[Mlir-commits] [mlir] [mlir][gpu] Preserve argument attributes during kernel outlining (PR #221465)

[Federico Bruzzone]

FedericoBruzzone wrote:

@purnima-nlp yes, this matches my understanding of @fabianmcg's suggestion.

I would also lean towards making the mapping between captured operands and their attr dictionaries explicit in `gpu.launch`, and then having the outlining pass propagate those attrs to the corresponding `gpu.func` args.
For the exact representation, I would defer to @fabianmcg, but a per-captured-operand attribute dictionary seems reasonable.

And yes, this is a separate prerequisite patch: first extend `gpu.launch` with the necessary representation, then update this PR to use it :D


https://github.com/llvm/llvm-project/pull/221465