[Mlir-commits] [mlir] [ROCDL] Use attached target on the GPU module when lowering GPU ops to ROCDL (PR #110735)

[Fabian Mora]

fabianmcg wrote:

Apologies, I forgot. 

> @fabianmcg Now that I'm back from vacation, care to expand on the infer/`--convert-to-llvm` incompatibility?
> 
> I'd think we could implement MLIR data layout stuff for the target attribute

The main source of incompatibility with `convert-to-llvm` is that the pass doesn't take options and uses `Pass::initialize` to get all the patterns from  the loaded dialects. Therefore it has no view of the IR and cannot check what's the target in the IR. At the conference Mehdi agreed that's an issue, now we need to find a performant way we all agree to solve the issue in `convert-to-llvm`.

So, I don't have any opposition to this patch if it helps downstream, but it won't work right away with convert-to-llvm until we fix it.

https://github.com/llvm/llvm-project/pull/110735