On Wed, 30 Mar 2022 12:06:39 GMT, Claes Redestad <redes...@openjdk.org> wrote:
> A few integer divisions and multiplications can be replaced with test + > addition, leading to a decent speed-up on java.time microbenchmarks such as > `GetYearBench`. Numbers from my local x86 workstation, seeing similar > speed-up on aarch64 and other x86 setups. > > Baseline: > > Benchmark Mode Cnt Score > Error Units > GetYearBench.getYearFromMillisZoneOffset thrpt 15 18.492 ± > 0.017 ops/ms > GetYearBench.getYearFromMillisZoneRegion thrpt 15 6.121 ± > 0.135 ops/ms > GetYearBench.getYearFromMillisZoneRegionNormalized thrpt 15 18.936 ± > 0.012 ops/ms > GetYearBench.getYearFromMillisZoneRegionUTC thrpt 15 9.283 ± > 0.222 ops/ms > > Patched: > > Benchmark Mode Cnt Score > Error Units > GetYearBench.getYearFromMillisZoneOffset thrpt 15 20.931 ± > 0.013 ops/ms > GetYearBench.getYearFromMillisZoneRegion thrpt 15 6.858 ± > 0.167 ops/ms > GetYearBench.getYearFromMillisZoneRegionNormalized thrpt 15 20.923 ± > 0.017 ops/ms > GetYearBench.getYearFromMillisZoneRegionUTC thrpt 15 10.028 ± > 0.182 ops/ms > > > Testing: java.time tests locally, CI tier1+2 ongoing. I'm happy, based on the assertion that it produces the same result and is faster. ------------- Marked as reviewed by scolebourne (Author). PR: https://git.openjdk.java.net/jdk/pull/8039
