RPCSX/llvm
1
0
Fork 0
mirror of https://github.com/RPCSX/llvm.git synced 2024-12-04 18:06:49 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Apply the InstCombine fptrunc sqrt optimization to llvm.sqrt

Browse Source 
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
This commit is contained in:
Hal Finkel 2013-11-16 21:29:08 +00:00
parent e6e811277f
commit 64fa501b10
2 changed files with 24 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
lib/Transforms/InstCombine/InstCombineCasts.cpp
View File

@ -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));
@ -1275,10 +1280,10 @@ 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(),
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");

13
test/Transforms/InstCombine/double-float-shrink-1.ll
View File

@ -263,6 +263,7 @@ define double @sin_test2(float %f) nounwind readnone {
ret double %call
; CHECK: call double @sin(double %conv)
}
define float @sqrt_test(float %f) nounwind readnone {
; CHECK: sqrt_test
%conv = fpext float %f to double
@ -272,6 +273,15 @@ define float @sqrt_test(float %f) nounwind readnone {
; CHECK: call float @sqrtf(float %f)
}
define float @sqrt_int_test(float %f) nounwind readnone {
; CHECK: sqrt_int_test
%conv = fpext float %f to double
%call = call double @llvm.sqrt.f64(double %conv)
%conv1 = fptrunc double %call to float
ret float %conv1
; CHECK: call float @llvm.sqrt.f32(float %f)
}
define double @sqrt_test2(float %f) nounwind readnone {
; CHECK: sqrt_test2
%conv = fpext float %f to double
@ -331,3 +341,6 @@ declare double @acos(double) nounwind readnone
declare double @acosh(double) nounwind readnone
declare double @asin(double) nounwind readnone
declare double @asinh(double) nounwind readnone
declare double @llvm.sqrt.f64(double) nounwind readnone