[X86][SSE] matchBinaryVectorShuffle - add support for different src/dst value shuffle...

Preparation for support for combining to PACKSS/PACKUS

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@314656 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index fb48b66..83fb1ca 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -27373,7 +27373,7 @@ static bool matchBinaryVectorShuffle(MVT MaskVT, ArrayRef<int> Mask,
                                      SDValue &V1, SDValue &V2, SDLoc &DL,
                                      SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget,
-                                     unsigned &Shuffle, MVT &ShuffleVT,
+                                     unsigned &Shuffle, MVT &SrcVT, MVT &DstVT,
                                      bool IsUnary) {
   unsigned EltSizeInBits = MaskVT.getScalarSizeInBits();
 
@@ -27381,26 +27381,26 @@ static bool matchBinaryVectorShuffle(MVT MaskVT, ArrayRef<int> Mask,
     if (isTargetShuffleEquivalent(Mask, {0, 0}) && AllowFloatDomain) {
       V2 = V1;
       Shuffle = X86ISD::MOVLHPS;
-      ShuffleVT = MVT::v4f32;
+      SrcVT = DstVT = MVT::v4f32;
       return true;
     }
     if (isTargetShuffleEquivalent(Mask, {1, 1}) && AllowFloatDomain) {
       V2 = V1;
       Shuffle = X86ISD::MOVHLPS;
-      ShuffleVT = MVT::v4f32;
+      SrcVT = DstVT = MVT::v4f32;
       return true;
     }
     if (isTargetShuffleEquivalent(Mask, {0, 3}) && Subtarget.hasSSE2() &&
         (AllowFloatDomain || !Subtarget.hasSSE41())) {
       std::swap(V1, V2);
       Shuffle = X86ISD::MOVSD;
-      ShuffleVT = MaskVT;
+      SrcVT = DstVT = MaskVT;
       return true;
     }
     if (isTargetShuffleEquivalent(Mask, {4, 1, 2, 3}) &&
         (AllowFloatDomain || !Subtarget.hasSSE41())) {
       Shuffle = X86ISD::MOVSS;
-      ShuffleVT = MaskVT;
+      SrcVT = DstVT = MaskVT;
       return true;
     }
   }
@@ -27413,9 +27413,9 @@ static bool matchBinaryVectorShuffle(MVT MaskVT, ArrayRef<int> Mask,
       (MaskVT.is512BitVector() && Subtarget.hasAVX512())) {
     if (matchVectorShuffleWithUNPCK(MaskVT, V1, V2, Shuffle, IsUnary, Mask, DL,
                                     DAG, Subtarget)) {
-      ShuffleVT = MaskVT;
-      if (ShuffleVT.is256BitVector() && !Subtarget.hasAVX2())
-        ShuffleVT = (32 == EltSizeInBits ? MVT::v8f32 : MVT::v4f64);
+      SrcVT = DstVT = MaskVT;
+      if (MaskVT.is256BitVector() && !Subtarget.hasAVX2())
+        SrcVT = DstVT = (32 == EltSizeInBits ? MVT::v8f32 : MVT::v4f64);
       return true;
     }
   }
@@ -27748,15 +27748,15 @@ static SDValue combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
   }
 
   if (matchBinaryVectorShuffle(MaskVT, Mask, AllowFloatDomain, AllowIntDomain,
-                               V1, V2, DL, DAG, Subtarget, Shuffle, ShuffleVT,
-                               UnaryShuffle)) {
+                               V1, V2, DL, DAG, Subtarget, Shuffle, ShuffleSrcVT,
+                               ShuffleVT, UnaryShuffle)) {
     if (Depth == 1 && Root.getOpcode() == Shuffle)
       return SDValue(); // Nothing to do!
     if (IsEVEXShuffle && (NumRootElts != ShuffleVT.getVectorNumElements()))
       return SDValue(); // AVX512 Writemask clash.
-    V1 = DAG.getBitcast(ShuffleVT, V1);
+    V1 = DAG.getBitcast(ShuffleSrcVT, V1);
     DCI.AddToWorklist(V1.getNode());
-    V2 = DAG.getBitcast(ShuffleVT, V2);
+    V2 = DAG.getBitcast(ShuffleSrcVT, V2);
     DCI.AddToWorklist(V2.getNode());
     Res = DAG.getNode(Shuffle, DL, ShuffleVT, V1, V2);
     DCI.AddToWorklist(Res.getNode());