[llvm] [AArch64] Improve lowering of truncating build vectors (PR #81960)

[David Green via llvm-commits]

================
@@ -19096,6 +19148,28 @@ static SDValue performBuildVectorCombine(SDNode *N,
   SDLoc DL(N);
   EVT VT = N->getValueType(0);
 
+  //    BUILD_VECTOR (extract_elt(Assert[S|Z]ext(x)))
+  // => BUILD_VECTOR (extract_elt(x))
+  SmallVector<SDValue, 8> Ops;
+  bool ExtractExtended = false;
+  for (SDValue Extr : N->ops()) {
+    if (Extr.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
+      ExtractExtended = false;
+      break;
+    }
+    SDValue ExtractBase = Extr.getOperand(0);
+    if (ExtractBase.getOpcode() == ISD::AssertSext ||
+        ExtractBase.getOpcode() == ISD::AssertZext) {
+      ExtractExtended = true;
+      Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
----------------
davemgreen wrote:

This loses information, which might not always be a good thing. In general it is better to check something can be transformed, then create nodes if we know it can be. But it may be better to look through AssertSext/AssertZext in the transform as opposed to removing the nodes.

https://github.com/llvm/llvm-project/pull/81960