnownBits.h" #include "llvm/Support/KnownFPClass.h" From 4988ffce43525d2dff759ede7d69b030d6700f9c Mon Sep 17 00:00:00 2001 From: pvanhout Date: Mon, 29 Sep 2025 11:43:11 +0200 Subject: [PATCH 4/4] Ccomments --- llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp index 74ea182270e1f..8e35ba77d69aa 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp @@ -245,6 +245,8 @@ class AMDGPUCodeGenPrepareImpl Value *emitSqrtIEEE2ULP(IRBuilder<> &Builder, Value *Src, FastMathFlags FMF) const; + bool tryNarrowMathIfNoOverflow(Instruction *I); + public: bool visitFDiv(BinaryOperator &I); @@ -284,6 +286,9 @@ bool AMDGPUCodeGenPrepareImpl::run() { BreakPhiNodesCache.clear(); bool MadeChange = false; + // Need to use make_early_inc_range because integer division expansion is + // handled by Transform/Utils, and it can delete instructions such as the + // terminator of the BB. for (BasicBlock &BB : reverse(F)) { for (Instruction &I : make_early_inc_range(reverse(BB))) { if (!isInstructionTriviallyDead(&I, TLI)) @@ -1295,11 +1300,7 @@ it will create `s_and_b32 s0, s0, 0xff`. We accept this change since the non-byte load assumes the upper bits within the byte are all 0. */ -static bool tryNarrowMathIfNoOverflow(Instruction *I, - const SITargetLowering *TLI, - const TargetTransformInfo &TTI, - const DataLayout &DL, - SmallVector &DeadVals) { +bool AMDGPUCodeGenPrepareImpl::tryNarrowMathIfNoOverflow(Instruction *I) { unsigned Opc = I->getOpcode(); Type *OldType = I->getType(); @@ -1324,6 +1325,7 @@ static bool tryNarrowMathIfNoOverflow(Instruction *I, NewType = I->getType()->getWithNewBitWidth(NewBit); // Old cost + const TargetTransformInfo &TTI = TM.getTargetTransformInfo(F); InstructionCost OldCost = TTI.getArithmeticInstrCost(Opc, OldType, TTI::TCK_RecipThroughput); // New cost of new op @@ -1364,8 +1366,7 @@ bool AMDGPUCodeGenPrepareImpl::visitBinaryOperator(BinaryOperator &I) { if (UseMul24Intrin && replaceMulWithMul24(I)) return true; - if (tryNarrowMathIfNoOverflow(&I, ST.getTargetLowering(), - TM.getTargetTransformInfo(F), DL, DeadVals)) + if (tryNarrowMathIfNoOverflow(&I)) return true; bool Changed = false;