Fix MergeInVectorType to check for vector types with the same alloc

size but different element types, so that it filters out the cases
that CreateShuffleVectorCast doesn't handle. This fixes rdar://9786827.

llvm-svn: 135721

2 files changed, 33 insertions, 0 deletions

diff --git a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index fbf309292d8..b3d7ef6ec10 100644
--- a/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/llvm/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -425,6 +425,12 @@ bool ConvertToScalarInfo::MergeInVectorType(VectorType *VInTy,
   Type *ElementTy = VectorTy->getElementType();
   Type *InElementTy = VInTy->getElementType();
 
+  // If they're the same alloc size, we'll be attempting to convert between
+  // them with a vector shuffle, which requires the element types to match.
+  if (TD.getTypeAllocSize(VectorTy) == TD.getTypeAllocSize(VInTy) &&
+      ElementTy != InElementTy)
+    return false;
+
   // Do not allow mixed integer and floating-point accesses from vectors of
   // different sizes.
   if (ElementTy->isFloatingPointTy() != InElementTy->isFloatingPointTy())
diff --git a/llvm/test/Transforms/ScalarRepl/vectors-with-mismatched-elements.ll b/llvm/test/Transforms/ScalarRepl/vectors-with-mismatched-elements.ll
new file mode 100644
index 00000000000..c3fbdf5f863
--- /dev/null
+++ b/llvm/test/Transforms/ScalarRepl/vectors-with-mismatched-elements.ll
@@ -0,0 +1,27 @@
+; RUN: opt -scalarrepl -S < %s | FileCheck %s
+; rdar://9786827
+
+; SROA should be able to handle the mixed types and eliminate the allocas here.
+
+; TODO: Currently it does this by falling back to integer "bags of bits".
+; With enough cleverness, it should be possible to convert between <3 x i32>
+; and <2 x i64> by using a combination of a bitcast and a shuffle.
+
+; CHECK: {
+; CHECK-NOT: alloca
+; CHECK: }
+
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32"
+target triple = "i386-apple-darwin11.0.0"
+
+define <2 x i64> @foo() nounwind {
+entry:
+  %retval = alloca <3 x i32>, align 16
+  %z = alloca <4 x i32>, align 16
+  %tmp = load <4 x i32>* %z
+  %tmp1 = shufflevector <4 x i32> %tmp, <4 x i32> undef, <3 x i32> <i32 0, i32 1, i32 2>
+  store <3 x i32> %tmp1, <3 x i32>* %retval
+  %0 = bitcast <3 x i32>* %retval to <2 x i64>*
+  %1 = load <2 x i64>* %0, align 1
+  ret <2 x i64> %1
+}