================
@@ -302,17 +303,26 @@ deriveNeonSISDIntrinsicOperandTypes(CIRGenFunction &cgf,
unsigned modifier,
else if (vecArgTy && !(modifier & AddRetType))
funcResTy = wrapAsVector(resultTy);
- // When VectorizeArgTypes is set, wrap every operand that has the same
- // scalar type as arg0 into a vector. This covers intrinsics with multiple
- // data operands of the same type (e.g. vsri takes two data operands,
- // both of which must be wrapped into the same vector type).
+ // LLVMExtendedType<0> preserves the result vector's lane count while
+ // widening its element type. Reconstruct that source type from the Clang
+ // builtin's scalar data type.
+ if (modifier & WidenArgs) {
+ auto resVecTy = mlir::dyn_cast<cir::VectorType>(funcResTy);
+ assert(resVecTy && "widened SISD arguments require a vector result");
+ vecArgTy = cir::VectorType::get(arg0Ty, resVecTy.getSize());
+ }
+
+ // Wrap every non-immediate data operand that has the same scalar type as
+ // arg0. Checking the ICE bitmap is required when a data operand and an
+ // immediate both have i32 type (e.g. vqshrns_n_s32).
----------------
iamvickynguyen wrote:
Thank you for the refactor!
I find it would be better if this comment includes which flags make `vecArgTy`
non-empty; otherwise, we have to read all way up to know whether wrapping
happens.
Something like:
```suggestion
// Runs when VectorizeArgTypes or WidenArgs marked this intrinsic for
// vectorized arguments. Wrap every non-immediate data operand that has the
same scalar type as
// arg0. Checking the ICE bitmap is required when a data operand and an
// immediate both have i32 type (e.g. vqshrns_n_s32).
```
https://github.com/llvm/llvm-project/pull/209389
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
