[llvm] [TTI] Include scalarization overhead for icmp/fcmp result (PR #206697)

[Simon Pilgrim via llvm-commits]

================
@@ -1464,9 +1464,14 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
 
       // Return the cost of multiple scalar invocation plus the cost of
       // inserting and extracting the values.
-      return getScalarizationOverhead(ValVTy, /*Insert*/ true,
-                                      /*Extract*/ false, CostKind) +
-             Num * Cost;
+      InstructionCost Overhead =
+          getScalarizationOverhead(ValVTy, /*Insert*/ false,
+                                   /*Extract*/ true, CostKind);
+      if (Opcode == Instruction::ICmp || Opcode == Instruction::FCmp)
+        Overhead +=
+            getScalarizationOverhead(cast<VectorType>(CondTy), /*Insert*/ true,
+                                     /*Extract*/ false, CostKind);
+      return Overhead + Num * Cost;
----------------
RKSimon wrote:

Still not sure this is correct, for the scalarization overheads shouldn't it be:

icmp/fcmp: overhead = 2 * extract of the ValVTy + 1 * insert of CondTy

select: overhead = 2 * extract of the ValVTy + 1 * extract of CondTy + 1 * insert of ValVTy 

Ideally we could be using the getOperandsScalarizationOverhead helpers?

https://github.com/llvm/llvm-project/pull/206697