[Mlir-commits] [mlir] [mlir][linalg] Fix #93973 - linalg reduce verifier crash (PR #119871)

[Renato Golin]

https://github.com/rengolin commented:

I don't even know what does it mean to have multiple inputs and multiple outputs. Your test would just pass if I passed a bogus second output and had the same code you had originally with the second output untouched.

This means the semantics of `linalg.reduce` is broken.

I think we need to decide if we want a reduction that is no unary in the first place and then just check for `ins == outs == 1`.

Can you add an RFC to the forum to discuss that, please?

https://github.com/llvm/llvm-project/pull/119871