diff options
author | Alex MacLean <amaclean@nvidia.com> | 2024-06-22 09:50:21 -0700 |
---|---|---|
committer | GitHub <noreply@github.com> | 2024-06-22 09:50:21 -0700 |
commit | a4ca22506c96cd8508bcfb93a26941f19c7179ef (patch) | |
tree | e30472afbd937cd68302f62f6f46e46df3f5b64a /llvm | |
parent | 485d7eaefd93c4f6bc8c51c9a169ffb22ce3a898 (diff) | |
download | llvm-a4ca22506c96cd8508bcfb93a26941f19c7179ef.zip llvm-a4ca22506c96cd8508bcfb93a26941f19c7179ef.tar.gz llvm-a4ca22506c96cd8508bcfb93a26941f19c7179ef.tar.bz2 |
[InstCombine] (uitofp bool X) * Y --> X ? Y : 0 (#96216)
Fold `mul (uitofp i1 X), Y` to `select i1 X, Y, 0.0` when the `mul` is
`nnan` and `nsz`
Proof: https://alive2.llvm.org/ce/z/_stiPm
Diffstat (limited to 'llvm')
-rw-r--r-- | llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp | 18 | ||||
-rw-r--r-- | llvm/test/Transforms/InstCombine/fmul-bool.ll | 35 |
2 files changed, 53 insertions, 0 deletions
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index 8fcb354..c3f1c12 100644 --- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -879,6 +879,24 @@ Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) { if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) return BinaryOperator::CreateFMulFMF(X, NegC, &I); + if (I.hasNoNaNs() && I.hasNoSignedZeros()) { + // (uitofp bool X) * Y --> X ? Y : 0 + // Y * (uitofp bool X) --> X ? Y : 0 + // Note INF * 0 is NaN. + if (match(Op0, m_UIToFP(m_Value(X))) && + X->getType()->isIntOrIntVectorTy(1)) { + auto *SI = SelectInst::Create(X, Op1, ConstantFP::get(I.getType(), 0.0)); + SI->copyFastMathFlags(I.getFastMathFlags()); + return SI; + } + if (match(Op1, m_UIToFP(m_Value(X))) && + X->getType()->isIntOrIntVectorTy(1)) { + auto *SI = SelectInst::Create(X, Op0, ConstantFP::get(I.getType(), 0.0)); + SI->copyFastMathFlags(I.getFastMathFlags()); + return SI; + } + } + // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1)) return replaceInstUsesWith(I, V); diff --git a/llvm/test/Transforms/InstCombine/fmul-bool.ll b/llvm/test/Transforms/InstCombine/fmul-bool.ll new file mode 100644 index 0000000..263ca53 --- /dev/null +++ b/llvm/test/Transforms/InstCombine/fmul-bool.ll @@ -0,0 +1,35 @@ +; RUN: opt < %s -passes=instcombine -S | FileCheck %s + +; X * Y (when Y is a boolean) --> Y ? X : 0 + +define float @fmul_bool(float %x, i1 %y) { +; CHECK-LABEL: @fmul_bool( +; CHECK-NEXT: [[M:%.*]] = select nnan nsz i1 [[Y:%.*]], float [[X:%.*]], float 0.000000e+00 +; CHECK-NEXT: ret float [[M]] +; + %z = uitofp i1 %y to float + %m = fmul nnan nsz float %z, %x + ret float %m +} + +define <2 x float> @fmul_bool_vec(<2 x float> %x, <2 x i1> %y) { +; CHECK-LABEL: @fmul_bool_vec( +; CHECK-NEXT: [[M:%.*]] = select nnan nsz <2 x i1> [[Y:%.*]], <2 x float> [[X:%.*]], <2 x float> zeroinitializer +; CHECK-NEXT: ret <2 x float> [[M]] +; + %z = uitofp <2 x i1> %y to <2 x float> + %m = fmul nnan nsz <2 x float> %z, %x + ret <2 x float> %m +} + +define <2 x float> @fmul_bool_vec_commute(<2 x float> %px, <2 x i1> %y) { +; CHECK-LABEL: @fmul_bool_vec_commute( +; CHECK-NEXT: [[X:%.*]] = fmul nnan nsz <2 x float> [[PX:%.*]], [[PX]] +; CHECK-NEXT: [[M:%.*]] = select nnan nsz <2 x i1> [[Y:%.*]], <2 x float> [[X]], <2 x float> zeroinitializer +; CHECK-NEXT: ret <2 x float> [[M]] +; + %x = fmul nnan nsz <2 x float> %px, %px ; thwart complexity-based canonicalization + %z = uitofp <2 x i1> %y to <2 x float> + %m = fmul nnan nsz <2 x float> %x, %z + ret <2 x float> %m +} |