[Mlir-commits] [mlir] [mlir][tosa] Stop support the custom simplified form of COND_IF (PR #139576)
Luke Hutton
llvmlistbot at llvm.org
Thu May 15 03:28:28 PDT 2025
lhutton1 wrote:
Thanks for the explanation.
> It seems to me that if we want to support this simplified form, a natural approach might be to relax the ERROR_IF checks in verify(). We could assume that when input_list and output_list are empty, their types and shapes match those of then_graph and else_graph.
Agreed, this is what I was alluding to.
> Another consideration is the output format. If we choose to print this simplified form, then any tools consuming the output must also support and understand this implication.
Good point, another instance I can think of here is allowing UI8 types in the validation pass e.g. https://github.com/llvm/llvm-project/blob/main/mlir/lib/Dialect/Tosa/Transforms/TosaValidation.cpp#L1194, where tools that consume the output must interpret and convert to signless. Perhaps these instances are something we can expose as an option in the validation pass which the user can decide to consciously opt-in/out of?
> However, it does require consensus from others. I’d prefer to pause support for it for now and revisit it after discussing with the team.
This seems reasonable to me while we're still pre 1.0, though I do wonder about the impact of downstream users. Is it worth highlighting on discourse and waiting a few days before merging?
https://github.com/llvm/llvm-project/pull/139576
More information about the Mlir-commits
mailing list