[PATCH] D105173: [WIP][NFC] factor out unrolling decision logic

Tue Jul 13 12:53:32 PDT 2021

Ali-Sed added a comment.

In D105173#2861546 <https://reviews.llvm.org/D105173#2861546>, @dmgreen wrote:

> The latest revision appears to be missing most of the changes. You may want to squash several commits together in git, and make sure the commit message properly explains the reasoning behind the patch.
>
> Do you intend this to go into LLVM, or are you still experimenting at this point?  It may be best to make sure what you come up with will have practical use before making a lot of changes in llvm itself. My understanding is that this code is (overly) complex but it is the order it is for deliberate reasons. All the existing tests have to keep passing, if they are not already.
>
> I would consider runtime unrolling to be the most interesting part of machine learning loop unroll factors, at least on many cpu's.

Thanks a lot for your feedback.
I just squashed all of the commits now and used "arc diff" for a group of commits and it seems it solved the problem.

I've been decoupling the logic, that unrolling pass uses, from the rest of it. So in a near future, I'll be able to replace it with a learning model.

For now, full unrolling and partial unrolling are the main goals. But, yes definitely! the logic behind the runtime unrolling could be a very interesting choice to be replaced with a learning model too.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105173/new/

https://reviews.llvm.org/D105173