[llvm] [AArch64] Lower masked.expandload intrinsic using SVE2p2/SME2p2 EXPAND (PR #190999)

[Paul Walker via llvm-commits]

================
@@ -7436,16 +7365,52 @@ SDValue AArch64TargetLowering::LowerMLOAD(SDValue Op, SelectionDAG &DAG) const {
   SDValue PassThru = LoadNode->getPassThru();
   SDValue Mask = LoadNode->getMask();
 
-  if (PassThru->isUndef() || isZerosVector(PassThru.getNode()))
+  if (!LoadNode->isExpandingLoad()) {
+    if (PassThru->isUndef() || isZerosVector(PassThru.getNode()))
+      return Op;
+
+    SDValue Load = DAG.getMaskedLoad(
+        VT, DL, LoadNode->getChain(), LoadNode->getBasePtr(),
+        LoadNode->getOffset(), Mask, DAG.getUNDEF(VT), LoadNode->getMemoryVT(),
+        LoadNode->getMemOperand(), LoadNode->getAddressingMode(),
+        LoadNode->getExtensionType());
+
+    SDValue Result = DAG.getSelect(DL, VT, Mask, Load, PassThru);
+    return DAG.getMergeValues({Result, Load.getValue(1)}, DL);
+  }
+
+  // Return if EXPAND instruction is not available.
+  if ((!Subtarget->isSVEAvailable() || !Subtarget->hasSVE2p2()) &&
+      (!Subtarget->isSVEorStreamingSVEAvailable() || !Subtarget->hasSME2p2()))
     return Op;
----------------
paulwalker-arm wrote:

This should return `SDValue();` to say the operation is not legal and we don't know how to lower it.

https://github.com/llvm/llvm-project/pull/190999