On Tue, 13 Jun 2023 20:42:49 GMT, Man Cao <m...@openjdk.org> wrote: >> Hi all, >> >> Could anyone review this small fix for a data race in >> java.io.ClassCache$CacheRef? This fix makes the code safer by making the >> code data-race-free. > > Man Cao has updated the pull request incrementally with one additional commit > since the last revision: > > Switch to using volatile instead of AtomicReference
Unfortunately, even the single `volatile` on that fast path seems to cost quite a bit :( On M1: # CONF=macosx-aarch64-server-release make images test TEST=micro:java.io.ObjectStreamClasses Benchmark Mode Cnt Score Error Units # Before ObjectStreamClasses.testLookup avgt 8 106,946 ± 0,691 ns/op # After ObjectStreamClasses.testLookup avgt 8 119,984 ± 0,588 ns/op So again, this looks like a benign data race, and we are penalizing the serialization a bit even with the lightweight fix. This is only to make TSAN happy, right? If so, I would rather keep the code as is. There are plenty of benign data races in JDK, we should be able to somehow tell TSAN which are actually fine. ------------- PR Comment: https://git.openjdk.org/jdk/pull/14386#issuecomment-1590869101
