[Mlir-commits] [mlir] [MLIR][XeVM] Update XeVM prefetch ops. (PR #166445)

[Sang Ik Lee]

silee2 wrote:

> Thanks for the example, now I see the issue.
> 
> I quickly checked on an example using only `xevm.prefetch` which originally doesn't define any side effects. Ops without known side effects are generally not eliminated due to conservative evaluation. Same applies to the XeGPU prefetch ops that also don't define side effects. As such, I think assigning write side effect might be stronger "guard" against unwanted modifications as it also forces passes to be more pessimistic.
> 
> Could you add a small test case just to avoid regression in the future?

Instead of adding new tests, updated XeGPU to XeVM conversion tests to call canonicalize pass.
This change will make sure XeVM  prefetch ops do not get folded away.

https://github.com/llvm/llvm-project/pull/166445