Use the splat index to generate the desired shuffle. Otherwise we

could only get undefs and the vector shuffle becomes an undef,
generating wrong code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@137295 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 217a7bcf5308f59b8c328fe18bf46c4bd546c05e..ae573284d7dc6a0af11c7e329d9d8ba819346c1e 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -4125,15 +4125,15 @@ static SDValue PromoteVectorToScalarSplat(ShuffleVectorSDNode *SV,
   int NumElems = SrcVT.getVectorNumElements();
 
   assert(SrcVT.is256BitVector() && "unknown howto handle vector type");
+  assert(SV->isSplat() && "shuffle must be a splat");
 
-  SmallVector<int, 4> Mask;
-  for (int i = 0; i < NumElems/2; ++i)
-    Mask.push_back(SV->getMaskElt(i));
+  int SplatIdx = SV->getSplatIndex();
+  const int Mask[4] = { SplatIdx, SplatIdx, SplatIdx, SplatIdx };
 
   EVT SVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
                              NumElems/2);
   SDValue SV1 = DAG.getVectorShuffle(SVT, dl, V1.getOperand(1),
-                                     DAG.getUNDEF(SVT), &Mask[0]);
+                                     DAG.getUNDEF(SVT), Mask);
   SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), SV1,
                                     DAG.getConstant(0, MVT::i32), DAG, dl);
 

diff --git a/test/CodeGen/X86/avx-splat.ll b/test/CodeGen/X86/avx-splat.ll
index 36d469417f93edde2987a1072b17d5337e3d0f6a..243ab9ba318f8a389787add8f46b7e2d512c07ab 100644 (file)
--- a/test/CodeGen/X86/avx-splat.ll
+++ b/test/CodeGen/X86/avx-splat.ll
@@ -77,3 +77,13 @@ __load_and_broadcast_32.exit1249:                 ; preds = %load.i1247, %for_ex
   ret void
 }
 
+; CHECK: vpshufd  $0
+; CHECK-NEXT: vinsertf128 $1
+define <8 x float> @funcF(i32* %ptr) nounwind {
+  %val = load i32* %ptr, align 4
+  %ret6 = insertelement <8 x i32> undef, i32 %val, i32 6
+  %ret7 = insertelement <8 x i32> %ret6, i32 %val, i32 7
+  %tmp = bitcast <8 x i32> %ret7 to <8 x float>
+  ret <8 x float> %tmp
+}
+