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.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@133338 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 24 insertions, 2 deletions

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index 38942dcb55..46ac948200 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -293,6 +293,11 @@ AllocaInst *ConvertToScalarInfo::TryConvert(AllocaInst *AI) {
   if (ScalarKind == Unknown)
     ScalarKind = Integer;
 
+  // FIXME: It should be possible to promote the vector type up to the alloca's
+  // size.
+  if (ScalarKind == Vector && VectorTy->getBitWidth() != AllocaSize * 8)
+    ScalarKind = Integer;
+
   // If we were able to find a vector type that can handle this with
   // insert/extract elements, and if there was at least one use that had
   // a vector type, promote this to a vector.  We don't want to promote
diff --git a/test/Transforms/ScalarRepl/2011-06-08-VectorExtractValue.ll b/test/Transforms/ScalarRepl/2011-06-08-VectorExtractValue.ll
index 32e67fbce7..98fa1c684c 100644
--- a/test/Transforms/ScalarRepl/2011-06-08-VectorExtractValue.ll
+++ b/test/Transforms/ScalarRepl/2011-06-08-VectorExtractValue.ll
@@ -10,7 +10,8 @@ target triple = "x86_64-apple-macosx10.7.0"
 
 ; CHECK: main
 ; CHECK-NOT: alloca
-; CHECK: extractelement <2 x float> zeroinitializer
+; CHECK: %[[A:[a-z0-9]*]] = and i128
+; CHECK: %[[B:[a-z0-9]*]] = trunc i128 %[[A]] to i32
 
 define void @main() uwtable ssp {
 entry:
@@ -27,7 +28,8 @@ entry:
 
 ; CHECK: test1
 ; CHECK-NOT: alloca
-; CHECK: extractelement <2 x float> zeroinitializer
+; CHECK: %[[A:[a-z0-9]*]] = and i128
+; CHECK: %[[B:[a-z0-9]*]] = trunc i128 %[[A]] to i32
 
 define void @test1() uwtable ssp {
 entry:
diff --git a/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll b/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll
index 29d195eba3..f8530d68a8 100644
--- a/test/Transforms/ScalarRepl/2011-06-17-VectorPartialMemset.ll
+++ b/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