When scalar replacement returns a vector type, only accept it if the vector

type's bitwidth matches the (allocated) size of the alloca. This severely
pessimizes vector scalar replacement when the only vector type being used is
something like <3 x float> on x86 or ARM whose allocated size matches a
<4 x float>.

I hope to fix some of the flawed assumptions about allocated size throughout
scalar replacement and reenable this in most cases.

llvm-svn: 133338

1 files changed, 15 insertions, 0 deletions

diff --git a/llvm/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll b/llvm/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll
index 29d195eba36..f8530d68a8f 100644
--- a/llvm/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll
+++ b/llvm/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll
@@ -19,4 +19,19 @@ entry:
   ret float %val
 }
 
+; CHECK: g
+; CHECK-NOT: alloca
+; CHECK: and i128
+
+define void @g() nounwind ssp {
+entry:
+  %a = alloca { <4 x float> }, align 16
+  %p = bitcast { <4 x float> }* %a to i8*
+  call void @llvm.memset.p0i8.i32(i8* %p, i8 0, i32 16, i32 16, i1 false)
+  %q = bitcast { <4 x float> }* %a to [2 x <2 x float>]*
+  %arrayidx = getelementptr inbounds [2 x <2 x float>]* %q, i32 0, i32 0
+  store <2 x float> undef, <2 x float>* %arrayidx, align 8
+  ret void
+}
+
 declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) nounwind