[Mlir-commits] [mlir] [mlir][tosa] Fail lowering of `tosa.pad` -> `tensor.pad` with -1 padding (PR #232221)

[Renato Golin]

rengolin wrote:

> IIRC the non-negative `tosa.pad` verifier checks didn't always exist and when it came to adding them, it broke some existing tests e.g. {...} , leading us to believe there was a use-case. I would support disallowing -1 in the `tosa.pad` verifier.

Right, I was going to ask this, because the test that this PR disables was hinting that it's accepted and legal.

> with the `tosa-validate` pass being used to check conformance against the specification.

This would work around the SSA value vs attribute problem, since in this pass you can inspect the graph, not just restricted to the op.

This would be prefered to lowering a run time check that shouldn't be there in the first place.

https://github.com/llvm/llvm-project/pull/232221