mcimadamore commented on PR #912: URL: https://github.com/apache/lucene/pull/912#issuecomment-1324947263
@sherman I believe there is a big difference in performance between Java 19 and Java 17. Java 19 is the first release of the API where loop optimizations and range check elimination are correctly implemented for loops over `long` induction variable. Up to Java 19, there used to be some black magic in the FFM API implementation to use a "fast path" when a segment was detected to have a size that would fit an integer. But this resulted in multiple code paths, and, sometimes, suboptimal optimizations. On top of that, IIRC, Java 17 is missing all the static `copy` methods (these were added in Java 18). These have been added precisely to address some performance concerns raised by the Lucene team (as copying data from heap arrays to off-heap memory would require, using the Java 17 API, creating intermediate memory segment instances). Finally, Java 18 also add instance dereference methods on the MemorySegment class. E.g. no more static methods on `MemoryAccess` (that class is now gone). Again, this provides some improvements, as C2 is better equipped to speculate on the receiver type of a memory segment. @uschindler can correct me if I'm wrong, but I believe that the numbers he got against Java 17 were not too great. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
