[X86] Fold (v2i64 (scalar_to_vector (i64 (bitcast (double))))) -> (bitcast (v2f64 scalar_to_vector))

This can occur more frequently after the MMX retirement if anyone is still using MMX intrinsics that now wrap to SSE

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -58359,10 +58359,16 @@ static SDValue combineScalarToVector(SDNode *N, SelectionDAG &DAG,
                                   DAG.getZExtOrTrunc(ZeroExt, DL, MVT::i32))));
   }
 
-  // Combine (v2i64 (scalar_to_vector (i64 (bitconvert (mmx))))) to MOVQ2DQ.
-  if (VT == MVT::v2i64 && Src.getOpcode() == ISD::BITCAST &&
-      Src.getOperand(0).getValueType() == MVT::x86mmx)
-    return DAG.getNode(X86ISD::MOVQ2DQ, DL, VT, Src.getOperand(0));
+  if (VT == MVT::v2i64 && Src.getOpcode() == ISD::BITCAST) {
+    SDValue SrcOp = Src.getOperand(0);
+    // Combine (v2i64 (scalar_to_vector (i64 (bitcast (double))))) to MOVQ.
+    if (SrcOp.getValueType() == MVT::f64)
+      return DAG.getBitcast(
+          VT, DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, SrcOp));
+    // Combine (v2i64 (scalar_to_vector (i64 (bitcast (mmx))))) to MOVQ2DQ.
+    if (SrcOp.getValueType() == MVT::x86mmx)
+      return DAG.getNode(X86ISD::MOVQ2DQ, DL, VT, SrcOp);
+  }
 
   // See if we're broadcasting the scalar value, in which case just reuse that.
   // Ensure the same SDValue from the SDNode use is being used.

llvm/test/CodeGen/X86/mmx-cvt.ll

@@ -287,8 +287,7 @@ define noundef <2 x i64> @cvt_f64_v2i64(double %a0) {
 ;
 ; X64-LABEL: cvt_f64_v2i64:
 ; X64:       # %bb.0:
-; X64-NEXT:    movq %xmm0, %rax
-; X64-NEXT:    movq %rax, %xmm0
+; X64-NEXT:    movq {{.*#+}} xmm0 = xmm0[0],zero
 ; X64-NEXT:    retq
   %bc = bitcast double %a0 to <1 x i64>
   %r = shufflevector <1 x i64> %bc, <1 x i64> zeroinitializer, <2 x i32> <i32 0, i32 1>