[PATCH] D110057: [LoopFlatten] Move it to a LoopPassManager

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

In D110057#3048181 <https://reviews.llvm.org/D110057#3048181>, @dmgreen wrote:

> New pass manager loop pass invalidation isn't exactly my area of expertise. My main issue with moving the pass was that performance was worse. I don't now remember why, but I have a vague memory of widening IV's in the flattening pass, if they have already been partially widened in indvarsimplify.

Would it be possible to reevaluate whether that issue still exists? If it is important for performance to have LoopFlatten in a standalone LPM, we should have a PhaseOrdering test that demonstrates that (which would have failed with this patch and made the issue obvious).


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D110057/new/

https://reviews.llvm.org/D110057