[Mlir-commits] [mlir] [mlir][vector] Deal with special patterns when emulating masked load/store (PR #75587)

[Jakub Kuderski]

kuhar wrote:

So what I suggest is to handle 3 cases for the mask:
- All ones --> generate an unmasked load/store. This should be done for the folders for `vector.maskedload`/`maskedstore`. 
- All zeros --> get the passthrough. This should be done for the folders for `vector.maskedload`/`maskedstore`.
- Everything else --> lower as-is and rely on the folders to simplify any known zero/one bits. Fix the other folders as necessary (`vector.create_mask`, `vector.extract` from mask.

https://github.com/llvm/llvm-project/pull/75587