1 files changed, 28 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ead00a9..b4299d5 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -42625,6 +42625,34 @@ static SDValue combineOr(SDNode *N, SelectionDAG &DAG,
   if (SDValue R = combineLogicBlendIntoPBLENDV(N, DAG, Subtarget))
     return R;
 
+  // Combine OR(X,KSHIFTL(Y,Elts/2)) -> CONCAT_VECTORS(X,Y) == KUNPCK(X,Y).
+  // Combine OR(KSHIFTL(X,Elts/2),Y) -> CONCAT_VECTORS(Y,X) == KUNPCK(Y,X).
+  // iff the upper elements of the non-shifted arg are zero.
+  // KUNPCK require 16+ bool vector elements.
+  if (N0.getOpcode() == X86ISD::KSHIFTL || N1.getOpcode() == X86ISD::KSHIFTL) {
+    unsigned NumElts = VT.getVectorNumElements();
+    unsigned HalfElts = NumElts / 2;
+    APInt UpperElts = APInt::getHighBitsSet(NumElts, HalfElts);
+    if (NumElts >= 16 && N1.getOpcode() == X86ISD::KSHIFTL &&
+        N1.getConstantOperandAPInt(1) == HalfElts &&
+        DAG.MaskedValueIsZero(N0, APInt(1, 1), UpperElts)) {
+      SDLoc dl(N);
+      return DAG.getNode(
+          ISD::CONCAT_VECTORS, dl, VT,
+          extractSubVector(N0, 0, DAG, dl, HalfElts),
+          extractSubVector(N1.getOperand(0), 0, DAG, dl, HalfElts));
+    }
+    if (NumElts >= 16 && N0.getOpcode() == X86ISD::KSHIFTL &&
+        N0.getConstantOperandAPInt(1) == HalfElts &&
+        DAG.MaskedValueIsZero(N1, APInt(1, 1), UpperElts)) {
+      SDLoc dl(N);
+      return DAG.getNode(
+          ISD::CONCAT_VECTORS, dl, VT,
+          extractSubVector(N1, 0, DAG, dl, HalfElts),
+          extractSubVector(N0.getOperand(0), 0, DAG, dl, HalfElts));
+    }
+  }
+
   // Attempt to recursively combine an OR of shuffles.
   if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) {
     SDValue Op(N, 0);