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

1 files changed, 19 insertions, 3 deletions

diff --git a/test/Transforms/InstCombine/sqrt.ll b/test/Transforms/InstCombine/sqrt.ll
index 6505382048..69e511bfb3 100644
--- a/test/Transforms/InstCombine/sqrt.ll
+++ b/test/Transforms/InstCombine/sqrt.ll
@@ -1,16 +1,32 @@
 ; RUN: opt -S -instcombine %s | FileCheck %s
 
-define float @foo(float %x) nounwind readnone ssp {
+define float @test1(float %x) nounwind readnone ssp {
 entry:
+; CHECK: @test1
 ; CHECK-NOT: fpext
 ; CHECK-NOT: sqrt(
 ; CHECK: sqrtf(
 ; CHECK-NOT: fptrunc
   %conv = fpext float %x to double                ; <double> [#uses=1]
-  %call = tail call double @sqrt(double %conv) nounwind ; <double> [#uses=1]
+  %call = tail call double @sqrt(double %conv) readnone nounwind ; <double> [#uses=1]
   %conv1 = fptrunc double %call to float          ; <float> [#uses=1]
 ; CHECK: ret float
   ret float %conv1
 }
 
-declare double @sqrt(double) readnone
+declare double @sqrt(double)
+
+; PR8096
+define float @test2(float %x) nounwind readnone ssp {
+entry:
+; CHECK: @test2
+; CHECK-NOT: fpext
+; CHECK-NOT: sqrt(
+; CHECK: sqrtf(
+; CHECK-NOT: fptrunc
+  %conv = fpext float %x to double                ; <double> [#uses=1]
+  %call = tail call double @sqrt(double %conv) nounwind ; <double> [#uses=1]
+  %conv1 = fptrunc double %call to float          ; <float> [#uses=1]
+; CHECK: ret float
+  ret float %conv1
+}