[Mlir-commits] [mlir] [mlir][arith] doc updates for ub semantics, and int representations (PR #72932)

[Jakub Kuderski]

kuhar wrote:

> For now, I only found a single bug, which is the canonicalization of `select`, which is more poisonous if we were to chose these semantics. I can report it if we decide to agree on these semantics.

@math-fehr Can you share the details? Are we over-defining the behavior in this PR, or is it something else with the canonicalization pattern?

> I see, that's fair enough. @joker-eph Am I correct in understanding that the only blocker is that we'd like to leave the overflow behaviour undefined for now, before we figure out if we want to adopt the LLVM-style opt-in poison semantics? (and otherwise okay?)

@pingshiyu Yes, if we can't resolve the disagreement we can take this bit out of the PR, land something we all agree on, and iterate on the rest separately.

> > LGTM. Let's give it a few days for other folks to weight in, especially with Thanksgiving just around the corner.
> 
> In case it's not clear: my comment is clearly blocking. You're misleading the author in thinking that this can be merged here: please be careful in providing explicit conditional approval ("LGTM but please wait for acknowledgement from XYZ").

@joker-eph I reread this comment and I think we are on the same page here. To me my message conveys the same meaning as what you put in parentheses.

https://github.com/llvm/llvm-project/pull/72932