On Tue, 15 Oct 2024 13:33:16 GMT, Martin Doerr <mdo...@openjdk.org> wrote:
>>> This version uses little endian accesses on big endian machines. This may >>> defeat the purpose of >>> [JDK-8339531](https://bugs.openjdk.org/browse/JDK-8339531) for those >>> platforms. Accesses are typically slower when using them with endianness >>> mismatch. But it looks at least correct. Thanks for providing a fix. >> >> Are you able to run the `SegmentBulkMismatch` benchmark on the various BE >> platforms? This will provide us with clear figures if there is an >> improvement or not. (The benchmark runs both the old and new code path) >> >> When we have the figures, we could (if necessary) open up a new issue >> targeting improved performance for BE platforms. I could help in such an >> effort but I can't drive it because I don't have access to a BE platform. > >> Are you able to run the `SegmentBulkMismatch` benchmark on the various BE >> platforms? > > Thanks for the hint. I've run it with the old implementation before > [JDK-8339531](https://bugs.openjdk.org/browse/JDK-8339531) but with the new > micro benchmark and the current implementation with your fix on Power8 Big > Endian. Some benchmarks are faster now, some a little slower, but I couldn't > see any significant regression. So, I guess we can live with it. Thanks, @TheRealMDoerr and @offamitkumar for helping me out with this one! ------------- PR Comment: https://git.openjdk.org/jdk/pull/21258#issuecomment-2414003523
