1 files changed, 28 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f654f1d7a72..9f6a0b76e06 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -44659,6 +44659,34 @@ static SDValue combinePMULDQ(SDNode *N, SelectionDAG &DAG,
   if (TLI.SimplifyDemandedBits(SDValue(N, 0), APInt::getAllOnesValue(64), DCI))
     return SDValue(N, 0);
 
+  // If the input is an extend_invec and the SimplifyDemandedBits call didn't
+  // convert it to any_extend_invec, due to the LegalOperations check, do the
+  // conversion directly to a vector shuffle manually. This exposes combine
+  // opportunities missed by combineExtInVec not calling
+  // combineX86ShufflesRecursively on SSE4.1 targets.
+  // FIXME: This is basically a hack around several other issues related to
+  // ANY_EXTEND_VECTOR_INREG.
+  if (N->getValueType(0) == MVT::v2i64 && LHS.hasOneUse() &&
+      (LHS.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG ||
+       LHS.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG) &&
+      LHS.getOperand(0).getValueType() == MVT::v4i32) {
+    SDLoc dl(N);
+    LHS = DAG.getVectorShuffle(MVT::v4i32, dl, LHS.getOperand(0),
+                               LHS.getOperand(0), { 0, -1, 1, -1 });
+    LHS = DAG.getBitcast(MVT::v2i64, LHS);
+    return DAG.getNode(N->getOpcode(), dl, MVT::v2i64, LHS, RHS);
+  }
+  if (N->getValueType(0) == MVT::v2i64 && RHS.hasOneUse() &&
+      (RHS.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG ||
+       RHS.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG) &&
+      RHS.getOperand(0).getValueType() == MVT::v4i32) {
+    SDLoc dl(N);
+    RHS = DAG.getVectorShuffle(MVT::v4i32, dl, RHS.getOperand(0),
+                               RHS.getOperand(0), { 0, -1, 1, -1 });
+    RHS = DAG.getBitcast(MVT::v2i64, RHS);
+    return DAG.getNode(N->getOpcode(), dl, MVT::v2i64, LHS, RHS);
+  }
+
   return SDValue();
 }