[PATCH] D82998: [BasicAA] Enable -basic-aa-recphi by default

Dave Green via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Thu Jul 9 01:25:45 PDT 2020


dmgreen added a comment.

Thanks for the info. That's very useful.

I think I need to find a way to get better compile time numbers. Whenever I try to run them they come up so noisy as to be almost useless.  I am in the very lucky position where all of the benchmarks I run are bare-metal and deterministic, so I can very accurately get performance numbers with zero noise in them (at least of the "I run the same binary twice and get the same result" kind). Those benchmarks showed this to be a improvement, and codesize test we run at -Oz where essentially flat with some small ups and downs.

I had run the llvm-test-suite for testing, but not for performance. That was what pointed at the problems in D82987 <https://reviews.llvm.org/D82987>. Luckily there was a gcc torture test that pointed out concisely what the problem was without having to go find out which part of cjpeg was broken. With that fixed the rest of the suite passed, along with a bootstrap. I also ran some csmith testing overnight some times back with a this and a couple of other patches in it, but that didn't show any problems even when D82987 <https://reviews.llvm.org/D82987> wasn't fixed, so probably doesn't mean much.

I can try and get some statistics. I usually trust the "zero noise" benchmarks more, due to showing performance more directly than statistics or noisy results usually do. But I acknowledge that they may be different to the test-suite. Plus I guess for basic-aa checking the number of noalias vs mayalias calls can be a more direct indication of how AA on it's own is performing.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82998/new/

https://reviews.llvm.org/D82998





More information about the llvm-commits mailing list