[llvm] [AArch64][TTI] Improve `LegalVF` when computing gather-loads cost (PR #69617)

[Alexey Bataev via llvm-commits]

================
@@ -2989,7 +2989,14 @@ InstructionCost AArch64TTIImpl::getGatherScatterOpCost(
       ElementCount::getScalable(1))
     return InstructionCost::getInvalid();
 
-  ElementCount LegalVF = LT.second.getVectorElementCount();
+  ElementCount LegalVF;
+  if (LT.second.isVector()) {
+    LegalVF = LT.second.getVectorElementCount();
+  } else {
+    // If the legalized type is a simple type, treat it as a 1-element vector.
+    LegalVF = ElementCount::getFixed(1);
+  }
+
----------------
alexey-bataev wrote:

If thу vector type gets scalarized (like in your check), isLegalMaskedGather() shall return false for such types, even though useNeonVector() returns true. SLP shall not assume emitting masked gather here at all, the scalarization increases the cost significantly, we should end up with the simple gather here

https://github.com/llvm/llvm-project/pull/69617