procutil: don't assign stdin to None, use os.devnull instead It will be painful to take care of procutil.stdin being None everywhere. Thanks to Yuya who recommended it.

dispatch: move IOError handling and flushing of streams to `dispatch()` Instead of patching both dispatch code and commandserver code, we directly handle this in `dispatch.dispatch()`. Thanks to Yuya who recommended this.

simplemerge: write output only once it's complete `simplemerge()` can write either to `ui.fout` or to the file context (for in-memory merge). This patch simplifies the code a bit by making it build the output the same way regardless of where it's written, and then writes the whole output at once. I don't think it will be a problem that we don't output anything until the whole file is merged even if the file is large. Differential Revision: https://phab.mercurial-scm.org/D9550

simplemerge: avoid quadratic concatenation when building output text I haven't checked if the difference is measurable, but the new version is no less readable or idiomatic, so I don't think performance numbers are needed. Differential Revision: https://phab.mercurial-scm.org/D9549

simplemerge: work with opts as native strings instead of bytes There was little reason to use `pycompat.byteskwargs()` in `simplemerge()` as far as I could tell. Differential Revision: https://phab.mercurial-scm.org/D9548

hghave: update the check for virtualenv This started as `hghave --test-features` failing on Windows in `test-hghave.t`. IDK how this worked, as neither my Linux nor Windows machines have the old attribute with virtualenv 20.2.2, even on py2. I think this was noticed recently because 357d8415aa27 mentioned an AttributeError, and mitigated by making this py2 only. But as mentioned, this is also a problem on py2 (where the failure was observed). When I got this working by removing the attribute reference, the command in the test failed because the `--no-site-package` argument was removed some time ago. Therefore, this backs out 357d8415aa27 and references a known good attribute (which was done to suppress the warning about an unused import) that also ensures the command does not need the argument. Since there appears to be (minor) broken stuff on py3, manually apply the `no-py3` guard that was backed out of the check itself. Differential Revision: https://phab.mercurial-scm.org/D9547

singlehead: introduce option to restrict to public changes The new experimental.single-head-per-branch:public-changes-only option restricts the single-head-per-branch filter to public changesets. This is useful when serving one repository with different views as publishing and non-publishing repository. Differential Revision: https://phab.mercurial-scm.org/D9525

treemanifest: stop storing full path for each item in manifest._lazydirs This information is obtainable, if needed, based on the lazydirs key (which is the entry name) and the manifest's `dir()` method. ### Performance This is actually both a memory and a performance improvement, but it's likely to be a very small one in most situations. In the pathological repo I've been using for testing other performance work I've done recently, this reduced the time for a rebase operation (rebasing two commits across a public-phase change that touches a sibling of one of my tracked directories where the common parent is massive (>>10k entries)): #### Before ``` Time (mean ± σ): 4.059 s ± 0.121 s [User: 0.9 ms, System: 0.6 ms] Range (min … max): 3.941 s … 4.352 s 10 runs ``` #### After ``` Time (mean ± σ): 3.707 s ± 0.060 s [User: 0.8 ms, System: 0.8 ms] Range (min … max): 3.648 s … 3.818 s 10 runs ``` Differential Revision: https://phab.mercurial-scm.org/D9553

extensions: avoid including `__index__` in the disabled extension list This generated module contains a dictionary of all bundled extension names and their help for builds that cannot enumerate extensions in the filesystem. The disabled list gets displayed in `hg help extensions`, and is also used by `setup.py` to populate `__index__.py` when building. I haven't seen it sneak into either py2exe or PyOxidizer builds, but it does show up when running tests locally after having created an installer. Differential Revision: https://phab.mercurial-scm.org/D9544

windows: continue looking at `%HOME%` for user config files with py3.8+ The `%HOME%` variable is explicitly called out in `hg help config` as a location that is consulted when reading user files, but python stopped looking at it when expanding '~' in py3.8+.[1] Restore that old functionality by copying in the old implementation (and simplifying it to just use bytes). It could be simplfied further, since only '~' is passed, but I'm not sure yet if we need to make this a generic utility function on Windows. There are other uses of `os.path.expanduser()`, but this is the only case I know of that documents `%HOME%` usage. (The reason for removing it was that it typically isn't set, but it actually is set in MSYS and PowerShell, and `%HOME%` and `%USERPROFILE%` are different in MSYS. I could be convinced to just replace all uses with this as a general utility, so we don't have to think too hard about BC.) [1] https://bugs.python.org/issue36264 Differential Revision: https://phab.mercurial-scm.org/D9559