From 1dd6f2a5ba0b92bad890505317dde76b056855e0 Mon Sep 17 00:00:00 2001 From: Michael Ilseman Date: Thu, 7 Feb 2013 01:40:15 +0000 Subject: [PATCH] Preserve fast-math flags after reassociation and commutation. Update test cases llvm-svn: 174571 --- .../InstCombine/InstructionCombining.cpp | 25 +++++++++++++++++----- llvm/test/Transforms/InstCombine/fast-math.ll | 8 +++---- 2 files changed, 24 insertions(+), 9 deletions(-) diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp index dc7fe5c..c6115e3 100644 --- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp +++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp @@ -162,6 +162,21 @@ static bool MaintainNoSignedWrap(BinaryOperator &I, Value *B, Value *C) { return !Overflow; } +/// Conservatively clears subclassOptionalData after a reassociation or +/// commutation. We preserve fast-math flags when applicable as they can be +/// preserved. +static void ClearSubclassDataAfterReassociation(BinaryOperator &I) { + FPMathOperator *FPMO = dyn_cast(&I); + if (!FPMO) { + I.clearSubclassOptionalData(); + return; + } + + FastMathFlags FMF = I.getFastMathFlags(); + I.clearSubclassOptionalData(); + I.setFastMathFlags(FMF); +} + /// SimplifyAssociativeOrCommutative - This performs a few simplifications for /// operators which are associative or commutative: // @@ -219,7 +234,7 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { I.clearSubclassOptionalData(); I.setHasNoSignedWrap(true); } else { - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); } Changed = true; @@ -241,7 +256,7 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { I.setOperand(1, C); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -263,7 +278,7 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { I.setOperand(1, B); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -283,7 +298,7 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { I.setOperand(1, V); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; ++NumReassoc; continue; @@ -310,7 +325,7 @@ bool InstCombiner::SimplifyAssociativeOrCommutative(BinaryOperator &I) { I.setOperand(1, Folded); // Conservatively clear the optional flags, since they may not be // preserved by the reassociation. - I.clearSubclassOptionalData(); + ClearSubclassDataAfterReassociation(I); Changed = true; continue; diff --git a/llvm/test/Transforms/InstCombine/fast-math.ll b/llvm/test/Transforms/InstCombine/fast-math.ll index 4de0c82..c97bd28 100644 --- a/llvm/test/Transforms/InstCombine/fast-math.ll +++ b/llvm/test/Transforms/InstCombine/fast-math.ll @@ -7,7 +7,7 @@ define float @fold(float %a) { %mul1 = fmul fast float %mul, 0x4002666660000000 ret float %mul1 ; CHECK: @fold -; CHECK: fmul float %a, 0x4006147AE0000000 +; CHECK: fmul fast float %a, 0x4006147AE0000000 } ; Same testing-case as the one used in fold() except that the operators have @@ -22,7 +22,7 @@ define float @notfold(float %a) { define float @fold2(float %a) { ; CHECK: @fold2 -; CHECK: fmul float %a, 0x4006147AE0000000 +; CHECK: fmul fast float %a, 0x4006147AE0000000 %mul = fmul float %a, 0x3FF3333340000000 %mul1 = fmul fast float %mul, 0x4002666660000000 ret float %mul1 @@ -54,7 +54,7 @@ define float @fold5(float %f1, float %f2) { %add1 = fadd fast float %add, 5.000000e+00 ret float %add1 ; CHECK: @fold5 -; CHECK: fadd float %f1, 9.000000e+00 +; CHECK: fadd fast float %f1, 9.000000e+00 } ; (X + X) + X => 3.0 * X @@ -107,7 +107,7 @@ define float @fold10(float %f1, float %f2) { %t3 = fadd fast float %t1, %t2 ret float %t3 ; CHECK: @fold10 -; CHECK: %t3 = fadd float %t2, -1.000000e+00 +; CHECK: %t3 = fadd fast float %t2, -1.000000e+00 ; CHECK: ret float %t3 } -- 2.7.4