Note: This patch also added the rayon crate as a Cargo dependency. It will
help us immensely in making Rust code parallel and easy to maintain. It is
a stable, well-known, and supported crate maintained by people on the Rust
The current dirstate.status method has grown over the years through bug
reports and new features to the point where it got too big and too complex.
This series does not yet improve the logic, but adds a Rust fast-path to speed
up certain cases.
Tested on mozilla-try-2019-02-18 with zstd compression:
- hg diff on an empty working copy:
- c: 1.64(+-)0.04s
- rust+c before this change: 2.84(+-)0.1s
- rust+c: 849(+-)40ms
- hg commit when creating a file:
- c: 5.960s
- rust+c before this change: 5.828s
- rust+c: 4.668s
- hg commit when updating a file:
- c: 4.866s
- rust+c before this change: 4.371s
- rust+c: 3.855s
- hg status -mard
- c: 1.82(+-)0.04s
- rust+c before this change: 2.64(+-)0.1s
- rust+c: 896(+-)30ms
The numbers are clear: the current Rust dirstatemap implementation is super
slow, its performance needs to be addressed.
This will be done in a future series, immediately after this one, with the goal
of getting Rust to be at least to the speed of the Python + C implementation
in all cases before the 5.2 freeze. At worse, we gate dirstatemap to only be used
in those cases.
Cases where the fast-path is not executed:
- for commands that need ignore support (status, for example)
- if subrepos are found (should not be hard to add, but winter is coming)
- any other matcher than an alwaysmatcher, like patterns, etc.
- with extensions like sparse and fsmonitor
The next step after this is to rethink the logic to be closer to
Jane Street's Valentin Gatien-Baron's Rust fast-path which does a lot less
work when possible.