https://github.com/harrisonGPU updated https://github.com/llvm/llvm-project/pull/227737
>From da21cd29a4d5c9d12fdf2e434dde7469cf2fea0b Mon Sep 17 00:00:00 2001 From: Harrison Hao <[email protected]> Date: Wed, 30 Sep 2026 22:24:25 +0800 Subject: [PATCH 1/5] [InstCombine] Fold select of pow into select of the differing operand When both arms of a select are calls to `llvm.pow` with one use that differ in exactly one operand, sink the select into that operand: $$ \mathrm{select}(c,\ x^{y},\ x^{z}) \rightarrow x^{\mathrm{select}(c,\ y,\ z)} $$ $$ \mathrm{select}(c,\ x^{z},\ y^{z}) \rightarrow \mathrm{select}(c,\ x,\ y)^{z} $$ This removes one `pow` call. FMF are intersected and debug locations are merged. The transform is skipped when a differing operand is a constant, since that may enable a cheaper lowering (e.g. $x^2 \rightarrow x \cdot x$). --- .../InstCombine/InstCombineSelect.cpp | 35 ++++++++++++ .../InstCombine/intrinsic-select.ll | 54 ++++++++----------- 2 files changed, 58 insertions(+), 31 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index e7ade71ef562ae..594bd93a9ef76b 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -507,6 +507,39 @@ Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, return nullptr; } +/// select C, f(..., X, ...), f(..., Y, ...) --> f(..., select C, X, Y, ...) +static Instruction * +foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, + IntrinsicInst *FII, + InstCombiner::BuilderTy &Builder) { + std::optional<unsigned> DiffArgNo; + for (unsigned I = 0, E = TII->arg_size(); I != E; ++I) { + if (TII->getArgOperand(I) == FII->getArgOperand(I)) + continue; + if (DiffArgNo) + return nullptr; + DiffArgNo = I; + } + if (!DiffArgNo) + return nullptr; + + Value *TV = TII->getArgOperand(*DiffArgNo); + Value *FV = FII->getArgOperand(*DiffArgNo); + // Constant args may enable cheaper lowering, e.g. pow(X, 2.0) -> X * X. + if (isa<Constant>(TV) || isa<Constant>(FV)) + return nullptr; + + SmallVector<Value *> Args(TII->args()); + Args[*DiffArgNo] = + Builder.CreateSelect(SI.getCondition(), TV, FV, SI.getName() + ".v", &SI); + auto *NewCall = CallInst::Create(TII->getCalledFunction(), Args); + NewCall->copyIRFlags(TII); + NewCall->andIRFlags(FII); + NewCall->setDebugLoc( + DebugLoc::getMergedLocation(TII->getDebugLoc(), FII->getDebugLoc())); + return NewCall; +} + /// This transforms patterns of the form: /// select cond, intrinsic(x, ...), intrinsic(y, ...) /// into: @@ -547,6 +580,8 @@ Instruction *InstCombinerImpl::foldSelectIntrinsic(SelectInst &SI) { return replaceInstUsesWith(SI, NewCall); } + case Intrinsic::pow: + return foldSelectIntoIntrinsicArg(SI, LHSIntrinsic, RHSIntrinsic, Builder); default: return nullptr; } diff --git a/llvm/test/Transforms/InstCombine/intrinsic-select.ll b/llvm/test/Transforms/InstCombine/intrinsic-select.ll index 6f77165d910677..3dca90da5aa32f 100644 --- a/llvm/test/Transforms/InstCombine/intrinsic-select.ll +++ b/llvm/test/Transforms/InstCombine/intrinsic-select.ll @@ -688,10 +688,9 @@ define i32 @select_of_abs(i1 %cond, i32 %x, i32 %y) { define float @select_of_pow_diff_exp(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_pow_diff_exp( -; CHECK-NEXT: [[POW1:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call float @llvm.pow.f32(float [[X]], float [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1]], float [[POW2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1:%.*]], float [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[SEL]]) +; CHECK-NEXT: ret float [[SEL1]] ; %pow1 = call float @llvm.pow.f32(float %x, float %y) %pow2 = call float @llvm.pow.f32(float %x, float %z) @@ -701,10 +700,9 @@ define float @select_of_pow_diff_exp(i1 %cond, float %x, float %y, float %z) { define float @select_of_pow_diff_base(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_pow_diff_base( -; CHECK-NEXT: [[POW1:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Z:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call float @llvm.pow.f32(float [[Y:%.*]], float [[Z]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1]], float [[POW2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1:%.*]], float [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call float @llvm.pow.f32(float [[SEL]], float [[Z:%.*]]) +; CHECK-NEXT: ret float [[SEL1]] ; %pow1 = call float @llvm.pow.f32(float %x, float %z) %pow2 = call float @llvm.pow.f32(float %y, float %z) @@ -714,10 +712,9 @@ define float @select_of_pow_diff_base(i1 %cond, float %x, float %y, float %z) { define double @select_of_pow_f64(i1 %cond, double %x, double %y, double %z) { ; CHECK-LABEL: @select_of_pow_f64( -; CHECK-NEXT: [[POW1:%.*]] = call double @llvm.pow.f64(double [[X:%.*]], double [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call double @llvm.pow.f64(double [[X]], double [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], double [[POW1]], double [[POW2]] -; CHECK-NEXT: ret double [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], double [[POW1:%.*]], double [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call double @llvm.pow.f64(double [[X:%.*]], double [[SEL]]) +; CHECK-NEXT: ret double [[SEL1]] ; %pow1 = call double @llvm.pow.f64(double %x, double %y) %pow2 = call double @llvm.pow.f64(double %x, double %z) @@ -727,10 +724,9 @@ define double @select_of_pow_f64(i1 %cond, double %x, double %y, double %z) { define <2 x float> @select_of_pow_vec(<2 x i1> %cond, <2 x float> %x, <2 x float> %y, <2 x float> %z) { ; CHECK-LABEL: @select_of_pow_vec( -; CHECK-NEXT: [[POW1:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X:%.*]], <2 x float> [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X]], <2 x float> [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[POW1]], <2 x float> [[POW2]] -; CHECK-NEXT: ret <2 x float> [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[POW1:%.*]], <2 x float> [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X:%.*]], <2 x float> [[SEL]]) +; CHECK-NEXT: ret <2 x float> [[SEL1]] ; %pow1 = call <2 x float> @llvm.pow.v2f32(<2 x float> %x, <2 x float> %y) %pow2 = call <2 x float> @llvm.pow.v2f32(<2 x float> %x, <2 x float> %z) @@ -740,10 +736,9 @@ define <2 x float> @select_of_pow_vec(<2 x i1> %cond, <2 x float> %x, <2 x float define <2 x float> @select_of_pow_vec_scalar_cond(i1 %cond, <2 x float> %x, <2 x float> %y, <2 x float> %z) { ; CHECK-LABEL: @select_of_pow_vec_scalar_cond( -; CHECK-NEXT: [[POW1:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X:%.*]], <2 x float> [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X]], <2 x float> [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], <2 x float> [[POW1]], <2 x float> [[POW2]] -; CHECK-NEXT: ret <2 x float> [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], <2 x float> [[POW1:%.*]], <2 x float> [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> [[X:%.*]], <2 x float> [[SEL]]) +; CHECK-NEXT: ret <2 x float> [[SEL1]] ; %pow1 = call <2 x float> @llvm.pow.v2f32(<2 x float> %x, <2 x float> %y) %pow2 = call <2 x float> @llvm.pow.v2f32(<2 x float> %x, <2 x float> %z) @@ -754,10 +749,9 @@ define <2 x float> @select_of_pow_vec_scalar_cond(i1 %cond, <2 x float> %x, <2 x ; Only FMF common to both calls are kept. define float @select_of_pow_fmf_intersect(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_pow_fmf_intersect( -; CHECK-NEXT: [[POW1:%.*]] = call nnan ninf float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call ninf nsz float @llvm.pow.f32(float [[X]], float [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1]], float [[POW2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1:%.*]], float [[POW2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call ninf float @llvm.pow.f32(float [[X:%.*]], float [[SEL]]) +; CHECK-NEXT: ret float [[SEL1]] ; %pow1 = call nnan ninf float @llvm.pow.f32(float %x, float %y) %pow2 = call ninf nsz float @llvm.pow.f32(float %x, float %z) @@ -767,10 +761,9 @@ define float @select_of_pow_fmf_intersect(i1 %cond, float %x, float %y, float %z define float @select_of_pow_select_fmf(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_pow_select_fmf( -; CHECK-NEXT: [[POW1:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call float @llvm.pow.f32(float [[X]], float [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select nnan i1 [[COND:%.*]], float [[POW1]], float [[POW2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[Z:%.*]] = select i1 [[COND:%.*]], float [[Y:%.*]], float [[Z1:%.*]] +; CHECK-NEXT: [[POW2:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Z]]) +; CHECK-NEXT: ret float [[POW2]] ; %pow1 = call float @llvm.pow.f32(float %x, float %y) %pow2 = call float @llvm.pow.f32(float %x, float %z) @@ -780,9 +773,8 @@ define float @select_of_pow_select_fmf(i1 %cond, float %x, float %y, float %z) { define float @select_of_pow_prof(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_pow_prof( -; CHECK-NEXT: [[POW1:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[Y:%.*]]) -; CHECK-NEXT: [[POW2:%.*]] = call float @llvm.pow.f32(float [[X]], float [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[POW1]], float [[POW2]], !prof [[PROF1]] +; CHECK-NEXT: [[SEL_V:%.*]] = select i1 [[COND:%.*]], float [[Y:%.*]], float [[Z:%.*]], !prof [[PROF1]] +; CHECK-NEXT: [[SEL:%.*]] = call float @llvm.pow.f32(float [[X:%.*]], float [[SEL_V]]) ; CHECK-NEXT: ret float [[SEL]] ; %pow1 = call float @llvm.pow.f32(float %x, float %y) >From 25ab0e432af10cc0c0320f3b5ed698ffc1bcfbfd Mon Sep 17 00:00:00 2001 From: Harrison Hao <[email protected]> Date: Thu, 1 Oct 2026 02:13:35 +0800 Subject: [PATCH 2/5] Use TriviallyVectorizable --- .../Transforms/InstCombine/InstCombineSelect.cpp | 14 ++++++++++---- .../select-fcmp-fmul-zero-absorbing-value.ll | 5 ++--- 2 files changed, 12 insertions(+), 7 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index 594bd93a9ef76b..531ec4871a8ca3 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -512,6 +512,9 @@ static Instruction * foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, IntrinsicInst *FII, InstCombiner::BuilderTy &Builder) { + if (TII->getCalledFunction() != FII->getCalledFunction()) + return nullptr; + std::optional<unsigned> DiffArgNo; for (unsigned I = 0, E = TII->arg_size(); I != E; ++I) { if (TII->getArgOperand(I) == FII->getArgOperand(I)) @@ -523,6 +526,10 @@ foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, if (!DiffArgNo) return nullptr; + // A select cannot feed a parameter that requires an immediate. + if (TII->paramHasAttr(*DiffArgNo, Attribute::ImmArg)) + return nullptr; + Value *TV = TII->getArgOperand(*DiffArgNo); Value *FV = FII->getArgOperand(*DiffArgNo); // Constant args may enable cheaper lowering, e.g. pow(X, 2.0) -> X * X. @@ -535,8 +542,6 @@ foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, auto *NewCall = CallInst::Create(TII->getCalledFunction(), Args); NewCall->copyIRFlags(TII); NewCall->andIRFlags(FII); - NewCall->setDebugLoc( - DebugLoc::getMergedLocation(TII->getDebugLoc(), FII->getDebugLoc())); return NewCall; } @@ -580,9 +585,10 @@ Instruction *InstCombinerImpl::foldSelectIntrinsic(SelectInst &SI) { return replaceInstUsesWith(SI, NewCall); } - case Intrinsic::pow: - return foldSelectIntoIntrinsicArg(SI, LHSIntrinsic, RHSIntrinsic, Builder); default: + if (isTriviallyVectorizable(IID)) + return foldSelectIntoIntrinsicArg(SI, LHSIntrinsic, RHSIntrinsic, + Builder); return nullptr; } } diff --git a/llvm/test/Transforms/InstCombine/select-fcmp-fmul-zero-absorbing-value.ll b/llvm/test/Transforms/InstCombine/select-fcmp-fmul-zero-absorbing-value.ll index a3f83c9605f336..7a662f0c7fdbe6 100644 --- a/llvm/test/Transforms/InstCombine/select-fcmp-fmul-zero-absorbing-value.ll +++ b/llvm/test/Transforms/InstCombine/select-fcmp-fmul-zero-absorbing-value.ll @@ -496,11 +496,10 @@ define float @select_oeq_fmul_nonconst_unknown_fabs_or_fabs_src(float %x, float define float @select_oeq_fmul_nonconst_ppos_fabs_or_fabs_src(float %x, float %y) { ; CHECK-LABEL: define float @select_oeq_fmul_nonconst_ppos_fabs_or_fabs_src( ; CHECK-SAME: float [[X:%.*]], float [[Y:%.*]]) { -; CHECK-NEXT: [[FABS_X:%.*]] = call float @llvm.fabs.f32(float [[X]]) ; CHECK-NEXT: [[TMP1:%.*]] = fmul float [[X]], [[Y]] -; CHECK-NEXT: [[MUL_FABS_X1:%.*]] = call float @llvm.fabs.f32(float [[TMP1]]) ; CHECK-NEXT: [[X_IS_ZERO:%.*]] = fcmp oeq float [[X]], 0.000000e+00 -; CHECK-NEXT: [[SELECT:%.*]] = select i1 [[X_IS_ZERO]], float [[MUL_FABS_X1]], float [[FABS_X]] +; CHECK-NEXT: [[SELECT_V:%.*]] = select i1 [[X_IS_ZERO]], float [[TMP1]], float [[X]] +; CHECK-NEXT: [[SELECT:%.*]] = call float @llvm.fabs.f32(float [[SELECT_V]]) ; CHECK-NEXT: ret float [[SELECT]] ; %fabs.x = call float @llvm.fabs.f32(float %x) >From d3b943e2c785eef62eeb0dd64d530356e8876981 Mon Sep 17 00:00:00 2001 From: Harrison Hao <[email protected]> Date: Thu, 1 Oct 2026 02:25:10 +0800 Subject: [PATCH 3/5] Update for comments --- .../InstCombine/InstCombineSelect.cpp | 4 +++ .../InstCombine/intrinsic-select.ll | 35 ++++++++----------- 2 files changed, 19 insertions(+), 20 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index 531ec4871a8ca3..dba320922d4b3a 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -532,6 +532,10 @@ foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, Value *TV = TII->getArgOperand(*DiffArgNo); Value *FV = FII->getArgOperand(*DiffArgNo); + // powi keeps a scalar exponent even in its vector form. + if (SI.getCondition()->getType()->isVectorTy() && + !TV->getType()->isVectorTy()) + return nullptr; // Constant args may enable cheaper lowering, e.g. pow(X, 2.0) -> X * X. if (isa<Constant>(TV) || isa<Constant>(FV)) return nullptr; diff --git a/llvm/test/Transforms/InstCombine/intrinsic-select.ll b/llvm/test/Transforms/InstCombine/intrinsic-select.ll index 3dca90da5aa32f..00882a49ee9fed 100644 --- a/llvm/test/Transforms/InstCombine/intrinsic-select.ll +++ b/llvm/test/Transforms/InstCombine/intrinsic-select.ll @@ -913,10 +913,9 @@ define float @select_of_pow_negative_not_intrinsic(i1 %cond, float %x, float %y, define <2 x float> @select_of_sqrt_vec(<2 x i1> %cond, <2 x float> %x, <2 x float> %y) { ; CHECK-LABEL: @select_of_sqrt_vec( -; CHECK-NEXT: [[SQRT1:%.*]] = call <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X:%.*]]) -; CHECK-NEXT: [[SQRT2:%.*]] = call <2 x float> @llvm.sqrt.v2f32(<2 x float> [[Y:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[SQRT1]], <2 x float> [[SQRT2]] -; CHECK-NEXT: ret <2 x float> [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[SQRT1:%.*]], <2 x float> [[SQRT2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call <2 x float> @llvm.sqrt.v2f32(<2 x float> [[SEL]]) +; CHECK-NEXT: ret <2 x float> [[SEL1]] ; %sqrt1 = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %x) %sqrt2 = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %y) @@ -926,10 +925,9 @@ define <2 x float> @select_of_sqrt_vec(<2 x i1> %cond, <2 x float> %x, <2 x floa define float @select_of_fma(i1 %cond, float %x, float %y, float %z, float %w) { ; CHECK-LABEL: @select_of_fma( -; CHECK-NEXT: [[FMA1:%.*]] = call float @llvm.fma.f32(float [[X:%.*]], float [[Y:%.*]], float [[Z:%.*]]) -; CHECK-NEXT: [[FMA2:%.*]] = call float @llvm.fma.f32(float [[X]], float [[Y]], float [[W:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[FMA1]], float [[FMA2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[FMA1:%.*]], float [[FMA2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call float @llvm.fma.f32(float [[X:%.*]], float [[Y:%.*]], float [[SEL]]) +; CHECK-NEXT: ret float [[SEL1]] ; %fma1 = call float @llvm.fma.f32(float %x, float %y, float %z) %fma2 = call float @llvm.fma.f32(float %x, float %y, float %w) @@ -939,10 +937,9 @@ define float @select_of_fma(i1 %cond, float %x, float %y, float %z, float %w) { define float @select_of_copysign(i1 %cond, float %x, float %y, float %z) { ; CHECK-LABEL: @select_of_copysign( -; CHECK-NEXT: [[COPYSIGN1:%.*]] = call float @llvm.copysign.f32(float [[X:%.*]], float [[Y:%.*]]) -; CHECK-NEXT: [[COPYSIGN2:%.*]] = call float @llvm.copysign.f32(float [[X]], float [[Z:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[COPYSIGN1]], float [[COPYSIGN2]] -; CHECK-NEXT: ret float [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], float [[COPYSIGN1:%.*]], float [[COPYSIGN2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call float @llvm.copysign.f32(float [[X:%.*]], float [[SEL]]) +; CHECK-NEXT: ret float [[SEL1]] ; %copysign1 = call float @llvm.copysign.f32(float %x, float %y) %copysign2 = call float @llvm.copysign.f32(float %x, float %z) @@ -952,10 +949,9 @@ define float @select_of_copysign(i1 %cond, float %x, float %y, float %z) { define i32 @select_of_fshl(i1 %cond, i32 %x, i32 %y, i32 %a, i32 %b) { ; CHECK-LABEL: @select_of_fshl( -; CHECK-NEXT: [[FSHL1:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[A:%.*]]) -; CHECK-NEXT: [[FSHL2:%.*]] = call i32 @llvm.fshl.i32(i32 [[X]], i32 [[Y]], i32 [[B:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], i32 [[FSHL1]], i32 [[FSHL2]] -; CHECK-NEXT: ret i32 [[SEL]] +; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], i32 [[FSHL1:%.*]], i32 [[FSHL2:%.*]] +; CHECK-NEXT: [[SEL1:%.*]] = call i32 @llvm.fshl.i32(i32 [[X:%.*]], i32 [[Y:%.*]], i32 [[SEL]]) +; CHECK-NEXT: ret i32 [[SEL1]] ; %fshl1 = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %a) %fshl2 = call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %b) @@ -965,10 +961,9 @@ define i32 @select_of_fshl(i1 %cond, i32 %x, i32 %y, i32 %a, i32 %b) { define <2 x float> @select_of_powi_vec_scalar_cond(i1 %cond, <2 x float> %x, i32 %n, i32 %m) { ; CHECK-LABEL: @select_of_powi_vec_scalar_cond( -; CHECK-NEXT: [[POWI1:%.*]] = call <2 x float> @llvm.powi.v2f32.i32(<2 x float> [[X:%.*]], i32 [[N:%.*]]) -; CHECK-NEXT: [[POWI2:%.*]] = call <2 x float> @llvm.powi.v2f32.i32(<2 x float> [[X]], i32 [[M:%.*]]) -; CHECK-NEXT: [[SEL:%.*]] = select i1 [[COND:%.*]], <2 x float> [[POWI1]], <2 x float> [[POWI2]] -; CHECK-NEXT: ret <2 x float> [[SEL]] +; CHECK-NEXT: [[M:%.*]] = select i1 [[COND:%.*]], i32 [[N:%.*]], i32 [[M1:%.*]] +; CHECK-NEXT: [[POWI2:%.*]] = call <2 x float> @llvm.powi.v2f32.i32(<2 x float> [[X:%.*]], i32 [[M]]) +; CHECK-NEXT: ret <2 x float> [[POWI2]] ; %powi1 = call <2 x float> @llvm.powi.v2f32.i32(<2 x float> %x, i32 %n) %powi2 = call <2 x float> @llvm.powi.v2f32.i32(<2 x float> %x, i32 %m) >From 45f0ed55950c02d83c893719e02639b54ff57be8 Mon Sep 17 00:00:00 2001 From: Harrison Hao <[email protected]> Date: Thu, 1 Oct 2026 09:50:50 +0800 Subject: [PATCH 4/5] Fix lit test issue. --- .../AMDGPU/amdgpu-simplify-libcall-pow.ll | 20 ++++--- .../AMDGPU/amdgpu-simplify-libcall-rootn.ll | 54 +++++++------------ llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll | 6 +-- 3 files changed, 29 insertions(+), 51 deletions(-) diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll index ecc9bc6f75d772..a91f127ad43b59 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow.ll @@ -4181,17 +4181,15 @@ define float @test_pow_afn_f32__y_poison(float %x) { ; NOPRELINK-NEXT: [[TMP1:%.*]] = call afn float @llvm.fabs.f32(float [[X]]) ; NOPRELINK-NEXT: [[TMP2:%.*]] = call afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = call afn float @llvm.exp2.f32(float [[TMP2]]) -; NOPRELINK-NEXT: [[TMP4:%.*]] = call afn float @llvm.copysign.f32(float [[TMP3]], float [[X]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call afn float @llvm.fabs.f32(float [[X]]) -; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp afn oeq float [[TMP5]], +inf -; NOPRELINK-NEXT: [[TMP7:%.*]] = fcmp afn oeq float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]] -; NOPRELINK-NEXT: [[TMP9:%.*]] = select afn i1 [[TMP7]], float 0.000000e+00, float +inf -; NOPRELINK-NEXT: [[TMP10:%.*]] = call afn float @llvm.copysign.f32(float [[TMP9]], float [[X]]) -; NOPRELINK-NEXT: [[TMP11:%.*]] = select afn i1 [[TMP8]], float [[TMP10]], float [[TMP4]] -; NOPRELINK-NEXT: [[TMP12:%.*]] = fcmp afn uno float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP13:%.*]] = select afn i1 [[TMP12]], float +qnan, float [[TMP11]] -; NOPRELINK-NEXT: ret float [[TMP13]] +; NOPRELINK-NEXT: [[TMP4:%.*]] = call afn float @llvm.fabs.f32(float [[X]]) +; NOPRELINK-NEXT: [[TMP5:%.*]] = fcmp afn oeq float [[TMP4]], +inf +; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp afn oeq float [[X]], 0.000000e+00 +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP5]], float +inf, float [[TMP3]] +; NOPRELINK-NEXT: [[DOTV:%.*]] = select i1 [[TMP6]], float 0.000000e+00, float [[TMP7]] +; NOPRELINK-NEXT: [[TMP8:%.*]] = call afn float @llvm.copysign.f32(float [[DOTV]], float [[X]]) +; NOPRELINK-NEXT: [[TMP9:%.*]] = fcmp afn uno float [[X]], 0.000000e+00 +; NOPRELINK-NEXT: [[TMP10:%.*]] = select afn i1 [[TMP9]], float +qnan, float [[TMP8]] +; NOPRELINK-NEXT: ret float [[TMP10]] ; %pow = tail call afn float @_Z3powff(float %x, float poison) ret float %pow diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll index 805e510bde37b0..fa0fa173d88925 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-rootn.ll @@ -1360,12 +1360,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_3(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], f0x3EAAAAAB ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn oeq float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float 0.000000e+00, float [[TMP4]] ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP8]], float [[TMP5]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 3) ret float %call @@ -1383,12 +1381,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_neg3(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], f0xBEAAAAAB ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn une float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float [[TMP4]], float +inf ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP5]], float [[TMP8]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 -3) ret float %call @@ -1450,12 +1446,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_5(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], 2.000000e-01 ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn oeq float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float 0.000000e+00, float [[TMP4]] ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP8]], float [[TMP5]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 5) ret float %call @@ -1473,12 +1467,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_neg5(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], -2.000000e-01 ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn une float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float [[TMP4]], float +inf ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP5]], float [[TMP8]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 -5) ret float %call @@ -1496,12 +1488,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_7(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], f0x3E124925 ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn oeq float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float 0.000000e+00, float [[TMP4]] ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP8]], float [[TMP5]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 7) ret float %call @@ -1519,12 +1509,10 @@ define float @test_rootn_afn_nnan_ninf_f32__y_neg7(float %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn float @llvm.log2.f32(float [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn float [[TMP2]], f0xBE124925 ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn float @llvm.exp2.f32(float [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP4]], float [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn une float [[X]], 0.000000e+00 -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn i1 [[TMP6]], float 0.000000e+00, float +inf +; NOPRELINK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], float [[TMP4]], float +inf ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn float @llvm.copysign.f32(float [[TMP7]], float [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn i1 [[TMP6]], float [[TMP5]], float [[TMP8]] -; NOPRELINK-NEXT: ret float [[TMP9]] +; NOPRELINK-NEXT: ret float [[TMP8]] ; %call = tail call nnan ninf afn float @_Z5rootnfi(float %x, i32 -7) ret float %call @@ -1588,12 +1576,10 @@ define <2 x float> @test_rootn_afn_nnan_ninf_v2f32__y_3(<2 x float> %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn <2 x float> @llvm.log2.v2f32(<2 x float> [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn <2 x float> [[TMP2]], splat (float f0x3EAAAAAB) ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn <2 x float> @llvm.exp2.v2f32(<2 x float> [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP4]], <2 x float> [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn oeq <2 x float> [[X]], zeroinitializer -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> zeroinitializer, <2 x float> splat (float +inf) +; NOPRELINK-NEXT: [[TMP7:%.*]] = select <2 x i1> [[TMP6]], <2 x float> zeroinitializer, <2 x float> [[TMP4]] ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP7]], <2 x float> [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> [[TMP8]], <2 x float> [[TMP5]] -; NOPRELINK-NEXT: ret <2 x float> [[TMP9]] +; NOPRELINK-NEXT: ret <2 x float> [[TMP8]] ; entry: %call = tail call afn nnan ninf <2 x float> @_Z5rootnDv2_fDv2_i(<2 x float> %x, <2 x i32> <i32 3, i32 3>) @@ -1639,12 +1625,10 @@ define <2 x float> @test_rootn_afn_nnan_ninf_v2f32__y_neg3(<2 x float> %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn <2 x float> @llvm.log2.v2f32(<2 x float> [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn <2 x float> [[TMP2]], splat (float f0xBEAAAAAB) ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn <2 x float> @llvm.exp2.v2f32(<2 x float> [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP4]], <2 x float> [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn une <2 x float> [[X]], zeroinitializer -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> zeroinitializer, <2 x float> splat (float +inf) +; NOPRELINK-NEXT: [[TMP7:%.*]] = select <2 x i1> [[TMP6]], <2 x float> [[TMP4]], <2 x float> splat (float +inf) ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP7]], <2 x float> [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> [[TMP5]], <2 x float> [[TMP8]] -; NOPRELINK-NEXT: ret <2 x float> [[TMP9]] +; NOPRELINK-NEXT: ret <2 x float> [[TMP8]] ; entry: %call = tail call afn nnan ninf <2 x float> @_Z5rootnDv2_fDv2_i(<2 x float> %x, <2 x i32> <i32 -3, i32 -3>) @@ -1690,12 +1674,10 @@ define <2 x float> @test_rootn_afn_nnan_ninf_v2f32__y_5(<2 x float> %x) { ; NOPRELINK-NEXT: [[TMP2:%.*]] = call nnan ninf afn <2 x float> @llvm.log2.v2f32(<2 x float> [[TMP1]]) ; NOPRELINK-NEXT: [[TMP3:%.*]] = fmul nnan ninf afn <2 x float> [[TMP2]], splat (float 2.000000e-01) ; NOPRELINK-NEXT: [[TMP4:%.*]] = call nnan ninf afn <2 x float> @llvm.exp2.v2f32(<2 x float> [[TMP3]]) -; NOPRELINK-NEXT: [[TMP5:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP4]], <2 x float> [[X]]) ; NOPRELINK-NEXT: [[TMP6:%.*]] = fcmp nnan ninf afn oeq <2 x float> [[X]], zeroinitializer -; NOPRELINK-NEXT: [[TMP7:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> zeroinitializer, <2 x float> splat (float +inf) +; NOPRELINK-NEXT: [[TMP7:%.*]] = select <2 x i1> [[TMP6]], <2 x float> zeroinitializer, <2 x float> [[TMP4]] ; NOPRELINK-NEXT: [[TMP8:%.*]] = call nnan ninf afn <2 x float> @llvm.copysign.v2f32(<2 x float> [[TMP7]], <2 x float> [[X]]) -; NOPRELINK-NEXT: [[TMP9:%.*]] = select nnan ninf afn <2 x i1> [[TMP6]], <2 x float> [[TMP8]], <2 x float> [[TMP5]] -; NOPRELINK-NEXT: ret <2 x float> [[TMP9]] +; NOPRELINK-NEXT: ret <2 x float> [[TMP8]] ; entry: %call = tail call afn nnan ninf <2 x float> @_Z5rootnDv2_fDv2_i(<2 x float> %x, <2 x i32> <i32 5, i32 5>) diff --git a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll index 3ae46dbf63b034..e779f1f3d6be87 100644 --- a/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll +++ b/llvm/test/CodeGen/AMDGPU/simplify-libcalls.ll @@ -1266,12 +1266,10 @@ define amdgpu_kernel void @test_rootn_3(ptr addrspace(1) nocapture %a) { ; GCN-POSTLINK-NEXT: [[TMP1:%.*]] = tail call fast float @llvm.log2.f32(float [[TMP0]]) ; GCN-POSTLINK-NEXT: [[TMP2:%.*]] = fmul fast float [[TMP1]], f0x3EAAAAAB ; GCN-POSTLINK-NEXT: [[TMP3:%.*]] = tail call fast float @llvm.exp2.f32(float [[TMP2]]) -; GCN-POSTLINK-NEXT: [[TMP4:%.*]] = tail call fast float @llvm.copysign.f32(float [[TMP3]], float [[TMP]]) ; GCN-POSTLINK-NEXT: [[TMP5:%.*]] = fcmp fast oeq float [[TMP]], 0.000000e+00 -; GCN-POSTLINK-NEXT: [[TMP6:%.*]] = select fast i1 [[TMP5]], float 0.000000e+00, float +inf +; GCN-POSTLINK-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], float 0.000000e+00, float [[TMP3]] ; GCN-POSTLINK-NEXT: [[TMP7:%.*]] = tail call fast float @llvm.copysign.f32(float [[TMP6]], float [[TMP]]) -; GCN-POSTLINK-NEXT: [[TMP8:%.*]] = select fast i1 [[TMP5]], float [[TMP7]], float [[TMP4]] -; GCN-POSTLINK-NEXT: store float [[TMP8]], ptr addrspace(1) [[A]], align 4 +; GCN-POSTLINK-NEXT: store float [[TMP7]], ptr addrspace(1) [[A]], align 4 ; GCN-POSTLINK-NEXT: ret void ; ; GCN-PRELINK-LABEL: define amdgpu_kernel void @test_rootn_3( >From 87cd6a7333f48a1717fe74ed77f790c9d2054852 Mon Sep 17 00:00:00 2001 From: Harrison Hao <[email protected]> Date: Wed, 14 Oct 2026 17:49:13 +0800 Subject: [PATCH 5/5] Update for comments --- .../lib/Transforms/InstCombine/InstCombineSelect.cpp | 12 +----------- 1 file changed, 1 insertion(+), 11 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp index dba320922d4b3a..6428a970557e22 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp @@ -526,8 +526,7 @@ foldSelectIntoIntrinsicArg(SelectInst &SI, IntrinsicInst *TII, if (!DiffArgNo) return nullptr; - // A select cannot feed a parameter that requires an immediate. - if (TII->paramHasAttr(*DiffArgNo, Attribute::ImmArg)) + if (!canReplaceOperandWithVariable(TII, *DiffArgNo)) return nullptr; Value *TV = TII->getArgOperand(*DiffArgNo); @@ -580,15 +579,6 @@ Instruction *InstCombinerImpl::foldSelectIntrinsic(SelectInst &SI) { return replaceInstUsesWith(SI, NewCall); } - case Intrinsic::ctpop: { - Value *TV = LHSIntrinsic->getArgOperand(0); - Value *FV = RHSIntrinsic->getArgOperand(0); - - Value *NewSel = Builder.CreateSelect(SI.getCondition(), TV, FV, "", &SI); - Value *NewCall = Builder.CreateUnaryIntrinsic(IID, NewSel); - - return replaceInstUsesWith(SI, NewCall); - } default: if (isTriviallyVectorizable(IID)) return foldSelectIntoIntrinsicArg(SI, LHSIntrinsic, RHSIntrinsic, _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
