[llvm] [AArch64][CostModel] Improve cost estimate of scalarizing a vector di… (PR #118055)

[Alexey Bataev via llvm-commits]

================
@@ -3472,29 +3472,41 @@ InstructionCost AArch64TTIImpl::getArithmeticInstrCost(
           Cost *= 4;
         return Cost;
       } else {
-        // If one of the operands is a uniform constant then the cost for each
-        // element is Cost for insertion, extraction and division.
-        // Insertion cost = 2, Extraction Cost = 2, Division = cost for the
-        // operation with scalar type
-        if ((Op1Info.isConstant() && Op1Info.isUniform()) ||
-            (Op2Info.isConstant() && Op2Info.isUniform())) {
-          if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
+        if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
+          if ((Op1Info.isConstant() && Op1Info.isUniform()) ||
+              (Op2Info.isConstant() && Op2Info.isUniform())) {
             InstructionCost DivCost = BaseT::getArithmeticInstrCost(
                 Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info);
-            return (4 + DivCost) * VTy->getNumElements();
+            // If #vector_elements = n then we need
+            // n inserts + 2n extracts + n divisions.
+            InstructionCost InsertExtractCost =
+                ST->getVectorInsertExtractBaseCost();
+            Cost = (3 * InsertExtractCost + DivCost) * VTy->getNumElements();
+          } else if (!Scalars.empty()) {
----------------
alexey-bataev wrote:

Why Args is not enough for this analysis?

https://github.com/llvm/llvm-project/pull/118055