[Mlir-commits] [mlir] [mlir][Vector] Reject scalable masks in maskedload/maskedstore emulation (PR #223316)

[Hung-Kuan Tseng]

https://github.com/Tim096 commented:

Thanks for fixing this! I checked it locally: without the patch, `--test-vector-emulate-masked-load-store` turns the `vector<[4]xi1>` maskedload into exactly four `scf.if` blocks with the insert positions fixed at 0..3, so lanes 4 and up are dropped whenever vscale > 1, and the new `CHECK: vector.maskedload` catches that.

On the tests: the MLIR Testing Guide asks for negative tests to be prefixed with `negative_` (https://mlir.llvm.org/getting_started/TestingGuide/#test-naming-convention), so something like `@negative_maskedload_scalable_mask` / `@negative_maskedstore_scalable_mask`, which would also cover @FedericoBruzzone's point about moving the comment into the name.

Unrelated to this PR, but the maskedload converter's existing rank check says "expected vector.maskedstore with 1-D mask" (L58). If you touch that area anyway, it could say `vector.maskedload`.


https://github.com/llvm/llvm-project/pull/223316