[Mlir-commits] [mlir] [mlir][arith] Add more canonicalization and integration tests coverage (PR #92272)

[Mehdi Amini]

joker-eph wrote:

>However, checking that "the translated code is equal to what we expect" does not give us confidence in correctness here - it is only checking the rewrite pattern has done its job. 

I don't quite follow the difference you're trying to make.

> I believe it's clearly desirable to test the functionality of the lowering, though. And I don't see how this can be done without something to the effect of an integration test.

I feel we're running in circle: we're checking this because the IR you generate to has a spec and you check manually the equivalence by inspecting the IR ("running mentally" is what you wrote).
Again virtually the entirety of transformations in LLVM are developed that way ~ forever, we could change our testing approach of course, but IMO that's an RFC for the LLVM discourse.

In the meantime, tests like [this](https://github.com/llvm/llvm-project/pull/96974/files) are out-of-scope of what I expect to see here.

https://github.com/llvm/llvm-project/pull/92272