[llvm] [SLP]Fix perfect diamond match with extractelements in scalars (PR #132466)

[Simon Pilgrim via llvm-commits]

================
@@ -5339,8 +5338,19 @@ static InstructionCost getScalarizationOverhead(const TargetTransformInfo &TTI,
     }
     return Cost;
   }
-  return TTI.getScalarizationOverhead(Ty, DemandedElts, Insert, Extract,
-                                      CostKind, VL);
+  APInt NewDemandedElts = DemandedElts;
+  InstructionCost Cost = 0;
+  if (!ForPoisonSrc && Insert) {
+    // Handle insert into non-poison vector.
----------------
RKSimon wrote:

I'm not certain this is entirely correct, afaict I can tell you are trying to cost inserting a number of elements into an existing (non-poison) vector, while the TTI::getScalarizationOverhead assumes Insert is for a ISD::BUILD_VECTOR style pattern. Would a SK_Select shuffle not be closer to what you're after? The isZero checks below suggest you've had to write this mainly for the single element diamond case, so maybe that need to be handled by a DemandedElts.isOneBitSet() special case instead?

https://github.com/llvm/llvm-project/pull/132466