diff options
author | Hal Finkel <hfinkel@anl.gov> | 2013-11-16 21:29:08 +0000 |
---|---|---|
committer | Hal Finkel <hfinkel@anl.gov> | 2013-11-16 21:29:08 +0000 |
commit | 64fa501b1081b5d5c25e5e9639075abb0cb724d9 (patch) | |
tree | a190f3ea48a03753cd6f44e382cf97464e79b63b /lib | |
parent | e6e811277f045ee3d61cd62622d71005c47eb48d (diff) | |
download | external_llvm-64fa501b1081b5d5c25e5e9639075abb0cb724d9.zip external_llvm-64fa501b1081b5d5c25e5e9639075abb0cb724d9.tar.gz external_llvm-64fa501b1081b5d5c25e5e9639075abb0cb724d9.tar.bz2 |
Apply the InstCombine fptrunc sqrt optimization to llvm.sqrt
InstCombine, in visitFPTrunc, applies the following optimization to sqrt calls:
(fptrunc (sqrt (fpext x))) -> (sqrtf x)
but does not apply the same optimization to llvm.sqrt. This is a problem
because, to enable vectorization, Clang generates llvm.sqrt instead of sqrt in
fast-math mode, and because this optimization is being applied to sqrt and not
applied to llvm.sqrt, sometimes the fast-math code is slower.
This change makes InstCombine apply this optimization to llvm.sqrt as well.
This fixes the specific problem in PR17758, although the same underlying issue
(optimizations applied to libcalls are not applied to intrinsics) exists for
other optimizations in SimplifyLibCalls.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@194935 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r-- | lib/Transforms/InstCombine/InstCombineCasts.cpp | 17 |
1 files changed, 11 insertions, 6 deletions
diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp index a1aedd4..72377dc 100644 --- a/lib/Transforms/InstCombine/InstCombineCasts.cpp +++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp @@ -1262,9 +1262,14 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { } // Fold (fptrunc (sqrt (fpext x))) -> (sqrtf x) + // Note that we restrict this transformation based on + // TLI->has(LibFunc::sqrtf), even for the sqrt intrinsic, because + // TLI->has(LibFunc::sqrtf) is sufficient to guarantee that the + // single-precision intrinsic can be expanded in the backend. CallInst *Call = dyn_cast<CallInst>(CI.getOperand(0)); if (Call && Call->getCalledFunction() && TLI->has(LibFunc::sqrtf) && - Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) && + (Call->getCalledFunction()->getName() == TLI->getName(LibFunc::sqrt) || + Call->getCalledFunction()->getIntrinsicID() == Intrinsic::sqrt) && Call->getNumArgOperands() == 1 && Call->hasOneUse()) { CastInst *Arg = dyn_cast<CastInst>(Call->getArgOperand(0)); @@ -1275,11 +1280,11 @@ Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) { Arg->getOperand(0)->getType()->isFloatTy()) { Function *Callee = Call->getCalledFunction(); Module *M = CI.getParent()->getParent()->getParent(); - Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf", - Callee->getAttributes(), - Builder->getFloatTy(), - Builder->getFloatTy(), - NULL); + Constant *SqrtfFunc = (Callee->getIntrinsicID() == Intrinsic::sqrt) ? + Intrinsic::getDeclaration(M, Intrinsic::sqrt, Builder->getFloatTy()) : + M->getOrInsertFunction("sqrtf", Callee->getAttributes(), + Builder->getFloatTy(), Builder->getFloatTy(), + NULL); CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0), "sqrtfcall"); ret->setAttributes(Callee->getAttributes()); |