rust-cpython: add macro for sharing references Following an experiment done by Georges Racinet, we now have a working way of sharing references between Python and Rust. This is needed in many points of the codebase, for example every time we need to expose an iterator to a Rust-backed Python class. In a few words, references are (unsafely) marked as `'static` and coupled with manual reference counting; we are doing manual borrow-checking. This changes introduces two declarative macro to help reduce boilerplate. While it is better than not using macros, they are not perfect. They need to: - Integrate with the garbage collector for container types (not needed as of yet), as stated in the docstring - Allow for leaking multiple attributes at the same time - Inject the `py_shared_state` data attribute in `py_class`-generated structs - Automatically namespace the functions and attributes they generate For at least the last two points, we will need to write a procedural macro instead of a declarative one. While this reference-sharing mechanism is being ironed out I thought it best not to implement it yet. Lastly, and implementation detail renders our Rust-backed Python iterators too strict to be proper drop-in replacements, as will be illustrated in a future patch: if the data structure referenced by a non-depleted iterator is mutated, an `AlreadyBorrowed` exception is raised, whereas Python would allow it, only to raise a `RuntimeError` if `next` is called on said iterator. This will have to be addressed at some point. Differential Revision: https://phab.mercurial-scm.org/D6631

rust-docstrings: add missing module docstrings Differential Revision: https://phab.mercurial-scm.org/D6630

rust-dirstate: improve API of `DirsMultiset` - Use opaque `Iterator` type instead of implementation-specific one from `HashMap` - Make `DirsMultiset` behave like a set both in Rust and from Python Differential Revision: https://phab.mercurial-scm.org/D6690

rust-dirstate: use EntryState enum instead of literals This improves code readability quite a bit, while also adding a layer of safety because we're checking the state byte against the enum. Differential Revision: https://phab.mercurial-scm.org/D6629

rust-parsers: switch to parse/pack_dirstate to mutate-on-loop Both `parse_dirstate` and `pack_dirstate` can operate directly on the data they're passed, which prevents the creation of intermediate data structures, simplifies the function signatures and reduces boilerplate. They are exposed directly to the Python for now, but a later patch will make use of them inside `hg-core`. Differential Revision: https://phab.mercurial-scm.org/D6628

rust-parsers: move parser bindings to their own file and Python module This tidies up the Rust side while simplifying the Python side. Differential Revision: https://phab.mercurial-scm.org/D6627

rust-dirstate: create dirstate submodule in hg-cpython This module will soon hold multiple files, this change is to make the review process easier. Differential Revision: https://phab.mercurial-scm.org/D6626

rust-discovery: using from Python code As previously done in other topics, the Rust version is used if it's been built. The version fully in Rust of the partialdiscovery class has the performance advantage over the Python version (actually using the Rust MissingAncestor) if the undecided set is big enough. Otherwise no sampling occurs, and the discovery is reasonably fast anyway. Note: it's hard to predict the size of the initial undecided set, it can depend on the kind of topological changes between the local and remote graphs. The point of the Rust version is to make the bad cases acceptable. More specifically, the performance advantages are: - faster sampling, especially takefullsample() - much faster addmissings() in almost all cases (see commit message in grandparent of the present changeset) - no conversion cost of the undecided set at the interface between Rust and Python == Measurements with big undecided sets For an extreme example, discovery between mozilla-try and mozilla-unified (over one million undecided revisions, same case as in dbd0fcca6dfc), we get roughly a x2.5/x3 better performance: Growing sample size (5% starting with 200): time goes down from 210 to 72 seconds. Constant sample size of 200: time down from 1853 to 659 seconds. With a sample size computed from number of roots and heads of the undecided set (`respectsize` is `False`), here are perfdiscovery results: Before ! wall 9.358729 comb 9.360000 user 9.310000 sys 0.050000 (median of 50) After ! wall 3.793819 comb 3.790000 user 3.750000 sys 0.040000 (median of 50) In that later case, the sample sizes are routinely in the hundreds of thousands of revisions. While still faster, the Rust iteration in addmissings has less of an advantage than with smaller sample sizes, but one sees addcommons becoming faster, probably a consequence of not having to copy big sets back and forth. This example is not a goal in itself, but it showcases several different areas in which the process can become slow, due to different factors, and how this full Rust version can help. == Measurements with small undecided sets In cases the undecided set is small enough than no sampling occurs, the Rust version has a disadvantage at init if `targetheads` is really big (some time is lost in the translation to Rust data structures), and that is compensated by the faster `addmissings()`. On a private repository with over one million commits, we still get a minor improvement, of 6.8%: Before ! wall 0.593585 comb 0.590000 user 0.550000 sys 0.040000 (median of 50) After ! wall 0.553035 comb 0.550000 user 0.520000 sys 0.030000 (median of 50) What's interesting in that case is the first addinfo() at 180ms for Rust and 233ms for Python+C, mostly due to add_missings and the children cache computation being done in less than 0.2ms on the Rust side vs over 40ms on the Python side. The worst case we have on hand is with mozilla-try, prepared with discovery-helper.sh for 10 heads and depth 10, time goes up 2.2% on the median. In this case `targetheads` is really huge with 165842 server heads. Before ! wall 0.823884 comb 0.810000 user 0.790000 sys 0.020000 (median of 50) After ! wall 0.842607 comb 0.840000 user 0.800000 sys 0.040000 (median of 50) If that would be considered a problem, more adjustments can be made, which are prematurate at this stage: cooking special variants of methods of the inner MissingAncestors object, retrieving local heads directly from Rust to avoid the cost of conversion. Effort would probably be better spent at this point improving the surroundings if needed. Here's another data point with a smaller repository, pypy, where performance is almost identical Before ! wall 0.015121 comb 0.030000 user 0.020000 sys 0.010000 (median of 186) After ! wall 0.015009 comb 0.010000 user 0.010000 sys 0.000000 (median of 184) Differential Revision: https://phab.mercurial-scm.org/D6430

rust-discovery: optimization of add commons/missings for empty arguments These two cases have to be catched early for different reasons. In the case of add_missing_revisions, we don't want to trigger the computation of the undecided set (and the children cache) too early: the later the better. In the case of add_common_revisions, the inner `MissingAncestors` object wouldn't know that all ancestors of its bases have already been removed from the undecided. In principle, that would in itself be a lead for further improvement: this remove_ancestors_from could be more incremental, but the current performance seems to be good enough. Differential Revision: https://phab.mercurial-scm.org/D6429

rust-discovery: using the children cache in add_missing The DAG range computation often needs to get back to very old revisions, and turns out to be disproportionately long, given that the end goal is to remove the descendents of the given missing revisons from the undecided set. The fast iteration capabilities available in the Rust case make it possible to avoid the DAG range entirely, at the cost of precomputing the children cache, and to simply iterate on children of the given missing revisions. This is a case where staying on the same side of the interface between the two languages has clear benefits. On discoveries with initial undecided sets small enough to bypass sampling entirely, the total cost of computing the children cache and the subsequent iteration becomes better than the Python + C counterpart, which relies on reachableroots2. For example, on a repo with more than one million revisions with an initial undecided set of 11 elements, we get these figures: Rust version with simple iteration addcommons: 57.287us first undecided computation: 184.278334ms first children cache computation: 131.056us addmissings iteration: 42.766us first addinfo total: 185.24 ms Python + C version first addcommons: 0.29 ms addcommons 0.21 ms first undecided computation 191.35 ms addmissings 45.75 ms first addinfo total: 237.77 ms On discoveries with large undecided sets, the initial price paid makes the first addinfo slower than the Python + C version, but that's more than compensated by the gain in sampling and subsequent iterations. Here's an extreme example with an undecided set of a million revisions: Rust version: first undecided computation: 293.842629ms first children cache computation: 407.911297ms addmissings iteration: 34.312869ms first addinfo total: 776.02 ms taking initial sample query 2: sampling time: 1318.38 ms query 2; still undecided: 1005013, sample size is: 200 addmissings: 143.062us Python + C version: first undecided computation 298.13 ms addmissings 80.13 ms first addinfo total: 399.62 ms taking initial sample query 2: sampling time: 3957.23 ms query 2; still undecided: 1005013, sample size is: 200 addmissings 52.88 ms Differential Revision: https://phab.mercurial-scm.org/D6428