On Wed, 21 Sep 2022 18:29:08 GMT, Aleksey Shipilev <sh...@openjdk.org> wrote:
>> @shipilev Good points and I agree with them. >> >>> Are you getting +0.8% throughput on microbenchmarks? >> >> Yes, I got it a few times. I tested on MacOS/M1 and Linux/AMD for >> 64/512/2048/16384 bytes. As the improvement is not significant, the number >> was impacted by the platforms status/loading. I can see improvement, but I >> did not get a stable +0.8%. >> >> I got a more than +1% improvement in another algorithm implementation (SM3). >> That's the reason I want to check out the JDK implementations. But you let >> me convinced that this may not be worthy of the effort. Thank you so much! >> >>> https://bugs.openjdk.org/browse/JDK-8256523 >> >> Thank you so much for pointing this out. From it, I learned something new >> that the Java inline may not work in some circumstance. Impressed PR >> description! > >> But you let me convinced that this may not be worthy of the effort. > > Let's think about it and not discount it right away. 1% improvement on > OpenJDK scale is still quite significant, if we can show the non-interpreter > modes benefit from it. Maybe older X86 and AArch64 are our targets, since it > is not guaranteed to have SHA hardware intrinsics? @shipilev Did you have further comment? Thanks! ------------- PR: https://git.openjdk.org/jdk/pull/10365