rust-index: remove one collect when converting back Turns out this is slightly faster. Sending the results back to Python is still the most costly (like 75% of the time) of the whole method, but it's about as fast as it can be now. hg perf::phases on mozilla-try-2023-03-22 before: 0.267114 after: 0.247101

rust-index: improve phase computation speed While less memory efficient, using an array is *much* faster than using a HashMap, especially with the default hasher. It even makes the code simpler, so I'm not really sure what I was thinking in the first place, maybe it's more obvious now. This fix a significant performance regression when using the rust version of the code. (however, the C code still outperform rust on this operation) hg perf::phases on mozilla-try-2023-03-22 - 6.6.3: 0.451239 seconds - before: 0.982495 seconds - after: 0.265347 seconds - C code: 0.183241 second

phases: directly update the phase sets in advanceboundary This is similar to what we do in retractboundary. There is no need to invalidate the cache if we have everything at hand to update it.

phases: large rework of advance boundary In a similar spirit as the rework of retractboundary, the new algorithm is doing an amount of work in the order of magnitude of the amount of changeset that changes phases. (except to find new roots in impacted higher phases if any may exists). This result in a very significant speedup for repository with many old draft like mozilla try. runtime of perf:unbundle for a bundle constaining a single changeset (C code): before 6.7 phase work: 14.497 seconds before this change: 6.311 seconds (-55%) with this change: 2.240 seconds (-85%) Combined with the other patches that fixes the phases computation in the Rust index, the rust code with a persistent nodemap get back to quite interresting performances with 2.026 seconds for the same operation, about 10% faster than the C code.

phases: apply similar early filtering to advanceboundary advanceboundary is called the push's unbundle (but not the other unbundle) so advanceboundary did not show up the profile I looked at so far. We start with simple pre-filtering to avoid doing any work if we don't needs too.

phases: filter revision that are already in the right phase No need to compute new roots if everything is already in order.

phases: invalidate the phases set less often on retract boundary We already have the information to update the phase set, so we do so directly instead of invalidating the cache. This show a sizeable speedup in our `perf::unbundle` benchmark on the many-draft mozilla-try repository. ### data-env-vars.name = mozilla-try-2023-03-22-zstd-sparse-revlog # benchmark.name = hg.perf.perf-unbundle # bin-env-vars.hg.flavor = no-rust # bin-env-vars.hg.py-re2-module = default # benchmark.variants.issue6528 = disabled # benchmark.variants.revs = last-10 before: 2.055259 seconds after: 1.887064 seconds (-8.18%) # benchmark.variants.revs = last-100 before: 2.409239 seconds after: 2.222429 seconds (-7.75%) # benchmark.variants.revs = last-1000 before: 3.945648 seconds after: 3.762480 seconds (-4.64%)

phases: incrementally update the phase sets when reasonable When the amount of manual walking is small, we update the phases set manually instead of computing them from scratch. This should help small update. The next changesets will make this used more often by reducing the amount of full invalidation we do on roots upgrade. The criteria for using an incremental upgrade are arbitrary, however, it "should never hurt".