[llvm] [CostModel] getInstructionCost - improve estimation of costs for length changing shuffles (PR #84156)

[Alexey Bataev via llvm-commits]

================
@@ -1356,7 +1357,35 @@ class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
               DemandedDstElts, CostKind);
         }
 
-        return CostKind == TTI::TCK_RecipThroughput ? -1 : 1;
+        bool IsUnary = isa<UndefValue>(Operands[1]);
+        NumSubElts = VecSrcTy->getElementCount().getKnownMinValue();
+        SmallVector<int, 16> AdjustMask(Mask.begin(), Mask.end());
+
+        // Widening shuffle - widening the source(s) to the new length
+        // (treated as free - see above), and then perform the adjusted
+        // shuffle at that width.
+        if (Shuffle->increasesLength()) {
+          for (int &M : AdjustMask)
+            M = M >= NumSubElts ? (M + (Mask.size() - NumSubElts)) : M;
+
+          return TargetTTI->getShuffleCost(
+              IsUnary ? TTI::SK_PermuteSingleSrc : TTI::SK_PermuteTwoSrc, VecTy,
+              AdjustMask, CostKind, 0, nullptr);
+        }
+
+        // Narrowing shuffle - perform shuffle at original wider width and
+        // then extract the lower elements.
+        AdjustMask.append(NumSubElts - Mask.size(), -1);
----------------
alexey-bataev wrote:

```suggestion
        AdjustMask.append(NumSubElts - Mask.size(), PoisonMaskElem);
```

https://github.com/llvm/llvm-project/pull/84156