indygreg added a comment.
I can reproduce the results. xxhash needs more than 4-8 bytes to outperform the old hasher. I also suspect more hash collisions are present with xxhash. Although I haven't verified. The input showing the regression effectively turns the code into a hash table benchmark. But my results on actual repo data show an improvement. FWIW, memchr+djb hash perform about the same as the original code on the perfbdiff -m case. REPOSITORY rHG Mercurial REVISION DETAIL https://phab.mercurial-scm.org/D2626 To: indygreg, #hg-reviewers, quark Cc: quark, mercurial-devel _______________________________________________ Mercurial-devel mailing list Mercurial-devel@mercurial-scm.org https://www.mercurial-scm.org/mailman/listinfo/mercurial-devel