[PATCH] D69686: [LVI][CVP] Use block value when simplifying icmps

Nikita Popov via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Sun Aug 30 08:32:20 PDT 2020


nikic added a comment.

Numbers for the new version of the patch:

Compile-time: https://llvm-compile-time-tracker.com/compare.php?from=5ec2b757cc7d37ff0d03b36ee863b0962fe78108&to=66e77d206c673aa4ed932b95b75256455daa968b&stat=instructions
Text size: https://llvm-compile-time-tracker.com/compare.php?from=5ec2b757cc7d37ff0d03b36ee863b0962fe78108&to=66e77d206c673aa4ed932b95b75256455daa968b&stat=size-text

Compile-time ends up being fairly neutral in geomean, with some benchmarks becoming slower and some faster, in part depending on whether we generate more or less IR as a result of this.

Text size reduces by 0.1% geomean, with some fairly heft outliers (-1.9% on consumer-typeset LTO).

I think that some of the motivation for this patch has been subsumed by the range support in SCCP in the meantime, but the numbers do suggest that doing this in CVP is still useful. It would be a shame to compute all this yummy range information and then just ... not use it :)


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69686/new/

https://reviews.llvm.org/D69686



More information about the llvm-commits mailing list