On Wed, 7 Dec 2022 20:25:53 GMT, Xin Liu <x...@openjdk.org> wrote: >> It's probably just a statistical fluke involving a very large return value, >> but to know for sure I'd need a way to gather statistics about the return >> values and how they correlate with running-time outliers. > >> It's probably just a statistical fluke involving a very large return value, >> but to know for sure I'd need a way to gather statistics about the return >> values and how they correlate with running-time outliers. > > Yeah, it's likely that p100 data are outliers. I think average and p50 data > make more sense. > I am not a reviewer and lack domain-specific knowledge in the algorithm. I > check your code and it seems reasonable. We need reviewers to look into it.
@navyxliu p50 isn't meaningful, because this change doesn't make a difference in the ~98% of cases where only the lookup table is used. It only affects the tails. ------------- PR: https://git.openjdk.org/jdk/pull/8131
