From 112aac4e8961b9626bb84f36deeaa5a674f03f5a Mon Sep 17 00:00:00 2001 From: David Green Date: Wed, 18 Sep 2024 09:38:28 +0100 Subject: [PATCH] [InstCombine] Fold fmod to frem if we know it does not set errno. (#107912) fmod will be folded to frem in clang under -fno-math-errno and can be constant folded in llvm if the operands are known. It can be relatively common to have fp code that handles special values before doing some calculation: ``` if (isnan(f)) return handlenan; if (isinf(f)) return handleinf; .. fmod(f, 2.0) ``` This patch enables the folding of fmod to frem in instcombine if the first parameter is not inf and the second is not zero. Other combinations do not set errno. The same transform is performed for fmod with the nnan flag, which implies the input is known to not be inf/zero. --- .../llvm/Transforms/Utils/SimplifyLibCalls.h | 1 + .../lib/Transforms/Utils/SimplifyLibCalls.cpp | 33 +++++++++++++++++++ llvm/test/Transforms/InstCombine/fmod.ll | 30 ++++++++++++----- 3 files changed, 55 insertions(+), 9 deletions(-) diff --git a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h index 2e7a0ec29ed999..2d3d2ada6183a7 100644 --- a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h +++ b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h @@ -206,6 +206,7 @@ class LibCallSimplifier { Value *optimizeFMinFMax(CallInst *CI, IRBuilderBase &B); Value *optimizeLog(CallInst *CI, IRBuilderBase &B); Value *optimizeSqrt(CallInst *CI, IRBuilderBase &B); + Value *optimizeFMod(CallInst *CI, IRBuilderBase &B); Value *mergeSqrtToExp(CallInst *CI, IRBuilderBase &B); Value *optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B); Value *optimizeTrigInversionPairs(CallInst *CI, IRBuilderBase &B); diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp index 917f81863cf673..4933b5bf60eea8 100644 --- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp +++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp @@ -2796,6 +2796,35 @@ Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) { return copyFlags(*CI, FabsCall); } +Value *LibCallSimplifier::optimizeFMod(CallInst *CI, IRBuilderBase &B) { + SimplifyQuery SQ(DL, TLI, DT, AC, CI, true, true, DC); + + // fmod(x,y) can set errno if y == 0 or x == +/-inf, and returns Nan in those + // case. If we know those do not happen, then we can convert the fmod into + // frem. + bool IsNoNan = CI->hasNoNaNs(); + if (!IsNoNan) { + KnownFPClass Known0 = computeKnownFPClass(CI->getOperand(0), fcInf, + /*Depth=*/0, SQ); + if (Known0.isKnownNeverInfinity()) { + KnownFPClass Known1 = + computeKnownFPClass(CI->getOperand(1), fcZero | fcSubnormal, + /*Depth=*/0, SQ); + Function *F = CI->getParent()->getParent(); + if (Known1.isKnownNeverLogicalZero(*F, CI->getType())) + IsNoNan = true; + } + } + + if (IsNoNan) { + Value *FRem = B.CreateFRemFMF(CI->getOperand(0), CI->getOperand(1), CI); + if (auto *FRemI = dyn_cast(FRem)) + FRemI->setHasNoNaNs(true); + substituteInParent(CI, FRem); + } + return nullptr; +} + Value *LibCallSimplifier::optimizeTrigInversionPairs(CallInst *CI, IRBuilderBase &B) { Module *M = CI->getModule(); @@ -3945,6 +3974,10 @@ Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI, case LibFunc_sqrt: case LibFunc_sqrtl: return optimizeSqrt(CI, Builder); + case LibFunc_fmod: + case LibFunc_fmodf: + case LibFunc_fmodl: + return optimizeFMod(CI, Builder); case LibFunc_logf: case LibFunc_log: case LibFunc_logl: diff --git a/llvm/test/Transforms/InstCombine/fmod.ll b/llvm/test/Transforms/InstCombine/fmod.ll index c021d27e95fa5e..10cff189b8dfca 100644 --- a/llvm/test/Transforms/InstCombine/fmod.ll +++ b/llvm/test/Transforms/InstCombine/fmod.ll @@ -9,7 +9,7 @@ define float @test_inf_const(float %f) { ; CHECK-NEXT: [[ISINF:%.*]] = fcmp oeq float [[ABS]], 0x7FF0000000000000 ; CHECK-NEXT: br i1 [[ISINF]], label [[RETURN:%.*]], label [[IF_END:%.*]] ; CHECK: if.end: -; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float [[F]], float 2.000000e+00) +; CHECK-NEXT: [[CALL:%.*]] = frem nnan float [[F]], 2.000000e+00 ; CHECK-NEXT: ret float [[CALL]] ; CHECK: return: ; CHECK-NEXT: ret float 0.000000e+00 @@ -34,7 +34,7 @@ define float @test_const_zero(float %f) { ; CHECK-NEXT: [[ISZERO:%.*]] = fcmp oeq float [[F]], 0.000000e+00 ; CHECK-NEXT: br i1 [[ISZERO]], label [[RETURN:%.*]], label [[IF_END:%.*]] ; CHECK: if.end: -; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float 2.000000e+00, float [[F]]) +; CHECK-NEXT: [[CALL:%.*]] = frem nnan float 2.000000e+00, [[F]] ; CHECK-NEXT: ret float [[CALL]] ; CHECK: return: ; CHECK-NEXT: ret float 0.000000e+00 @@ -67,11 +67,11 @@ define float @test_noinf_nozero(float nofpclass(inf) %f, float nofpclass(zero) % ; CHECK-LABEL: define float @test_noinf_nozero( ; CHECK-SAME: float nofpclass(inf) [[F:%.*]], float nofpclass(zero) [[G:%.*]]) { ; CHECK-NEXT: entry: -; CHECK-NEXT: [[CALL:%.*]] = tail call nnan float @fmodf(float [[F]], float [[G]]) +; CHECK-NEXT: [[CALL:%.*]] = frem nnan float [[F]], [[G]] ; CHECK-NEXT: ret float [[CALL]] ; entry: - %call = tail call nnan float @fmodf(float %f, float %g) + %call = tail call float @fmodf(float %f, float %g) ret float %call } @@ -79,7 +79,7 @@ define double @test_double(double nofpclass(inf) %f, double nofpclass(zero) %g) ; CHECK-LABEL: define double @test_double( ; CHECK-SAME: double nofpclass(inf) [[F:%.*]], double nofpclass(zero) [[G:%.*]]) { ; CHECK-NEXT: entry: -; CHECK-NEXT: [[CALL:%.*]] = tail call double @fmod(double [[F]], double [[G]]) +; CHECK-NEXT: [[CALL:%.*]] = frem nnan double [[F]], [[G]] ; CHECK-NEXT: ret double [[CALL]] ; entry: @@ -91,7 +91,7 @@ define fp128 @test_fp128(fp128 nofpclass(inf) %f, fp128 nofpclass(zero) %g) { ; CHECK-LABEL: define fp128 @test_fp128( ; CHECK-SAME: fp128 nofpclass(inf) [[F:%.*]], fp128 nofpclass(zero) [[G:%.*]]) { ; CHECK-NEXT: entry: -; CHECK-NEXT: [[CALL:%.*]] = tail call fp128 @fmodl(fp128 [[F]], fp128 [[G]]) +; CHECK-NEXT: [[CALL:%.*]] = frem nnan fp128 [[F]], [[G]] ; CHECK-NEXT: ret fp128 [[CALL]] ; entry: @@ -103,11 +103,11 @@ define float @test_noinf_nozero_dazpreservesign(float nofpclass(inf) %f, float n ; CHECK-LABEL: define float @test_noinf_nozero_dazpreservesign( ; CHECK-SAME: float nofpclass(inf) [[F:%.*]], float nofpclass(zero) [[G:%.*]]) #[[ATTR0:[0-9]+]] { ; CHECK-NEXT: entry: -; CHECK-NEXT: [[CALL:%.*]] = tail call nnan float @fmodf(float [[F]], float [[G]]) +; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float [[F]], float [[G]]) ; CHECK-NEXT: ret float [[CALL]] ; entry: - %call = tail call nnan float @fmodf(float %f, float %g) + %call = tail call float @fmodf(float %f, float %g) ret float %call } @@ -115,7 +115,19 @@ define float @test_noinf_nozero_dazdynamic(float nofpclass(inf) %f, float nofpcl ; CHECK-LABEL: define float @test_noinf_nozero_dazdynamic( ; CHECK-SAME: float nofpclass(inf) [[F:%.*]], float nofpclass(zero) [[G:%.*]]) #[[ATTR1:[0-9]+]] { ; CHECK-NEXT: entry: -; CHECK-NEXT: [[CALL:%.*]] = tail call nnan float @fmodf(float [[F]], float [[G]]) +; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float [[F]], float [[G]]) +; CHECK-NEXT: ret float [[CALL]] +; +entry: + %call = tail call float @fmodf(float %f, float %g) + ret float %call +} + +define float @test_nnan(float %f, float %g) { +; CHECK-LABEL: define float @test_nnan( +; CHECK-SAME: float [[F:%.*]], float [[G:%.*]]) { +; CHECK-NEXT: entry: +; CHECK-NEXT: [[CALL:%.*]] = frem nnan float [[F]], [[G]] ; CHECK-NEXT: ret float [[CALL]] ; entry: