[Mlir-commits] [mlir] [mlir][linalg] Extend `FuseElementwiseOps` pattern to work with named ops (PR #144922)

[llvmlistbot at llvm.org]

srcarroll wrote:

> > Having to change user code due to upstream changes is just how using open source code works. That's not a good reason to avoid useful changes in upstream. I don't consider this "breaking the world", and that's a pretty alarmist response that isn't helpful code review. And by the way I have to make non-trivial changes all the time when updating upstream, and I'm the only compiler engineer in my company. So I have no sympathy for other users having to change code, especially those from big companies that have plenty of engineering resources.
> > If there's a specific concern, then please elaborate. Don't just expect the worst and guess that things will break.
> 
> I dont understand the problem here. Its just a new entry point that reuses the code. Dont see why so much push back against something that straight-forward. You are saying this is a small change. As a long-time maintainer (and someone who wrote most of this code), I am saying this is a fairly substantial change, and its better to just add a new entry point that handles named ops + generic ops. I think thats quite reasonable.

I'm more objecting to the reason for your suggestion more than pushing back on the change itself.  Your suggestion is predicated on the fact that you don't want to make changes in your downstream project.   This will never be a valid reason to me and I think most MLIR contributors that aren't invested in their org's self interests would agree with me as well.

If the reason was more of a matter of actual functionality, then fine.  If this PR in its current state makes it impossible for you to make changes to keep the same behavior, then I'd be happy make further changes to avoid that.  Moreover, if your suggestion is useful in isolation as more of a code organization thing, then I'm also happy to accommodate.

But the only reason you've given me for the change is that "I dont want to update my downstream code", and sorry but I couldn't care less about that weak excuse.  I get this excuse all the time, and I've long since lost my patience for hearing it, so I'm sorry for being blunt about this.

Nevertheless, I will put in some effort to think about your suggestion more and decide for myself if it makes sense on its own outside the context of "I dont want to update my code".

https://github.com/llvm/llvm-project/pull/144922