[AArch64] Combing scalar_to_reg into DUP if the DUP already exists (llvm#160499)

If we already have a dup(x) as part of the DAG along with a
scalar_to_vec(x), we can re-use the result of the dup to the
scalar_to_vec(x).

Author: davemgreen

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -27602,6 +27602,15 @@ static SDValue performPTestFirstCombine(SDNode *N,
 static SDValue
 performScalarToVectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                              SelectionDAG &DAG) {
+  SDLoc DL(N);
+
+  // If a DUP(Op0) already exists, reuse it for the scalar_to_vector.
+  if (DCI.isAfterLegalizeDAG()) {
+    if (SDNode *LN = DCI.DAG.getNodeIfExists(AArch64ISD::DUP, N->getVTList(),
+                                             N->getOperand(0)))
+      return SDValue(LN, 0);
+  }
+
   // Let's do below transform.
   //
   //         t34: v4i32 = AArch64ISD::UADDLV t2
@@ -27638,7 +27647,6 @@ performScalarToVectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
     return SDValue();
 
   // Let's generate new sequence with AArch64ISD::NVCAST.
-  SDLoc DL(N);
   SDValue EXTRACT_SUBVEC =
       DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, UADDLV,
                   DAG.getConstant(0, DL, MVT::i64));

llvm/test/CodeGen/AArch64/aarch64-matrix-umull-smull.ll

@@ -1371,11 +1371,10 @@ define noundef <8 x i16> @cmplx_mul_combined_re_im(<8 x i16> noundef %a, i64 %sc
 ; CHECK-SD-NEXT:    lsr x9, x0, #16
 ; CHECK-SD-NEXT:    adrp x8, .LCPI14_0
 ; CHECK-SD-NEXT:    dup v4.8h, w0
-; CHECK-SD-NEXT:    dup v1.8h, w9
-; CHECK-SD-NEXT:    fmov s3, w9
-; CHECK-SD-NEXT:    sqneg v2.8h, v1.8h
-; CHECK-SD-NEXT:    ldr q1, [x8, :lo12:.LCPI14_0]
-; CHECK-SD-NEXT:    tbl v1.16b, { v2.16b, v3.16b }, v1.16b
+; CHECK-SD-NEXT:    ldr q3, [x8, :lo12:.LCPI14_0]
+; CHECK-SD-NEXT:    dup v2.8h, w9
+; CHECK-SD-NEXT:    sqneg v1.8h, v2.8h
+; CHECK-SD-NEXT:    tbl v1.16b, { v1.16b, v2.16b }, v3.16b
 ; CHECK-SD-NEXT:    rev32 v2.8h, v0.8h
 ; CHECK-SD-NEXT:    sqdmull v3.4s, v0.4h, v4.4h
 ; CHECK-SD-NEXT:    sqdmull2 v0.4s, v0.8h, v4.8h