Fixed a bug in LowerVECTOR_SHUFFLE and LowerBUILD_VECTOR.

Matching MOVLP mask for AVX (265-bit vectors) was wrong.
The failure was detected by conformance tests.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147308 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 40 insertions, 5 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 16a3d090e9..00b46d2cf6 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -3448,6 +3448,11 @@ bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
+  EVT VT = N->getValueType(0);
+
+  if (VT.getSizeInBits() != 128)
+    return false;
+
   unsigned NumElems = N->getValueType(0).getVectorNumElements();
 
   if (NumElems != 2 && NumElems != 4)
@@ -3666,6 +3671,8 @@ bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2) {
 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
   if (VT.getVectorElementType().getSizeInBits() < 32)
     return false;
+  if (VT.getSizeInBits() == 256)
+    return false;
 
   int NumElts = VT.getVectorNumElements();
 
@@ -5158,16 +5165,30 @@ X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
         return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
       } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
           (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
+        if (VT.getSizeInBits() == 256) {
+          
+          EVT VT128 = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems / 2);
+          Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Item);
+          SDValue ZeroVec = getZeroVector(VT, true, DAG, dl);              
+          return Insert128BitVector(ZeroVec, Item, DAG.getConstant(0, MVT::i32),
+                              DAG, dl);
+        }
         Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
         // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
         return getShuffleVectorZeroOrUndef(Item, 0, true,Subtarget->hasXMMInt(),
                                            DAG);
       } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
         Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
-        unsigned NumBits = VT.getSizeInBits();
-        assert((NumBits == 128 || NumBits == 256) && 
-               "Expected an SSE or AVX value type!");
-        EVT MiddleVT = NumBits == 128 ? MVT::v4i32 : MVT::v8i32;
+        if (VT.getSizeInBits() == 256) {
+          
+          EVT VT128 = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems / 2);
+          Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Item);
+          SDValue ZeroVec = getZeroVector(VT, true, DAG, dl);              
+          return Insert128BitVector(ZeroVec, Item, DAG.getConstant(0, MVT::i32),
+                              DAG, dl);
+        }
+        assert (VT.getSizeInBits() == 128 || "Expected an SSE value type!");
+        EVT MiddleVT = MVT::v4i32;
         Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
         Item = getShuffleVectorZeroOrUndef(Item, 0, true,
                                            Subtarget->hasXMMInt(), DAG);
diff --git a/test/CodeGen/X86/avx-shuffle.ll b/test/CodeGen/X86/avx-shuffle.ll
index e9392ae607..8532b40613 100644
--- a/test/CodeGen/X86/avx-shuffle.ll
+++ b/test/CodeGen/X86/avx-shuffle.ll
@@ -13,8 +13,22 @@ define <4 x float> @test1(<4 x float> %a) nounwind {
 define <3 x i64> @test2(<2 x i64> %v) nounwind readnone {
 ; CHECK: test2:
 ; CHECK: vxorpd
-; CHECK: vmovsd
+; CHECK: vperm2f128
   %1 = shufflevector <2 x i64> %v, <2 x i64> %v, <3 x i32> <i32 0, i32 1, i32 undef>
   %2 = shufflevector <3 x i64> zeroinitializer, <3 x i64> %1, <3 x i32> <i32 3, i32 4, i32 2>
   ret <3 x i64> %2
 }
+
+define <4 x i64> @test3(<4 x i64> %a, <4 x i64> %b) nounwind {
+  %c = shufflevector <4 x i64> %a, <4 x i64> %b, <4 x i32> <i32 4, i32 5, i32 2, i32 undef>
+  ret <4 x i64> %c
+; CHECK: test3:
+; CHECK: vperm2f128
+}
+
+define <8 x float> @test4(float %a) nounwind {
+  %b = insertelement <8 x float> zeroinitializer, float %a, i32 0
+  ret <8 x float> %b
+; CHECK: test4:
+; CHECK: vinsertf128
+}
\ No newline at end of file