This commit represents the initial work to get hyperlinks working and
was submitted as part of PR #2483. Subsequent commits largely retain the
functionality and structure of the hyperlink support added here, but
rejigger some things around.
This represents yet another iteration on how `ignore` enqueues and
distributes work in parallel. The original implementation used a
multi-producer/multi-consumer thread safe queue from crossbeam. At some
point, I migrated to a simple `Arc<Mutex<Vec<_>>>` and treated it as a
stack so that we did depth first traversal. This helped with memory
usage in very wide directories.
But it turns out that a naive stack-behind-a-mutex can be quite a bit
slower than something that's a little smarter, such as a work-stealing
stack used in this commit. My hypothesis for why this helps is that
without the stealing component, work distribution can get stuck in
sub-optimal configurations that depend on which directory entries get
assigned to a particular worker. It's likely that this can result in
some workers getting "more" work than others, just by chance, and thus
remain idle. But the work-stealing approach heads that off.
This does re-introduce a dependency on parts of crossbeam which is kind
of a bummer, but it's carrying its weight for now.
Closes#1823, Closes#2591
Ref https://github.com/sharkdp/fd/issues/28
This brings in aarch64 SIMD support for Teddy[1]. In effect, it means
searches that are multiple (but a small number of) literals extracted
will likely get much faster on aarch64 (i.e., Apple silicon). For
example, from the PR, on my M2 mac mini:
$ time rg-before-teddy-aarch64 -i -c 'Sherlock Holmes' OpenSubtitles2018.half.en
3055
real 8.196
user 7.726
sys 0.469
maxmem 5728 MB
faults 17
$ time rg-after-teddy-aarch64 -i -c 'Sherlock Holmes' OpenSubtitles2018.half.en
3055
real 1.127
user 0.701
sys 0.425
maxmem 4880 MB
faults 13
w00t.
[1]: https://github.com/BurntSushi/aho-corasick/pull/129
As of the memchr 2.6 release, its Iterator::count method is specialized
to only count the number of occurrences instead of finding the offset of
each occurrence. This replaces ripgrep's use of the bytecount crate.
While micro-benchmarks suggest that memchr's method has better
throughput than bytecount, it turned out to be an illusion. Namely, on a
~13GB haystack prior to this change:
$ time rg-bytecount 'You killed my friend, my best friend, my lifelong friend!' OpenSubtitles2018.raw.en --line-number
441450441:- You killed my friend, my best friend, my lifelong friend!
real 1.473
user 1.186
sys 0.286
maxmem 12512 MB
faults 0
And then after:
$ time rg 'You killed my friend, my best friend, my lifelong friend!' OpenSubtitles2018.raw.en --line-number
441450441:- You killed my friend, my best friend, my lifelong friend!
real 1.532
user 1.280
sys 0.250
maxmem 12512 MB
faults 0
But perf is just about in the same ballpark. That's good enough for me
at the moment in order to drop the extra dependency.
I did this because the marginal cost of adding the Iterator::count()
specialization to memchr was extremely small.
This increases the limits a bit for when the regex engine will build and
use a fully compiled DFA. They can faster in some circumstances. For
example, '(?-u)^\w{30,}$' gets a nice speed boost from state
acceleration.
We are also able to remove `regex` proper as a dependency. Wow.
Previously, ripgrep core was responsible for escaping regex patterns and
implementing the --line-regexp flag. This commit moves that
responsibility down into the matchers such that ripgrep just needs to
hand the patterns it gets off to the matcher builder. The builder will
then take care of escaping and all that.
This was done to make pattern construction completely owned by the
matcher builders. With the arrival regex-automata, this means we can
move to the HIR very quickly and then never move back to the concrete
syntax. We can then build our regex directly from the HIR. This overall
can save quite a bit of time, especially when searching for large
dictionaries.
We still aren't quite as fast as GNU grep when searching something on
the scale of /usr/share/dict/words, but we are basically within spitting
distance. Prior to this, we were about an order of magnitude slower.
This architecture in particular lets us write a pretty simple fast path
that avoids AST parsing and HIR translation entirely: the case where one
is just searching for a literal. In that case, we can hand construct the
HIR directly.
0.2.4 updates to PCRE2 10.42 and has a few other nice changes. For
example, when `utf` is enabled, the crate will always set the
PCRE2_MATCH_INVALID_UTF option. That means we no longer need to do
transcoding or UTF-8 validity checks.
Because of this, we actually get to remove one of the two uses of
`unsafe` in ripgrep's `main` program.
(This also updates a couple other dependencies for convenience.)
Just some small polishing. We also get rid of thread_local in favor of
using regex-automata, mostly just in the name of reducing dependencies.
(We should eventually be able to drop thread_local completely.)
This leaves the grep-regex crate in tatters. Pretty much the entire
thing needs to be re-worked. The upshot is that it should result in some
big simplifications. I hope.
The idea here is to drop down and actually use regex-automata 0.3
instead of the regex crate itself.
Note that this adds a new dependency, 'unicode-ident', and removes
'unicode-xid'. I looked briefly at 'unicode-ident' and all looks okay.
It is also permissively licensed.
