[Mlir-commits] [mlir] [mlir] Add normalize pass to MLIR (PR #186647)

[Renato Golin]

rengolin wrote:

> Under the assumption that an operation cannot define differing semantics for captured vs non-captured, then moving pure operations out of a region is basically always legal.

Yes, but if the semantics isn't clear (ie. there _isn't_ an interface that defines it), then different compilers can _assume_ different things and they're not _technically wrong_. If there's no way to define it, than it will break those assumptions and there will be no way to "fix" it. But as I said above, since this pass is just for diffing output, it should be fine?

> I think this is important to note such that no one accidentally uses it instead of the canonicalization pass (a pass that aims for increasing semantic equivalence instead).

This is key and would alleviate most if not all problems, I think. Clear documentation on what the pass is doing and the dangers within.

https://github.com/llvm/llvm-project/pull/186647