[Mlir-commits] [mlir] [mlir][gpu] Avoid duplicate attrs when lowering gpu.func (PR #205035)

[Fabian Mora]

fabianmcg wrote:

> I don't quite see why? This does not seem to be in line with any of the other dialects and lowering.
>
> Aligning with the expected historical and current behavior in MLIR, the lowering needs to be changed to drop these discardable attributes.
> 
> You're proposing a change that may makes sense, but does not belong to a specific op verifier IMO (why would we have this here and not for `arith` dialect with all the other inherent attributes?)

That's why I prefer the general fix in `Verifier.cpp`. My issue is, `gpu.known_block_size` will trip users even outside of this lowering, eg. int range analysis will not produce the expected results and the user will not know why, so from a user perspective I'd prefer an error as early as possible.

https://github.com/llvm/llvm-project/pull/205035