log: make -frREV PATH detect missing files before falling back to slow path If -rREV isn't specified, "log --follow" would abort on nonexistent paths. Let's implement this behavior for "-frREV" case as we have ctx.hasdir() now. Otherwise "log -frREV PATH" would silently fall back to slow path and files wouldn't be followed across renames. The loop is quadratic (as before), but the size of the startctxs and match.files() should be small in general. Some tests are marked as BROKEN since file renames aren't tracked in the slow path. This is a known limitation of the current history traversal function.

log: fix -fr'wdir()' PATH to follow newly added file Testing filelog doesn't make sense in this case because the file existence is tested against the specified changectxs. If the filelog is empty and if startctxs != [wctx], 'f not in c' should be triggered.

log: reorganize if-else and for loop in logcmdutil._makematcher() The test conditions are branchy depending on --follow and --rev options, so it should be better to switch first by --follow --rev. Note that revs is not empty so "if follow and startctxs" can be replaced with "if follow and opts.get(b'rev')".

merge: update commitinfo from all mergeresults during bid merge During bid merge, it's not clear which commitinfo should be stored and which one should not. This depends on which side the bid merge chooses for a file. For this we will need to refactor bid merge code and commitinfo handling. For now, we just blindly updates info since we hardly have any users of commitinfo and this will help us in testing and clearing out further path. Differential Revision: https://phab.mercurial-scm.org/D8965

merge: add missing ACTION_KEEP when both remote and ancestor are not present Previous patch may lead to confusion that the related criss-cross merge is consistent when done from any of the parents. However this is not true and we were missing setting an ACTION_KEEP. This patch now exposes that bid-merge favors ACTION_KEEP always and the result of merge is different when started from different parents. This change also effects a test case above where bid merge was wrongly picking `r` because it was missing keep related information from one of the ancestor. After this test, we are back in a state in the criss-cross merge tests where the result depends on which parent we are merging from. Differential Revision: https://phab.mercurial-scm.org/D8941

merge: store ACTION_KEEP_ABSENT when we are keeping the file absent locally If a file is not present on the local side, and it's unchanged between other merge parent and ancestor, we don't use any action, neither we had a if-else branch for that condition. This leads to bid-merge missing that there is a such action possible which can be performed. As test changes demonstrate, we now choose the locally deleted side instead of choosing the remote one consistently. This is also wrong behavior which is resulted because of missing possible action. It will be fixed in next patch. This whole logic is not acurrate as we should prompt user on what to do when this kind of criss-cross merge is in play. Differential Revision: https://phab.mercurial-scm.org/D8940

merge: add `ACTION_KEEP_ABSENT` to represent files we want to keep absent There are files which were deleted/not present in working copy parent but were present on other side of merge. On merge, we might decide to keep them deleted. We want to track such cases more closely, rather all kind of cases which results from some kind of merging logic. We do have `ACTION_KEEP` but having a dedicated action for the absent case is more cleaner. Initially I named the action as `ACTION_KEEP_DELETED` but later realized that file can be not-present because of other reasons than deletion like rename, hence decided to use more generic name `ACTION_KEEP_ABSENT`. Differential Revision: https://phab.mercurial-scm.org/D8974

mergeresult: introduce dedicated tuple for no-op actions This will help us in adding more no-op actions in next patch while keeping the code cleaner.

tests: add criss cross merging tests whose behavior need to be fixed Merging two changesets can mark a file as removed post merge. However, in some cases, a user might not want to remove that file and they revert the removal back and commit the merge. All this works perfectly well. However, when we do criss-cross merges with such merge where user explicitly choose to revert the removal with one where another user choose the removal, we does not get any conflict. The intent here is conflicting and merge should result in conflicts. One user merged and want to keep the file while other user merged and want to remove the file. Merging those merges should result in conflicts. This patch adds test cases for these cases. Differential Revision: https://phab.mercurial-scm.org/D8939

repo: avoid copying/updating a dict on every `repo.__getitem__` This has some mild performance benefits. I'm looking into a pathological case where one of our `hg log` invocations takes several seconds, and according to hyperfine this reduces the wall time of the entire operation (running in chg) from: ``` Time (mean ± σ): 7.390 s ± 0.106 s [User: 7.058 s, System: 0.271 s] Range (min … max): 7.300 s … 7.625 s ``` to: ``` Time (mean ± σ): 7.046 s ± 0.091 s [User: 6.714 s, System: 0.279 s] Range (min … max): 6.916 s … 7.169 s ``` Note: the log command is slow due to an issue in our custom stuff executing `repo[<arg>]` 298,800 times. This performance improvement is likely not noticeable during normal operation, but I don't feel like it's making the code more difficult to understand, and every small bit helps. Differential Revision: https://phab.mercurial-scm.org/D9022