[PATCH] D140698: [LoopUnroll] Be more permissive to high-cost loop trip count SCEV's

Tue Dec 27 13:12:29 PST 2022

nikic added a reviewer: fhahn.
nikic added a comment.

Compile-time: https://llvm-compile-time-tracker.com/compare.php?from=781eabeb40b8e47e3a46b0b927784e63f0aad9ab&to=c0e5cd413681c1622e6d454bba677cf8847c3db3&stat=instructions:u ThinLTO +0.5% geomean with +2.7% tramp3d-v4

Code-size: https://llvm-compile-time-tracker.com/compare.php?from=781eabeb40b8e47e3a46b0b927784e63f0aad9ab&to=c0e5cd413681c1622e6d454bba677cf8847c3db3&stat=size-text ThinLTO +1.0% geomean with +6.7% tramp3d-v4

I checked, and using a factor of two instead of four only makes a relatively small difference.

Looks like this change has a higher impact than I would have expected. The code size increase //might// be worthwhile, but it's not possible to say without performance data.

It's possible that D115261 <https://reviews.llvm.org/D115261> would mitigate this somewhat, by avoiding at least some of the known non-profitable unrolling.

Ultimately this (once again) comes back to our lack of a runtime unrolling cost model :/

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D140698/new/

https://reviews.llvm.org/D140698