copies: filter out copies from non-existent source later in _chain() _changesetforwardcopies() repeatedly calls _chain(). That is very expensive because _chain() does lookups in the manifest. I hope to split up the function in two parts: 1) simple chaining, not considering end points, and 2) filter out files that don't exist in the end points (and ping-pong copies/renames). This patches gets us closer to that by moving the check for non-existent source later in the function. Now there are no more checks for "src" and "dst" in the first loop; all the filtering of invalid copies is done in the second loop. The code also looks much more consistent now. No measureable impact on `hg debugpathcopies 4.0 4.8`. That shouldn't be surprising since the only case we're doing more checks now is in case of chained copies/renames, which are quire rare in practice. Differential Revision: https://phab.mercurial-scm.org/D6277

copies: clarify mutually exclusive cases in _chain() with a s/if/elif/ If the 'b' dict has a rename from 'x' to 'y', it shouldn't be possible for 'x' to be both (a key) in 'a' and in 'src'. That would mean that 'x' is a file in the source commit and also a rename destination in the intermediate commit. But we currently don't allow renaming files onto existing files, so that shouldn't happen. So let's clarify that by using an "elif" instead of an "if". And if we did allow renaming files onto existing files, we should prefer to use the rename destination in the intermediate commit as source anyway. Differential Revision: https://phab.mercurial-scm.org/D6276

copies: delete a redundant cleanup step in _chain() The check is redundant since d5edb5d3a337 (copies: filter out copies when target is not in destination manifest, 2019-02-14). To test that hypothesis, I made this change in the commit that commit, but all tests still passed. I think the case was necessary before then, we just didn't have tests for it. Differential Revision: https://phab.mercurial-scm.org/D6275

copies: document cases in _chain() Differential Revision: https://phab.mercurial-scm.org/D6274

copies: ignore heuristics copytracing when using changeset-centric algos Differential Revision: https://phab.mercurial-scm.org/D6269

copies: move check for experimental.copytrace==<falsy> earlier I'm going to ignore experimental.copytrace when changeset-centric algorithms are required. This little refactoring makes that easier to add. Differential Revision: https://phab.mercurial-scm.org/D6268

copies: replace .items() by .values() where appropriate As pointed out by Pierre-Yves. Differential Revision: https://phab.mercurial-scm.org/D6266

copies: inline _computenonoverlap() in mergecopies() We now call pathcopies() from the base to each of the commits, and that calls _computeforwardmissing(), which does file prefetching (in the remotefilelog override). So the call to _computenonoverlap() is now pointless (the sets of files from _computenonoverlap() are subsets of the sets of files from _computeforwardmissing()). This somehow also fixes a broken remotefilelog test. Differential Revision: https://phab.mercurial-scm.org/D6256

copies: calculate mergecopies() based on pathcopies() When copies are stored in changesets, we need a changeset-centric version of mergecopies() just like we have a changeset-centric version of pathcopies(). I think the natural way of thinking about mergecopies() is in terms of pathcopies() from the base to each of the commits. So if we can rewrite mergecopies() based on two such pathcopies() calls, we'll get the changeset-centric version for free. That's what this patch does. A nice bonus is that it ends up being a lot simpler. mergecopies() has accumulated a lot of technical debt over time. One good example is the code for dealing with grafts (the "partial/incomplete/dirty" stuff). Since pathcopies() already deals with backwards renames and ping-pong renames, we get that for free. I've run tests with hard-coded debug logging for "fullcopy" and while I haven't looked at every difference it produces, all the ones I have looked at seemed reasonable to me. I'm a little surprised that no more tests fail when run with '--extra-config-opt experimental.copies.read-from=compatibility' compared to before this patch. This patch also fixes the broken cases in test-annotate.t and test-fastannotate.t. It also enables the part of test-copies.t that was previously disabled exactly because mergecopies() needed to get a changeset-centric version. One drawback of the rewritten code is that we may now make remotefilelog prefetch more files. We used to prefetch files that were unique to either side of the merge compared to the other. We now prefetch files that are unique to either side of the merge compared to the base. This means that if you added the same file to each side, we would not prefetch it before, but we would now. Such cases are probably quite rare, but one likely scenario where they happen is when moving from a commit to its successor (or the other way around). The user will probably already have the files in the cache in such cases, so it's probably not a big deal. Some timings for calculating mergecopies between two revisions (revisions shown on each line, all using the common ancestor as base): In the hg repo: 4.8 4.9: 0.21s -> 0.21s 4.0 4.8: 0.35s -> 0.63s In and old copy of the mozilla-unified repo: FIREFOX_BETA_60_BASE^ FIREFOX_BETA_60_BASE: 0.82s -> 0.82s FIREFOX_NIGHTLY_59_END FIREFOX_BETA_60_BASE: 2.5s -> 2.6s FIREFOX_BETA_59_END FIREFOX_BETA_60_BASE: 3.9s -> 4.1s FIREFOX_AURORA_50_BASE FIREFOX_BETA_60_BASE: 31s -> 33s So it's measurably slower in most cases. The most significant difference is in the hg repo between revisions 4.0 and 4.8. In that case it seems to come from the fact that pathcopies() uses fctx.isintroducedafter() (in _tracefile), while the old mergecopies() used fctx.linkrev() (in _checkcopies()). That results in a single call to filectx._adjustlinkrev(), which is responsible for the entire difference in time (in my repo). So we pay a performance penalty but we get more correct code (see change in test-mv-cp-st-diff.t). Deleting the "== f.filenode()" in _tracefile() recovers the lost performance in the hg repo. There were are few other optimizations in _checkcopies() that I could not measure any impact from. One was from the "seen" set. Another was from a "continue" when the file was not in the destination manifest (corresponding to "am" in _tracefile). Also note that merge copies are not calculated when updating with a clean working copy, which is probably the most common case. I therefore think the much simpler code is worth the slowdown. Differential Revision: https://phab.mercurial-scm.org/D6255

tests: add test where copy source is deleted and added back This shows another difference between pathcopies() and mergecopies(): mergecopies() considers files that have been deleted and then added back as different files, but pathcopies() does not. Differential Revision: https://phab.mercurial-scm.org/D6330