diff options
author | Shuxin Yang <shuxin.llvm@gmail.com> | 2013-09-19 21:13:46 +0000 |
---|---|---|
committer | Shuxin Yang <shuxin.llvm@gmail.com> | 2013-09-19 21:13:46 +0000 |
commit | b1ccfb3a548e122e282cd62c534c4d47f5310bf6 (patch) | |
tree | 39cebad965612ce61b118f5ad31abbd08026a31b | |
parent | bf22298093d7069fbf3a3149d26fbcb4ad453ceb (diff) | |
download | llvm-b1ccfb3a548e122e282cd62c534c4d47f5310bf6.tar.gz llvm-b1ccfb3a548e122e282cd62c534c4d47f5310bf6.tar.bz2 llvm-b1ccfb3a548e122e282cd62c534c4d47f5310bf6.tar.xz |
[Fast-math] Disable "(C1/X)*C2 => (C1*C2)/X" if C1/X has multiple uses.
If "C1/X" were having multiple uses, the only benefit of this
transformation is to potentially shorten critical path. But it is at the
cost of instroducing additional div.
The additional div may or may not incur cost depending on how div is
implemented. If it is implemented using Newton–Raphson iteration, it dosen't
seem to incur any cost (FIXME). However, if the div blocks the entire
pipeline, that sounds to be pretty expensive. Let CodeGen to take care
this transformation.
This patch sees 6% on a benchmark.
rdar://15032743
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@191037 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r-- | lib/Transforms/InstCombine/InstCombineMulDivRem.cpp | 9 | ||||
-rw-r--r-- | test/Transforms/InstCombine/fast-math.ll | 12 |
2 files changed, 18 insertions, 3 deletions
diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index cc6a301540..9c310f04dc 100644 --- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -374,9 +374,12 @@ Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, ConstantFP *C, } else { if (C0) { // (C0 / X) * C => (C0 * C) / X - ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C)); - if (isNormalFp(F)) - R = BinaryOperator::CreateFDiv(F, Opnd1); + if (FMulOrDiv->hasOneUse()) { + // It would otherwise introduce another div. + ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C)); + if (isNormalFp(F)) + R = BinaryOperator::CreateFDiv(F, Opnd1); + } } else { // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFDiv(C, C1)); diff --git a/test/Transforms/InstCombine/fast-math.ll b/test/Transforms/InstCombine/fast-math.ll index a9a7015f62..d8ba2a59ff 100644 --- a/test/Transforms/InstCombine/fast-math.ll +++ b/test/Transforms/InstCombine/fast-math.ll @@ -202,6 +202,18 @@ define float @fmul2(float %f1) { ; CHECK: fdiv fast float 1.200000e+07, %f1 } +; X/C1 * C2 => X * (C2/C1) is disabled if X/C1 has multiple uses +@fmul2_external = external global float +define float @fmul2_disable(float %f1) { + %div = fdiv fast float 1.000000e+00, %f1 + store float %div, float* @fmul2_external + %mul = fmul fast float %div, 2.000000e+00 + ret float %mul +; CHECK-LABEL: @fmul2_disable +; CHECK: store +; CHECK: fmul fast +} + ; X/C1 * C2 => X * (C2/C1) (if C2/C1 is normal Fp) define float @fmul3(float %f1, float %f2) { %t1 = fdiv float %f1, 2.0e+3 |