1 files changed, 34 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e81339d8eb7..2475dc46e02 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6676,6 +6676,40 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
     return true;
   }
   case ISD::OR: {
+    // Inspect each operand at the byte level. We can merge these into a
+    // blend shuffle mask if for each byte at least one is masked out (zero).
+    KnownBits Known0 = DAG.computeKnownBits(N.getOperand(0));
+    KnownBits Known1 = DAG.computeKnownBits(N.getOperand(1));
+    if (Known0.One.isNullValue() && Known1.One.isNullValue()) {
+      bool IsByteMask = true;
+      unsigned NumSizeInBytes = NumSizeInBits / 8;
+      unsigned NumBytesPerElt = NumBitsPerElt / 8;
+      APInt ZeroMask = APInt::getNullValue(NumBytesPerElt);
+      APInt SelectMask = APInt::getNullValue(NumBytesPerElt);
+      for (unsigned i = 0; i != NumBytesPerElt && IsByteMask; ++i) {
+        unsigned LHS = Known0.Zero.extractBits(8, i * 8).getZExtValue();
+        unsigned RHS = Known1.Zero.extractBits(8, i * 8).getZExtValue();
+        if (LHS == 255 && RHS == 0)
+          SelectMask.setBit(i);
+        else if (LHS == 255 && RHS == 255)
+          ZeroMask.setBit(i);
+        else if (!(LHS == 0 && RHS == 255))
+          IsByteMask = false;
+      }
+      if (IsByteMask) {
+        for (unsigned i = 0; i != NumSizeInBytes; i += NumBytesPerElt) {
+          for (unsigned j = 0; j != NumBytesPerElt; ++j) {
+            unsigned Ofs = (SelectMask[j] ? NumSizeInBytes : 0);
+            int Idx = (ZeroMask[j] ? SM_SentinelZero : (i + j + Ofs));
+            Mask.push_back(Idx);
+          }
+        }
+        Ops.push_back(N.getOperand(0));
+        Ops.push_back(N.getOperand(1));
+        return true;
+      }
+    }
+
     // Handle OR(SHUFFLE,SHUFFLE) case where one source is zero and the other
     // is a valid shuffle index.
     SDValue N0 = peekThroughOneUseBitcasts(N.getOperand(0));