On Thu, 21 Sep 2023 02:38:08 GMT, Chen Liang <li...@openjdk.org> wrote:
>> Like #12077, this uses JDK's internal utilities to speed up ASCII reading in >> Class files, where most identifiers, from conventions to attribute names, >> are ASCII. See the JBS issue for more in-depth explanations. >> >> Before: (Master) >> >> Benchmark Mode Cnt Score Error Units >> ReadMetadata.jdkReadMemberNames thrpt 4 167.623 ± 8.522 ops/s >> >> >> After: (This patch, first revision) >> >> Benchmark Mode Cnt Score Error Units >> ReadMetadata.jdkReadMemberNames thrpt 4 175.908 ± 4.766 ops/s > > Chen Liang has updated the pull request incrementally with one additional > commit since the last revision: > > Fix logical bug with char array filling I can verify the improvement locally in your added micro: Name Cnt Base Error Test Error Unit Diff% ReadMetadata.jdkReadMemberNames 4 108.234 ± 1.089 114.953 ± 0.709 ops/s 6.2% (p = 0.000*) I don't have domain expertise w.r.t the classfile API to value how much this will benefit typical use-cases, but it seem like a straightforward enough improvement to me. (At least the `vectorizedHashCode` has the nice property that it's always faster than a naive `31 * hash + v` loop, even when inputs are too small to actually benefit from SIMD instructions). ------------- PR Comment: https://git.openjdk.org/jdk/pull/15837#issuecomment-1729454730
