add aarch64 SIMD implementations of memchr and memmem (and other goodies) #129
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
BurntSushi
changed the title
BurntSushi
force-pushed
the
ag/refactor
branch
5 times, most recently
from
cfee283
to
49c143f
Compare
This PR doesn't just add `aarch64`-specific code, but it refactors pretty much everything about how the code is organized. There are big perf wins for `aarch64` (see benchmark results below), and also latency improvements across the board. A brief summary of the changes in this PR: * I've added `aarch64` NEON vector implementations for `memchr`, `memrchr`, `memchr2`, `memrchr2`, `memchr3`, `memrchr3` and `memmem`. This should lead to massive speed improvements on an increasing popular target, due in large part to Apple silicon. * I've added `wasm32` simd128 vector implementations for `memchr`, `memrchr`, `memchr2`, `memrchr2`, `memchr3` and `memrchr3`. (alexcrichton previously contributed a vector implementation for `memmem` and that remains.) * `x86_64` has no real additions other than the `memchr_iter(needle, haystack).count()` specialization. It already has SSE2 and AVX2 implementations of `memchr` (and friends) and `memmem`. It uses AVX2 automatically via runtime inspection of what the current CPU supports. There is no need to compile with the `avx2` feature enabled. * I've replaced the benchmark suite using Criterion with a benchmark suite using [rebar](https://github.com/BurntSushi/rebar). While I designed rebar to be used for regex engines, it can be used for [any substring or multi-substring search task](https://github.com/BurntSushi/rebar/blob/45afe89f437173d2dd970fee 7d7f1db5d0e05588/BYOB.md). * I've added a new `arch` sub-module that exposes a lot of the internal routines (including target specific routines) used to implement `memchr` and `memmem`. This module is part of a major refactoring of how this crate is organized and it seemed prudent to expose the internals as their APIs are pretty straight-forward. That is, there isn't a huge API design space IMO. This module includes scalar substring search implementations of Shift-Or, Rabin-Karp and Two-Way. * As a result of the refactoring mentioned above, most of the conditional compilation stuff has been pushed down and mostly abstracted away. Moreover, since each implementation now has its own proper API surface that is uniform across other implementations, each thing can be easily independently tested. Because of this, I was able to remove a reliance on the variety of custom `cfg` knobs that the previous version of `memchr` setup in its build script. This in turn **allowed me to remove the build script entirely.** Given the ubiquity of this crate, this may lead to compile time improvements downstream. (Likely small in each individual case but perhaps large in aggregate.) I can't promise that a build script will never re-appear, but I'll try to resist adding one in the future if possible. * Despite the above, compile times for this crate have sadly seemed to increase slightly. Namely, a fresh `time rebar build -e '^rust/memchrold/memmem/prebuilt$'` reports 0.944 seconds on my system while a fresh `time rebar build -e '^rust/memchr/memmem/prebuilt$'` reports 1.164 seconds. This is on `x86_64` where no real additional code was added. This could be because of the "nicer" abstractions now present in the `arch` sub-module or perhaps how the internals are structured. (Previously there were multiple monomorphic implementations of `memchr` for example and now there is a single generic implementation that is monomorphized automatically by the compiler via generics. Perhaps that is more expensive?) * I've specialized `memchr_iter(needle, haystack).count()` to use a different vector implementation that specifically only counts matches instead of reporting the offsets of each match. This can make *huge* (potentially over an order of magnitude) differences when counting the number of matches of a frequently (even semi-frequently) occurring byte in a large haystack. This is effectively what the [`bytecount`](https://crates.io/crates/bytecount) crate does (which is what ripgrep currently uses to compute line numbers for matches), but the marginal cost of adding it to the `memchr` crate was very low. So I did. And I plan to move ripgrep to using `memchr_iter(needle, haystack).count()`. (Also, the benchmarks below suggest that the counting implementation I wrote is faster than the one in `bytecount` in some cases which look like they'll be relevant for ripgrep. This was surprising to me.) * I've added an `alloc` feature which permits compiling this crate without the standard library but with the `alloc` crate. This crate is designed through-and-through to work in a core-only context, so this doesn't unlock much compared to just disabling the `std` feature. It adds a couple of APIs requiring allocation (like `memmem::Finder::into_owned`) and other things like `arch::all::shiftor` which really want an allocation to store its bit-parallel state machine. * The `libc` feature is **DEPRECATED** and is now a no-op. I don't think there is any real benefit to it any more. * A new disabled-by-default `logging` feature has been added. When enabled, this crate will emit a smattering of log messages. Usually these messages are used to indicate what kind of strategy is selected. For example, whether a vector or scalar algorithm is used for substring search. Differences across the board from the status quo. Showing only measurements with a 1.2x (or greater) difference. ``` $ rebar diff tmp/old.csv tmp/new.csv -t 1.2 -e memmem -E oneshot benchmark engine tmp/old.csv tmp/new.csv --------- ------ ----------- ----------- memmem/code/rust-library-never-fn-strength rust/memchr/memmem/prebuilt 42.8 GB/s (1.25x) 53.6 GB/s (1.00x) memmem/code/rust-library-never-fn-strength-paren rust/memchr/memmem/prebuilt 40.8 GB/s (1.32x) 53.8 GB/s (1.00x) memmem/code/rust-library-never-fn-quux rust/memchr/memmem/prebuilt 40.5 GB/s (1.37x) 55.6 GB/s (1.00x) memmem/code/rust-library-rare-fn-from-str rust/memchr/memmem/prebuilt 39.3 GB/s (1.37x) 53.8 GB/s (1.00x) memmem/code/rust-library-common-fn-is-empty rust/memchr/memmem/prebuilt 40.5 GB/s (1.30x) 52.6 GB/s (1.00x) memmem/code/rust-library-common-fn rust/memchr/memmem/prebuilt 21.6 GB/s (1.27x) 27.5 GB/s (1.00x) memmem/pathological/rare-repeated-huge-tricky rust/memchr/memmem/prebuilt 40.9 GB/s (1.55x) 63.4 GB/s (1.00x) memmem/pathological/rare-repeated-small-match rust/memchr/memmem/prebuilt 1468.7 MB/s (1.23x) 1811.4 MB/s (1.00x) memmem/sliceslice/short rust/memchr/memmem/prebuilt 14.74ms (2.08x) 7.08ms (1.00x) memmem/sliceslice/seemingly-random rust/memchr/memmem/prebuilt 9.1 MB/s (1.23x) 11.2 MB/s (1.00x) memmem/sliceslice/i386 rust/memchr/memmem/prebuilt 41.4 MB/s (1.35x) 55.8 MB/s (1.00x) memmem/subtitles/common/huge-en-you rust/memchr/memmem/prebuilt 10.7 GB/s (1.26x) 13.5 GB/s (1.00x) memmem/subtitles/common/huge-zh-that rust/memchr/memmem/prebuilt 25.2 GB/s (1.49x) 37.5 GB/s (1.00x) memmem/subtitles/never/huge-en-john-watson rust/memchr/memmem/prebuilt 42.9 GB/s (1.48x) 63.6 GB/s (1.00x) memmem/subtitles/never/huge-en-all-common-bytes rust/memchr/memmem/prebuilt 41.9 GB/s (1.26x) 52.7 GB/s (1.00x) memmem/subtitles/never/teeny-en-all-common-bytes rust/memchr/memmem/prebuilt 1161.0 MB/s (1.53x) 1780.2 MB/s (1.00x) memmem/subtitles/never/teeny-en-some-rare-bytes rust/memchr/memmem/prebuilt 1161.0 MB/s (1.53x) 1780.2 MB/s (1.00x) memmem/subtitles/never/teeny-en-two-space rust/memchr/memmem/prebuilt 1161.0 MB/s (1.53x) 1780.2 MB/s (1.00x) memmem/subtitles/never/huge-ru-john-watson rust/memchr/memmem/prebuilt 40.6 GB/s (1.56x) 63.5 GB/s (1.00x) memmem/subtitles/never/teeny-ru-john-watson rust/memchr/memmem/prebuilt 1741.5 MB/s (1.44x) 2.4 GB/s (1.00x) memmem/subtitles/never/huge-zh-john-watson rust/memchr/memmem/prebuilt 41.1 GB/s (1.46x) 59.9 GB/s (1.00x) memmem/subtitles/never/teeny-zh-john-watson rust/memchr/memmem/prebuilt 1285.4 MB/s (1.53x) 1970.9 MB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock-holmes rust/memchr/memmem/prebuilt 41.9 GB/s (1.52x) 63.5 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock rust/memchr/memmem/prebuilt 41.9 GB/s (1.46x) 61.3 GB/s (1.00x) memmem/subtitles/rare/huge-en-medium-needle rust/memchr/memmem/prebuilt 38.3 GB/s (1.46x) 55.9 GB/s (1.00x) memmem/subtitles/rare/huge-en-long-needle rust/memchr/memmem/prebuilt 2.5 GB/s (17.34x) 44.0 GB/s (1.00x) memmem/subtitles/rare/huge-en-huge-needle rust/memchr/memmem/prebuilt 2.3 GB/s (20.24x) 45.7 GB/s (1.00x) memmem/subtitles/rare/teeny-en-sherlock-holmes rust/memchr/memmem/prebuilt 1068.1 MB/s (1.47x) 1570.8 MB/s (1.00x) memmem/subtitles/rare/teeny-en-sherlock rust/memchr/memmem/prebuilt 953.7 MB/s (1.27x) 1213.8 MB/s (1.00x) memmem/subtitles/rare/teeny-ru-sherlock-holmes rust/memchr/memmem/prebuilt 1430.5 MB/s (1.47x) 2.1 GB/s (1.00x) memmem/subtitles/rare/teeny-ru-sherlock rust/memchr/memmem/prebuilt 1213.8 MB/s (1.32x) 1602.2 MB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock-holmes rust/memchr/memmem/prebuilt 41.8 GB/s (1.33x) 55.5 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock rust/memchr/memmem/prebuilt 43.0 GB/s (1.38x) 59.4 GB/s (1.00x) memmem/subtitles/rare/teeny-zh-sherlock rust/memchr/memmem/prebuilt 895.9 MB/s (1.27x) 1137.1 MB/s (1.00x) ``` A comparison with the [`sliceslice`](https://crates.io/crates/sliceslice) crate for just substring search. We only include measurements with a 1.2x difference or greater. ``` $ rebar cmp benchmarks/record/x86_64/2023-08-26.csv -e sliceslice/memmem/prebuilt -e rust/memchr/memmem/prebuilt -t 1.2 benchmark rust/memchr/memmem/prebuilt rust/sliceslice/memmem/prebuilt --------- --------------------------- ------------------------------- memmem/byterank/binary 4.4 GB/s (1.32x) 5.8 GB/s (1.00x) memmem/code/rust-library-never-fn-strength 53.6 GB/s (1.00x) 39.8 GB/s (1.35x) memmem/code/rust-library-never-fn-strength-paren 53.8 GB/s (1.00x) 39.7 GB/s (1.35x) memmem/code/rust-library-never-fn-quux 55.6 GB/s (1.00x) 38.7 GB/s (1.44x) memmem/code/rust-library-rare-fn-from-str 53.8 GB/s (2.65x) 142.7 GB/s (1.00x) memmem/pathological/md5-huge-no-hash 50.1 GB/s (1.00x) 25.7 GB/s (1.95x) memmem/pathological/md5-huge-last-hash 47.6 GB/s (1.00x) 27.7 GB/s (1.72x) memmem/pathological/rare-repeated-huge-tricky 63.4 GB/s (1.00x) 41.9 GB/s (1.51x) memmem/pathological/rare-repeated-small-tricky 25.2 GB/s (1.32x) 33.3 GB/s (1.00x) memmem/pathological/defeat-simple-vector-alphabet 4.1 GB/s (1.65x) 6.7 GB/s (1.00x) memmem/pathological/defeat-simple-vector-freq-alphabet 19.2 GB/s (1.00x) 2.6 GB/s (7.33x) memmem/pathological/defeat-simple-vector-repeated-alphabet 1234.5 MB/s (1.00x) 508.7 MB/s (2.43x) memmem/sliceslice/short 7.08ms (1.00x) 14.10ms (1.99x) memmem/sliceslice/i386 55.8 MB/s (1.00x) 39.6 MB/s (1.41x) memmem/subtitles/never/huge-en-john-watson 63.6 GB/s (1.00x) 41.7 GB/s (1.53x) memmem/subtitles/never/huge-en-all-common-bytes 52.7 GB/s (1.00x) 42.6 GB/s (1.24x) memmem/subtitles/never/teeny-en-john-watson 1027.0 MB/s (2.17x) 2.2 GB/s (1.00x) memmem/subtitles/never/teeny-en-all-common-bytes 1780.2 MB/s (1.25x) 2.2 GB/s (1.00x) memmem/subtitles/never/teeny-en-some-rare-bytes 1780.2 MB/s (1.25x) 2.2 GB/s (1.00x) memmem/subtitles/never/teeny-en-two-space 1780.2 MB/s (1.25x) 2.2 GB/s (1.00x) memmem/subtitles/never/huge-ru-john-watson 63.5 GB/s (1.00x) 12.7 GB/s (4.99x) memmem/subtitles/never/teeny-ru-john-watson 2.4 GB/s (1.23x) 3.0 GB/s (1.00x) memmem/subtitles/never/huge-zh-john-watson 59.9 GB/s (1.00x) 41.1 GB/s (1.46x) memmem/subtitles/never/teeny-zh-john-watson 1970.9 MB/s (1.25x) 2.4 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock-holmes 63.5 GB/s (1.00x) 41.6 GB/s (1.53x) memmem/subtitles/rare/huge-en-sherlock 61.3 GB/s (1.00x) 43.0 GB/s (1.42x) memmem/subtitles/rare/huge-en-medium-needle 55.9 GB/s (1.00x) 25.7 GB/s (2.17x) memmem/subtitles/rare/huge-en-long-needle 44.0 GB/s (1.00x) 25.9 GB/s (1.70x) memmem/subtitles/rare/huge-en-huge-needle 45.7 GB/s (1.00x) 29.3 GB/s (1.56x) memmem/subtitles/rare/teeny-en-sherlock 1213.8 MB/s (1.37x) 1668.9 MB/s (1.00x) memmem/subtitles/rare/huge-ru-sherlock-holmes 40.7 GB/s (1.00x) 15.2 GB/s (2.67x) memmem/subtitles/rare/teeny-ru-sherlock 1602.2 MB/s (1.56x) 2.4 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock-holmes 55.5 GB/s (1.00x) 26.6 GB/s (2.09x) memmem/subtitles/rare/huge-zh-sherlock 59.4 GB/s (1.00x) 42.4 GB/s (1.40x) memmem/subtitles/rare/teeny-zh-sherlock-holmes 1055.9 MB/s (1.87x) 1970.9 MB/s (1.00x) memmem/subtitles/rare/teeny-zh-sherlock 1137.1 MB/s (1.86x) 2.1 GB/s (1.00x) ``` Differences with the substring search implementation and `memmem` as provided by GNU libc. Showing only measurements with 2x difference or greater. ``` $ rebar cmp benchmarks/record/x86_64/2023-08-26.csv -e libc/memmem/oneshot -e rust/memchr/memmem/oneshot -t 2 benchmark libc/memmem/oneshot rust/memchr/memmem/oneshot --------- ------------------- -------------------------- memmem/code/rust-library-never-fn-strength 11.4 GB/s (4.75x) 54.1 GB/s (1.00x) memmem/code/rust-library-never-fn-strength-paren 12.4 GB/s (4.36x) 54.0 GB/s (1.00x) memmem/code/rust-library-never-fn-quux 8.1 GB/s (6.91x) 55.8 GB/s (1.00x) memmem/code/rust-library-rare-fn-from-str 15.0 GB/s (3.59x) 53.8 GB/s (1.00x) memmem/code/rust-library-common-fn-is-empty 12.5 GB/s (4.16x) 51.9 GB/s (1.00x) memmem/code/rust-library-common-fn 2.2 GB/s (5.89x) 13.0 GB/s (1.00x) memmem/code/rust-library-common-let 3.2 GB/s (2.65x) 8.5 GB/s (1.00x) memmem/pathological/rare-repeated-huge-tricky 17.8 GB/s (3.56x) 63.3 GB/s (1.00x) memmem/pathological/rare-repeated-huge-match 718.0 MB/s (1.00x) 289.1 MB/s (2.48x) memmem/pathological/rare-repeated-small-match 707.1 MB/s (1.00x) 303.1 MB/s (2.33x) memmem/subtitles/common/huge-en-that 3.7 GB/s (4.22x) 15.7 GB/s (1.00x) memmem/subtitles/common/huge-en-one-space 1543.9 MB/s (1.00x) 541.6 MB/s (2.85x) memmem/subtitles/common/huge-ru-that 2.7 GB/s (4.22x) 11.6 GB/s (1.00x) memmem/subtitles/common/huge-ru-not 2.0 GB/s (2.47x) 5.0 GB/s (1.00x) memmem/subtitles/common/huge-ru-one-space 2.9 GB/s (1.00x) 1081.0 MB/s (2.71x) memmem/subtitles/common/huge-zh-that 4.2 GB/s (3.20x) 13.4 GB/s (1.00x) memmem/subtitles/common/huge-zh-do-not 2.6 GB/s (2.40x) 6.3 GB/s (1.00x) memmem/subtitles/common/huge-zh-one-space 5.7 GB/s (1.00x) 2.4 GB/s (2.38x) memmem/subtitles/never/huge-en-john-watson 15.4 GB/s (4.12x) 63.3 GB/s (1.00x) memmem/subtitles/never/huge-en-all-common-bytes 11.9 GB/s (4.41x) 52.2 GB/s (1.00x) memmem/subtitles/never/huge-en-some-rare-bytes 11.0 GB/s (5.77x) 63.6 GB/s (1.00x) memmem/subtitles/never/huge-en-two-space 2.3 GB/s (27.77x) 63.5 GB/s (1.00x) memmem/subtitles/never/huge-ru-john-watson 5.2 GB/s (11.56x) 59.9 GB/s (1.00x) memmem/subtitles/never/huge-zh-john-watson 20.7 GB/s (2.86x) 59.2 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock-holmes 17.0 GB/s (3.71x) 63.1 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock 11.8 GB/s (5.18x) 60.9 GB/s (1.00x) memmem/subtitles/rare/huge-en-huge-needle 19.3 GB/s (2.02x) 38.9 GB/s (1.00x) memmem/subtitles/rare/huge-ru-sherlock-holmes 6.5 GB/s (9.47x) 61.5 GB/s (1.00x) memmem/subtitles/rare/huge-ru-sherlock 3.8 GB/s (16.23x) 61.6 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock 10.8 GB/s (5.48x) 59.1 GB/s (1.00x) ``` Differences with the [`bytecount`](https://crates.io/crates/bytecount) crate as `memchr_iter(needle, haystack).count()` is now specialized to its own vector implementation just for counting the number of matches (instead of reporting the offset of each match). The thoughput improvements as compared to `bytecount` on large haystacks are most interesting IMO. (I was somewhat surprised by this, as `bytecount` seems to do something clever while `memchr_iter(needle, haystack).count()` is basically just `memchr` but with the branching for reporting matches removed.) Either way, I expect this to translate directly to improvements in ripgrep, although I haven't measured that yet. ``` $ rebar cmp benchmarks/record/x86_64/2023-08-26.csv -e '^rust/bytecount/memchr/oneshot$' -e '^rust/memchr/memchr/onlycount$' benchmark rust/bytecount/memchr/oneshot rust/memchr/memchr/onlycount --------- ----------------------------- ---------------------------- memchr/sherlock/common/huge1 28.5 GB/s (1.94x) 55.3 GB/s (1.00x) memchr/sherlock/common/small1 17.7 GB/s (1.25x) 22.1 GB/s (1.00x) memchr/sherlock/common/tiny1 4.3 GB/s (1.00x) 3.8 GB/s (1.13x) memchr/sherlock/never/huge1 28.4 GB/s (2.09x) 59.3 GB/s (1.00x) memchr/sherlock/never/small1 17.7 GB/s (1.25x) 22.1 GB/s (1.00x) memchr/sherlock/never/tiny1 4.3 GB/s (1.00x) 3.8 GB/s (1.13x) memchr/sherlock/never/empty1 11.00ns (1.00x) 11.00ns (1.00x) memchr/sherlock/rare/huge1 28.5 GB/s (1.94x) 55.2 GB/s (1.00x) memchr/sherlock/rare/small1 17.7 GB/s (1.25x) 22.1 GB/s (1.00x) memchr/sherlock/rare/tiny1 4.3 GB/s (1.00x) 3.8 GB/s (1.13x) memchr/sherlock/uncommon/huge1 26.9 GB/s (2.20x) 59.3 GB/s (1.00x) memchr/sherlock/uncommon/small1 17.7 GB/s (1.25x) 22.1 GB/s (1.00x) memchr/sherlock/uncommon/tiny1 4.3 GB/s (1.00x) 3.8 GB/s (1.13x) memchr/sherlock/verycommon/huge1 28.4 GB/s (2.09x) 59.3 GB/s (1.00x) memchr/sherlock/verycommon/small1 17.7 GB/s (1.25x) 22.1 GB/s (1.00x) ``` Differences across the board from the status quo. Note that here, I've only included measurements with a 4x difference from the old memchr crate. Otherwise, pretty much every benchmark has a pretty sizeable improvement from the old version. (Because previously, `aarch64` had no vector implementations at all.) ``` $ rebar diff tmp/old-aarch64.csv tmp/new-aarch64.csv -t 4 -E oneshot benchmark engine tmp/old-aarch64.csv tmp/new-aarch64.csv --------- ------ ------------------- ------------------- memchr/sherlock/never/huge2 rust/memchr/memchr2 10.8 GB/s (4.27x) 46.3 GB/s (1.00x) memchr/sherlock/never/small1 rust/memchr/memchr/prebuilt 15.1 GB/s (41.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/small1 rust/memchr/memrchr 14.7 GB/s (42.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/small2 rust/memchr/memchr2 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/small2 rust/memchr/memrchr2 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/small3 rust/memchr/memchr3 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/small3 rust/memchr/memrchr3 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/rare/small1 rust/memchr/memchr/prebuilt 14.7 GB/s (42.00x) 618.4 GB/s (1.00x) memchr/sherlock/rare/small1 rust/memchr/memrchr 14.7 GB/s (42.00x) 618.4 GB/s (1.00x) memchr/sherlock/rare/small2 rust/memchr/memchr2 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/rare/small2 rust/memchr/memrchr2 7.5 GB/s (83.00x) 618.4 GB/s (1.00x) memchr/sherlock/uncommon/tiny1 rust/memchr/memchr/prebuilt 1605.0 MB/s (41.00x) 64.3 GB/s (1.00x) memchr/sherlock/uncommon/tiny1 rust/memchr/memrchr 1605.0 MB/s (41.00x) 64.3 GB/s (1.00x) memmem/code/rust-library-never-fn-strength rust/memchr/memmem/prebuilt 7.1 GB/s (4.17x) 29.6 GB/s (1.00x) memmem/code/rust-library-never-fn-strength-paren rust/memchr/memmem/prebuilt 6.9 GB/s (4.19x) 29.0 GB/s (1.00x) memmem/code/rust-library-rare-fn-from-str rust/memchr/memmem/prebuilt 6.5 GB/s (4.42x) 28.7 GB/s (1.00x) memmem/code/rust-library-common-fn rust/memchr/memmem/prebuilt 3.2 GB/s (5.58x) 18.0 GB/s (1.00x) memmem/code/rust-library-common-let rust/memchr/memmem/prebuilt 2012.9 MB/s (6.45x) 12.7 GB/s (1.00x) memmem/pathological/md5-huge-no-hash rust/memchr/memmem/prebuilt 1070.2 MB/s (24.69x) 25.8 GB/s (1.00x) memmem/pathological/md5-huge-last-hash rust/memchr/memmem/prebuilt 1148.2 MB/s (22.85x) 25.6 GB/s (1.00x) memmem/pathological/rare-repeated-huge-tricky rust/memchr/memmem/prebuilt 1299.3 MB/s (23.87x) 30.3 GB/s (1.00x) memmem/pathological/rare-repeated-small-tricky rust/memchr/memmem/prebuilt 1146.0 MB/s (19.83x) 22.2 GB/s (1.00x) memmem/sliceslice/seemingly-random rust/memchr/memmem/prebuilt 1485.7 KB/s (4.13x) 6.0 MB/s (1.00x) memmem/sliceslice/i386 rust/memchr/memmem/prebuilt 6.0 MB/s (5.07x) 30.3 MB/s (1.00x) memmem/subtitles/common/huge-en-that rust/memchr/memmem/prebuilt 1418.2 MB/s (11.50x) 15.9 GB/s (1.00x) memmem/subtitles/common/huge-ru-that rust/memchr/memmem/prebuilt 1389.1 MB/s (13.44x) 18.2 GB/s (1.00x) memmem/subtitles/common/huge-ru-not rust/memchr/memmem/prebuilt 1482.7 MB/s (7.06x) 10.2 GB/s (1.00x) memmem/subtitles/never/huge-en-all-common-bytes rust/memchr/memmem/prebuilt 1813.7 MB/s (12.81x) 22.7 GB/s (1.00x) memmem/subtitles/never/huge-en-two-space rust/memchr/memmem/prebuilt 1370.2 MB/s (25.23x) 33.8 GB/s (1.00x) memmem/subtitles/never/teeny-en-two-space rust/memchr/memmem/prebuilt 651.3 MB/s (41.00x) 26.1 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock rust/memchr/memmem/prebuilt 7.0 GB/s (4.40x) 30.6 GB/s (1.00x) memmem/subtitles/rare/huge-en-medium-needle rust/memchr/memmem/prebuilt 6.4 GB/s (4.43x) 28.3 GB/s (1.00x) memmem/subtitles/rare/huge-en-long-needle rust/memchr/memmem/prebuilt 7.1 GB/s (4.64x) 32.8 GB/s (1.00x) memmem/subtitles/rare/teeny-en-sherlock-holmes rust/memchr/memmem/prebuilt 651.3 MB/s (41.00x) 26.1 GB/s (1.00x) memmem/subtitles/rare/teeny-en-sherlock rust/memchr/memmem/prebuilt 651.3 MB/s (41.00x) 26.1 GB/s (1.00x) memmem/subtitles/rare/teeny-ru-sherlock-holmes rust/memchr/memmem/prebuilt 953.7 MB/s (42.00x) 39.1 GB/s (1.00x) memmem/subtitles/rare/teeny-ru-sherlock rust/memchr/memmem/prebuilt 976.9 MB/s (41.00x) 39.1 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock-holmes rust/memchr/memmem/prebuilt 4.1 GB/s (7.06x) 28.8 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock rust/memchr/memmem/prebuilt 6.1 GB/s (4.81x) 29.6 GB/s (1.00x) memmem/subtitles/rare/teeny-zh-sherlock-holmes rust/memchr/memmem/prebuilt 721.1 MB/s (41.00x) 28.9 GB/s (1.00x) memmem/subtitles/rare/teeny-zh-sherlock rust/memchr/memmem/prebuilt 721.1 MB/s (41.00x) 28.9 GB/s (1.00x) ``` A comparison with the [`sliceslice`](https://crates.io/crates/sliceslice) crate, which has its own custom `aarch64` vector implementation of substring search. We only show measurements with 1.2x or greater difference. ``` $ rebar cmp benchmarks/record/aarch64/2023-08-26.csv -e sliceslice/memmem/prebuilt -e rust/memchr/memmem/prebuilt -t 1.2 benchmark rust/memchr/memmem/prebuilt rust/sliceslice/memmem/prebuilt --------- --------------------------- ------------------------------- memmem/byterank/binary 3.1 GB/s (1.00x) 1586.4 MB/s (2.01x) memmem/code/rust-library-never-fn-strength 29.6 GB/s (1.00x) 16.1 GB/s (1.84x) memmem/code/rust-library-never-fn-strength-paren 29.0 GB/s (1.00x) 15.6 GB/s (1.86x) memmem/code/rust-library-never-fn-quux 30.2 GB/s (1.00x) 15.1 GB/s (2.00x) memmem/code/rust-library-rare-fn-from-str 28.7 GB/s (1.93x) 55.5 GB/s (1.00x) memmem/pathological/md5-huge-no-hash 25.8 GB/s (1.00x) 13.6 GB/s (1.89x) memmem/pathological/md5-huge-last-hash 25.6 GB/s (1.00x) 13.5 GB/s (1.90x) memmem/pathological/rare-repeated-huge-tricky 30.3 GB/s (1.00x) 16.6 GB/s (1.83x) memmem/pathological/rare-repeated-small-tricky 22.2 GB/s (1.00x) 11.2 GB/s (1.98x) memmem/pathological/defeat-simple-vector-alphabet 3.0 GB/s (1.00x) 1114.1 MB/s (2.77x) memmem/pathological/defeat-simple-vector-freq-alphabet 14.8 GB/s (1.00x) 2.2 GB/s (6.72x) memmem/pathological/defeat-simple-vector-repeated-alphabet 835.1 MB/s (1.00x) 173.8 MB/s (4.80x) memmem/sliceslice/short 7.33ms (1.00x) 36.55ms (4.99x) memmem/sliceslice/seemingly-random 6.0 MB/s (1.00x) 3.6 MB/s (1.67x) memmem/sliceslice/i386 30.3 MB/s (1.00x) 15.1 MB/s (2.00x) memmem/subtitles/never/huge-en-john-watson 30.9 GB/s (1.00x) 16.6 GB/s (1.86x) memmem/subtitles/never/huge-en-all-common-bytes 22.7 GB/s (1.00x) 13.8 GB/s (1.64x) memmem/subtitles/never/huge-en-some-rare-bytes 30.9 GB/s (1.00x) 16.6 GB/s (1.86x) memmem/subtitles/never/huge-en-two-space 33.8 GB/s (1.00x) 16.6 GB/s (2.03x) memmem/subtitles/never/huge-ru-john-watson 30.3 GB/s (1.00x) 7.1 GB/s (4.25x) memmem/subtitles/never/huge-zh-john-watson 29.2 GB/s (1.00x) 16.0 GB/s (1.83x) memmem/subtitles/rare/huge-en-sherlock-holmes 30.3 GB/s (1.00x) 16.3 GB/s (1.86x) memmem/subtitles/rare/huge-en-sherlock 30.6 GB/s (1.00x) 16.6 GB/s (1.85x) memmem/subtitles/rare/huge-en-medium-needle 28.3 GB/s (1.00x) 12.4 GB/s (2.28x) memmem/subtitles/rare/huge-en-long-needle 32.8 GB/s (1.00x) 15.7 GB/s (2.08x) memmem/subtitles/rare/huge-en-huge-needle 32.9 GB/s (1.00x) 16.1 GB/s (2.05x) memmem/subtitles/rare/huge-ru-sherlock-holmes 30.3 GB/s (1.00x) 8.0 GB/s (3.80x) memmem/subtitles/rare/huge-ru-sherlock 30.2 GB/s (1.00x) 10.1 GB/s (3.00x) memmem/subtitles/rare/huge-zh-sherlock-holmes 28.8 GB/s (1.00x) 14.7 GB/s (1.95x) memmem/subtitles/rare/huge-zh-sherlock 29.6 GB/s (1.00x) 14.0 GB/s (2.12x) ``` Differences with the substring search implementation and `memmem` as provided by macOS's libc. Showing only measurements with 2x difference or greater. This is what utter destruction looks like. (I'm not sure what's going on in benchmarks like `memmem/subtitles/rare/teeny-en-sherlock-holmes`. It's a tiny haystack and macOS seems to either measure 1ns or 41ns. I wonder if there's something odd about time precision on macOS? You can see the reverse happen in `memmem/subtitles/rare/teeny-zh-sherlock`.) ``` $ rebar cmp benchmarks/record/aarch64/2023-08-26.csv -e libc/memmem/oneshot -e rust/memchr/memmem/oneshot -t 2 benchmark libc/memmem/oneshot rust/memchr/memmem/oneshot --------- ------------------- -------------------------- memmem/byterank/binary 626.1 MB/s (5.11x) 3.1 GB/s (1.00x) memmem/code/rust-library-never-fn-strength 1320.8 MB/s (22.98x) 29.6 GB/s (1.00x) memmem/code/rust-library-never-fn-strength-paren 1320.8 MB/s (22.49x) 29.0 GB/s (1.00x) memmem/code/rust-library-never-fn-quux 1332.0 MB/s (23.25x) 30.2 GB/s (1.00x) memmem/code/rust-library-rare-fn-from-str 1442.0 MB/s (20.37x) 28.7 GB/s (1.00x) memmem/code/rust-library-common-fn-is-empty 1320.8 MB/s (22.02x) 28.4 GB/s (1.00x) memmem/code/rust-library-common-fn 1320.8 MB/s (11.44x) 14.8 GB/s (1.00x) memmem/code/rust-library-common-let 1114.7 MB/s (8.59x) 9.4 GB/s (1.00x) memmem/pathological/md5-huge-no-hash 994.0 MB/s (26.39x) 25.6 GB/s (1.00x) memmem/pathological/md5-huge-last-hash 994.3 MB/s (26.39x) 25.6 GB/s (1.00x) memmem/pathological/rare-repeated-huge-tricky 1670.8 MB/s (18.56x) 30.3 GB/s (1.00x) memmem/pathological/rare-repeated-huge-match 1353.0 MB/s (1.00x) 378.5 MB/s (3.57x) memmem/pathological/rare-repeated-small-tricky 1637.4 MB/s (13.88x) 22.2 GB/s (1.00x) memmem/pathological/rare-repeated-small-match 1348.3 MB/s (1.00x) 394.5 MB/s (3.42x) memmem/pathological/defeat-simple-vector-alphabet 568.1 MB/s (5.43x) 3.0 GB/s (1.00x) memmem/pathological/defeat-simple-vector-freq-alphabet 1027.2 MB/s (14.55x) 14.6 GB/s (1.00x) memmem/pathological/defeat-simple-vector-repeated-alphabet 173.8 MB/s (4.80x) 834.2 MB/s (1.00x) memmem/subtitles/common/huge-en-that 841.6 MB/s (13.19x) 10.8 GB/s (1.00x) memmem/subtitles/common/huge-en-you 1161.7 MB/s (4.00x) 4.5 GB/s (1.00x) memmem/subtitles/common/huge-ru-that 590.9 MB/s (19.48x) 11.2 GB/s (1.00x) memmem/subtitles/common/huge-ru-not 334.3 MB/s (18.62x) 6.1 GB/s (1.00x) memmem/subtitles/common/huge-zh-that 1340.1 MB/s (11.49x) 15.0 GB/s (1.00x) memmem/subtitles/common/huge-zh-do-not 858.5 MB/s (9.15x) 7.7 GB/s (1.00x) memmem/subtitles/never/huge-en-john-watson 1648.3 MB/s (19.14x) 30.8 GB/s (1.00x) memmem/subtitles/never/huge-en-all-common-bytes 1075.4 MB/s (21.65x) 22.7 GB/s (1.00x) memmem/subtitles/never/huge-en-some-rare-bytes 1655.7 MB/s (19.10x) 30.9 GB/s (1.00x) memmem/subtitles/never/huge-en-two-space 541.6 MB/s (63.83x) 33.8 GB/s (1.00x) memmem/subtitles/never/teeny-en-two-space 651.3 MB/s (41.00x) 26.1 GB/s (1.00x) memmem/subtitles/never/huge-ru-john-watson 427.0 MB/s (72.56x) 30.3 GB/s (1.00x) memmem/subtitles/never/huge-zh-john-watson 1155.4 MB/s (25.81x) 29.1 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock-holmes 1577.4 MB/s (19.60x) 30.2 GB/s (1.00x) memmem/subtitles/rare/huge-en-sherlock 1577.4 MB/s (19.78x) 30.5 GB/s (1.00x) memmem/subtitles/rare/huge-en-medium-needle 1155.6 MB/s (24.95x) 28.2 GB/s (1.00x) memmem/subtitles/rare/huge-en-long-needle 1488.8 MB/s (20.77x) 30.2 GB/s (1.00x) memmem/subtitles/rare/huge-en-huge-needle 1609.5 MB/s (17.27x) 27.1 GB/s (1.00x) memmem/subtitles/rare/teeny-en-sherlock-holmes 26.1 GB/s (1.00x) 651.3 MB/s (41.00x) memmem/subtitles/rare/huge-ru-sherlock-holmes 427.0 MB/s (72.41x) 30.2 GB/s (1.00x) memmem/subtitles/rare/huge-ru-sherlock 348.2 MB/s (91.21x) 31.0 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock-holmes 955.8 MB/s (31.66x) 29.6 GB/s (1.00x) memmem/subtitles/rare/huge-zh-sherlock 853.4 MB/s (35.46x) 29.6 GB/s (1.00x) memmem/subtitles/rare/teeny-zh-sherlock-holmes 28.9 GB/s (1.00x) 721.1 MB/s (41.00x) memmem/subtitles/rare/teeny-zh-sherlock 721.1 MB/s (41.00x) 28.9 GB/s (1.00x) ``` Differences with the [`bytecount`](https://crates.io/crates/bytecount) crate as `memchr_iter(needle, haystack).count()` is now specialized to its own vector implementation just for counting the number of matches (instead of reporting the offset of each match). ``` $ rebar cmp benchmarks/record/aarch64/2023-08-26.csv -e '^rust/bytecount/memchr/oneshot$' -e '^rust/memchr/memchr/onlycount$' benchmark rust/bytecount/memchr/oneshot rust/memchr/memchr/onlycount --------- ----------------------------- ---------------------------- memchr/sherlock/common/huge1 29.5 GB/s (1.40x) 41.4 GB/s (1.00x) memchr/sherlock/common/small1 618.4 GB/s (1.00x) 618.4 GB/s (1.00x) memchr/sherlock/common/tiny1 64.3 GB/s (1.00x) 64.3 GB/s (1.00x) memchr/sherlock/never/huge1 29.5 GB/s (1.40x) 41.4 GB/s (1.00x) memchr/sherlock/never/small1 618.4 GB/s (1.00x) 618.4 GB/s (1.00x) memchr/sherlock/never/tiny1 64.3 GB/s (1.00x) 64.3 GB/s (1.00x) memchr/sherlock/never/empty1 1.00ns (1.00x) 1.00ns (1.00x) memchr/sherlock/rare/huge1 29.5 GB/s (1.40x) 41.4 GB/s (1.00x) memchr/sherlock/rare/small1 618.4 GB/s (1.00x) 618.4 GB/s (1.00x) memchr/sherlock/rare/tiny1 64.3 GB/s (1.00x) 64.3 GB/s (1.00x) memchr/sherlock/uncommon/huge1 29.5 GB/s (1.40x) 41.4 GB/s (1.00x) memchr/sherlock/uncommon/small1 618.4 GB/s (1.00x) 618.4 GB/s (1.00x) memchr/sherlock/uncommon/tiny1 64.3 GB/s (1.00x) 64.3 GB/s (1.00x) memchr/sherlock/verycommon/huge1 28.7 GB/s (1.44x) 41.4 GB/s (1.00x) memchr/sherlock/verycommon/small1 618.4 GB/s (1.00x) 618.4 GB/s (1.00x) ```
BurntSushi
force-pushed
the
ag/refactor
branch
from
49c143f
to
5dd2de9
Compare
|
Cool! Before replacing bytecount with |
osiewicz
added a commit
to zed-industries/zed
that referenced
this pull request
Aug 28, 2023
Fresh off the press, memchr 2.6.0 adds vector search routines for aarch64. That directly improves our search performance for both text and regex searches. Per BurntSushi's claims, the simple string searches in ripgrep got ~2 times faster (more details available in BurntSushi/memchr#129).
osiewicz
added a commit
to zed-industries/zed
that referenced
this pull request
Aug 28, 2023
Fresh off the press, memchr 2.6.0 adds vector search routines for aarch64. That directly improves our search performance for both text and regex searches. Per BurntSushi's claims, the simple string searches in ripgrep got ~2 times faster (more details available in BurntSushi/memchr#129). Release Notes: - N/A
|
@llogiq Note that it's not just for |
|
@llogiq If you want to run memchr's benchmarks with a specific focus on |
This was referenced Aug 28, 2023
BurntSushi
added a commit
to BurntSushi/ripgrep
that referenced
this pull request
Aug 29, 2023
This in particular brings in a PR[1] that provides huge speedups on aarch64 (e.g., Apple silicon). [1]: BurntSushi/memchr#129
tomtau
added a commit
to tomtau/pest
that referenced
this pull request
Aug 29, 2023
memchr now supports aarch64: BurntSushi/memchr#129 cargo lib (not-bootstrap-in-src) did not respect umask: GHSA-j3xp-wfr4-hx87 (probably not an issue here)
BurntSushi
added a commit
to rust-lang/regex
that referenced
this pull request
Aug 29, 2023
This bumps the minimum memchr version to 2.6, which brings in massive improvements to aarch64 for single substring search. We also can now enable the new `alloc` feature in `memchr` when `alloc` is enable for `regex` and `regex-automata`. We also squash some warnings. [1]: BurntSushi/memchr#129
BurntSushi
added a commit
to rust-lang/regex
that referenced
this pull request
Aug 29, 2023
This bumps the minimum memchr version to 2.6, which brings in massive improvements to aarch64 for single substring search. We also can now enable the new `alloc` feature in `memchr` when `alloc` is enable for `regex` and `regex-automata`. We also squash some warnings. [1]: BurntSushi/memchr#129
BurntSushi
added a commit
to rust-lang/regex
that referenced
this pull request
Aug 29, 2023
This bumps the minimum memchr version to 2.6, which brings in massive improvements to aarch64 for single substring search. We also can now enable the new `alloc` feature in `memchr` when `alloc` is enable for `regex` and `regex-automata`. We also squash some warnings. [1]: BurntSushi/memchr#129
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR doesn't just add
aarch64-specific code, but it refactors pretty much everything about how the code is organized. There are big perf wins foraarch64(see benchmark results below), and also latency improvements across the board. A brief summary of the changes in this PR:aarch64NEON vector implementations formemchr,memrchr,memchr2,memrchr2,memchr3,memrchr3andmemmem. This should lead to massive speed improvements on an increasing popular target, due in large part to Apple silicon.wasm32simd128 vector implementations formemchr,memrchr,memchr2,memrchr2,memchr3andmemrchr3. (@alexcrichton previously contributed a vector implementation formemmemand that remains.)x86_64has no real additions other than thememchr_iter(needle, haystack).count()specialization. It already has SSE2 and AVX2 implementations ofmemchr(and friends) andmemmem. It uses AVX2 automatically via runtime inspection of what the current CPU supports. There is no need to compile with theavx2feature enabled.archsub-module that exposes a lot of the internal routines (including target specific routines) used to implementmemchrandmemmem. This module is part of a major refactoring of how this crate is organized and it seemed prudent to expose the internals as their APIs are pretty straight-forward. That is, there isn't a huge API design space IMO. This module includes scalar substring search implementations of Shift-Or, Rabin-Karp and Two-Way.cfgknobs that the previous version ofmemchrsetup in its build script. This in turn allowed me to remove the build script entirely. Given the ubiquity of this crate, this may lead to compile time improvements downstream. (Likely small in each individual case but perhaps large in aggregate.) I can't promise that a build script will never re-appear, but I'll try to resist adding one in the future if possible.time rebar build -e '^rust/memchrold/memmem/prebuilt$'reports 0.944 seconds on my system while a freshtime rebar build -e '^rust/memchr/memmem/prebuilt$'reports 1.164 seconds. This is onx86_64where no real additional code was added. This could be because of the "nicer" abstractions now present in thearchsub-module or perhaps how the internals are structured. (Previously there were multiple monomorphic implementations ofmemchrfor example and now there is a single generic implementation that is monomorphized automatically by the compiler via generics. Perhaps that is more expensive?)memchr_iter(needle, haystack).count()to use a different vector implementation that specifically only counts matches instead of reporting the offsets of each match. This can make huge (potentially over an order of magnitude) differences when counting the number of matches of a frequently (even semi-frequently) occurring byte in a large haystack. This is effectively what thebytecountcrate does (which is what ripgrep currently uses to compute line numbers for matches), but the marginal cost of adding it to thememchrcrate was very low. So I did. And I plan to move ripgrep to usingmemchr_iter(needle, haystack).count(). (Also, the benchmarks below suggest that the counting implementation I wrote is faster than the one inbytecountin some cases which look like they'll be relevant for ripgrep. This was surprising to me.)allocfeature which permits compiling this crate without the standard library but with thealloccrate. This crate is designed through-and-through to work in a core-only context, so this doesn't unlock much compared to just disabling thestdfeature. It adds a couple of APIs requiring allocation (likememmem::Finder::into_owned) and other things likearch::all::shiftorwhich really want an allocation to store its bit-parallel state machine.libcfeature is DEPRECATED and is now a no-op. I don't think there is any real benefit to it any more.loggingfeature has been added. When enabled, this crate will emit a smattering of log messages. Usually these messages are used to indicate what kind of strategy is selected. For example, whether a vector or scalar algorithm is used for substring search.Selected benchmark differences for
x86_64Differences across the board from the status quo. Showing only measurements with a 1.2x (or greater) difference.
A comparison with the
sliceslicecrate for just substring search. We only include measurements with a 1.2x difference or greater.Differences with the substring search implementation and
memmemas provided by GNU libc. Showing only measurements with 2x difference or greater.Differences with the
bytecountcrate asmemchr_iter(needle, haystack).count()is now specialized to its own vector implementation just for counting the number of matches (instead of reporting the offset of each match). The thoughput improvements as compared tobytecounton large haystacks are most interesting IMO. (I was somewhat surprised by this, asbytecountseems to do something clever whilememchr_iter(needle, haystack).count()is basically justmemchrbut with the branching for reporting matches removed.) Either way, I expect this to translate directly to improvements in ripgrep, although I haven't measured that yet.Selected benchmark differences for
aarch64Differences across the board from the status quo. Note that here, I've only included measurements with a 4x difference from the old memchr crate. Otherwise, pretty much every benchmark has a pretty sizeable improvement from the old version. (Because previously,
aarch64had no vector implementations at all.)A comparison with the
sliceslicecrate, which has its own customaarch64vector implementation of substring search. We only show measurements with 1.2x or greater difference.Differences with the substring search implementation and
memmemas provided by macOS's libc. Showing only measurements with 2x difference or greater. This is what utter destruction looks like. (I'm not sure what's going on in benchmarks likememmem/subtitles/rare/teeny-en-sherlock-holmes. It's a tiny haystack and macOS seems to either measure 1ns or 41ns. I wonder if there's something odd about time precision on macOS? You can see the reverse happen inmemmem/subtitles/rare/teeny-zh-sherlock.)Differences with the
bytecountcrate asmemchr_iter(needle, haystack).count()is now specialized to its own vector implementation just for counting the number of matches (instead of reporting the offset of each match).Selected Benchmark differences for
regexonaarch64This shows benchmark results before and after this change for the
regexcrate. We only show results with a difference of 1.2x or greater.Improvements to ripgrep on
aarch64In short, simple ripgrep searches (likely the most common kind) get about twice as fast on Apple silicon now.