[PATCH] D41953: [LoopUnroll] Unroll and Jam

[Sjoerd Meijer via Phabricator via llvm-commits]

SjoerdMeijer added a comment.

> Any suggestions on what would be at the top of the list? :)

For me the first thing would probably be a rather straightforward thing: more OptimizationRemarks. I think it currently reports only something after a successful transformation. But perhaps more interesting to know is why it isn't triggering (if that's what you expect): so optimisation remarks about the legality, profitability, or any other reason, would be crucial to understand what's going on. Then, based on this optimisation report, and if the transformation really needs to be triggered, the user can resort to pragmas (for which you have another patch) and/or tweaking options for treshhold values.

This is an implementation detail, but I am still not sure about some of the terminology that is used, e.g. "Fore", "Subloops", etc., but I've not come up yet with convincing alternatives...

Also, we probably want to revise some of the threshold values, and see if we need to tweak them further.

But assuming that the dependence analysis is now conservative/correct, these are the sort of things I was thinking of that we can improve after this lands and we get some more experience  with the pass.


Repository:
  rL LLVM

https://reviews.llvm.org/D41953