Fix a serious performance regression introduced by r108687 on linux:
turning (fptrunc (sqrt (fpext x))) -> (sqrtf x) is great, but we have
to delete the original sqrt as well. Not doing so causes us to do
two sqrt's when building with -fmath-errno (the default on linux).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@113260 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/InstCombine/InstCombineCasts.cpp b/lib/Transforms/InstCombine/InstCombineCasts.cpp
index 6c2477d..79a9b09 100644
--- a/lib/Transforms/InstCombine/InstCombineCasts.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCasts.cpp
@@ -1139,7 +1139,7 @@
Arg->getOperand(0)->getType()->isFloatTy()) {
Function *Callee = Call->getCalledFunction();
Module *M = CI.getParent()->getParent()->getParent();
- Constant* SqrtfFunc = M->getOrInsertFunction("sqrtf",
+ Constant *SqrtfFunc = M->getOrInsertFunction("sqrtf",
Callee->getAttributes(),
Builder->getFloatTy(),
Builder->getFloatTy(),
@@ -1147,6 +1147,11 @@
CallInst *ret = CallInst::Create(SqrtfFunc, Arg->getOperand(0),
"sqrtfcall");
ret->setAttributes(Callee->getAttributes());
+
+
+ // Remove the old Call. With -fmath-errno, it won't get marked readnone.
+ Call->replaceAllUsesWith(UndefValue::get(Call->getType()));
+ EraseInstFromFunction(*Call);
return ret;
}
}