[llvm] [AArch64] Fold scalar-to-vector shuffles into DUP/FMOV (PR #166962)

[Amina Chabane via llvm-commits]

================
@@ -15654,6 +15654,41 @@ SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
     }
   }
 
+  // 128-bit NEON vectors: writing to the 64-bit low half with
+  // DUP/SCALAR_TO_VECTOR already zeroes the other 64 bits, so if the low half
+  // is a splat and the upper half is zero/undef we can materialise just that
+  // low half.
+  if (VT.isFixedLengthVector() && VT.getSizeInBits() == 128) {
+    EVT LaneVT = VT.getVectorElementType();
+    const unsigned HalfElts = NumElts >> 1;
+    SDValue FirstVal = Op.getOperand(0);
+
+    auto IsZero = [&](SDValue V) {
+      if (isNullConstant(V))
+        return true;
+      return V.getValueType().isFloatingPoint() && isNullFPConstant(V);
----------------
Amichaxx wrote:

Thanks.

https://github.com/llvm/llvm-project/pull/166962