[llvm] [LoopFusion] Assert failure in the issue 80301 (PR #167837)

[Ehsan Amiri via llvm-commits]

amehsan wrote:

> Running testing and fuzzers on loop fusion was a first exercise to judge the quality of the pass, and see how much work it would be. As I mentioned, we are mostly focused on interchange, but we are happy to support you with fusion. For example, we have data where fusion improves workloads if we need to justify enabling this and if it is worth the compile-time. And we can help with point fixes like this, although you seem to suggest the pass may need a bit of an overhaul?

Yes, we are interested in loop fusion. This past week we have been busy with DA discussions and internal stuff but I expect we speed up on loop fusion soon.

With regard to major changes in fusion. yes, I think at  least two major improvements is needed (and both should help with compile time). One is the comment above, the other the changes in how we check dependency that I previously commented on one of the issues. But I think before we finalize the second one, we need to wait for the current investigations in DA to conclude (basically it is a discussion of what are the conditions for correctness of DA)

If you are interested in merging this fix as is, it is fine with me. we can merge this and we will address the issue of multiple different control flow checks separately.

And sure, we can discuss the status both in the loop opt meeting and also here or elsewhere on github.


https://github.com/llvm/llvm-project/pull/167837