[llvm] [AArch64] Improve cost model for legal subvec insert/extract (PR #81135)

[Paul Walker via llvm-commits]

================
@@ -568,6 +573,39 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
     }
     return Cost;
   }
+  case Intrinsic::vector_extract:
+  case Intrinsic::vector_insert: {
+    // If both the vector and subvector types are legal types and the index
+    // is 0, then this should be a no-op or simple operation; return a
+    // relatively low cost.
+
+    // If arguments aren't actually supplied, then we cannot determine the
+    // value of the index. We also want to skip predicate types.
+    if (ICA.getArgs().size() != ICA.getArgTypes().size() ||
+        ICA.getReturnType()->getScalarType()->isIntegerTy(1))
+      break;
+
+    LLVMContext &C = RetTy->getContext();
+    EVT VecVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
+    bool IsExtract = ICA.getID() == Intrinsic::vector_extract;
+    EVT SubVecVT = IsExtract ? getTLI()->getValueType(DL, RetTy)
+                             : getTLI()->getValueType(DL, ICA.getArgTypes()[1]);
+    // Skip this if either the vector or subvector types are unpacked
+    // SVE types; they may get lowered to stack stores and loads.
+    if (isUnpackedVectorVT(VecVT) || isUnpackedVectorVT(SubVecVT))
+      break;
+
+    TargetLoweringBase::LegalizeKind SubVecLK =
+        getTLI()->getTypeConversion(C, SubVecVT);
+    TargetLoweringBase::LegalizeKind VecLK =
+        getTLI()->getTypeConversion(C, VecVT);
+    const Value *Idx = IsExtract ? ICA.getArgs()[1] : ICA.getArgs()[2];
+    const ConstantInt *CIdx = dyn_cast<ConstantInt>(Idx);
----------------
paulwalker-arm wrote:

Can this be just `cast<>` and the null check removed?  I ask because the LangRef says the Idx operand must be a constant integer.

https://github.com/llvm/llvm-project/pull/81135