| commit | e0f3915da0802c59f1b5a4ed7b2efaf8ce7c35d2 | [log] [tgz] |
|---|---|---|
| author | Jeff Vander Stoep <jeffv@google.com> | Fri Feb 02 21:18:13 2024 +0000 |
| committer | Automerger Merge Worker <android-build-automerger-merge-worker@system.gserviceaccount.com> | Fri Feb 02 21:18:13 2024 +0000 |
| tree | 5f12d185796d37de3d623e6fe7cc8c4d55802e46 | |
| parent | 798849bc13a1f88c49dc9f48e6e22d6e01c466b7 [diff] | |
| parent | 7e8ed9a1367d663f60d148ba2a7aa58ddafa034c [diff] |
Upgrade memchr to 2.7.1 am: 7e8ed9a136 Original change: https://android-review.googlesource.com/c/platform/external/rust/crates/memchr/+/2944244 Change-Id: Id79cc69eb51c871084ffcfaaf6dd5554de6bf819 Signed-off-by: Automerger Merge Worker <android-build-automerger-merge-worker@system.gserviceaccount.com>
This library provides heavily optimized routines for string search primitives.
Dual-licensed under MIT or the UNLICENSE.
memmem sub-module provides forward and reverse substring search routines.In all such cases, routines operate on &[u8] without regard to encoding. This is exactly what you want when searching either UTF-8 or arbitrary bytes.
memchr links to the standard library by default, but you can disable the std feature if you want to use it in a #![no_std] crate:
[dependencies] memchr = { version = "2", default-features = false }
On x86_64 platforms, when the std feature is disabled, the SSE2 accelerated implementations will be used. When std is enabled, AVX2 accelerated implementations will be used if the CPU is determined to support it at runtime.
SIMD accelerated routines are also available on the wasm32 and aarch64 targets. The std feature is not required to use them.
When a SIMD version is not available, then this crate falls back to SWAR techniques.
This crate's minimum supported rustc version is 1.61.0.
The current policy is that the minimum Rust version required to use this crate can be increased in minor version updates. For example, if crate 1.0 requires Rust 1.20.0, then crate 1.0.z for all values of z will also require Rust 1.20.0 or newer. However, crate 1.y for y > 0 may require a newer minimum version of Rust.
In general, this crate will be conservative with respect to the minimum supported version of Rust.
Given the complexity of the code in this crate, along with the pervasive use of unsafe, this crate has an extensive testing strategy. It combines multiple approaches:
quickcheck.cargo fuzz.Improvements to the testing infrastructure are very welcome.
At time of writing, this crate's implementation of substring search actually has a few different algorithms to choose from depending on the situation.
We'll start by establishing what the difference in performance actually is. There are two relevant benchmark classes to consider: prebuilt and oneshot. The prebuilt benchmarks are designed to measure---to the extent possible---search time only. That is, the benchmark first starts by building a searcher and then only tracking the time for using the searcher:
$ rebar rank benchmarks/record/x86_64/2023-08-26.csv --intersection -e memchr/memmem/prebuilt -e std/memmem/prebuilt Engine Version Geometric mean of speed ratios Benchmark count ------ ------- ------------------------------ --------------- rust/memchr/memmem/prebuilt 2.5.0 1.03 53 rust/std/memmem/prebuilt 1.73.0-nightly 180dffba1 6.50 53
Conversely, the oneshot benchmark class measures the time it takes to both build the searcher and use it:
$ rebar rank benchmarks/record/x86_64/2023-08-26.csv --intersection -e memchr/memmem/oneshot -e std/memmem/oneshot Engine Version Geometric mean of speed ratios Benchmark count ------ ------- ------------------------------ --------------- rust/memchr/memmem/oneshot 2.5.0 1.04 53 rust/std/memmem/oneshot 1.73.0-nightly 180dffba1 5.26 53
NOTE: Replace rebar rank with rebar cmp in the above commands to explore the specific benchmarks and their differences.
So in both cases, this crate is quite a bit faster over a broad sampling of benchmarks regardless of whether you measure only search time or search time plus construction time. The difference is a little smaller when you include construction time in your measurements.
These two different types of benchmark classes make for a nice segue into one reason why the standard library‘s substring search can be slower: API design. In the standard library, the only APIs available to you require one to re-construct the searcher for every search. While you can benefit from building a searcher once and iterating over all matches in a single string, you cannot reuse that searcher to search other strings. This might come up when, for example, searching a file one line at a time. You’ll need to re-build the searcher for every line searched, and this can really matter.
NOTE: The prebuilt benchmark for the standard library can‘t actually avoid measuring searcher construction at some level, because there is no API for it. Instead, the benchmark consists of building the searcher once and then finding all matches in a single string via an iterator. This tends to approximate a benchmark where searcher construction isn’t measured, but it isn't perfect. While this means the comparison is not strictly apples-to-apples, it does reflect what is maximally possible with the standard library, and thus reflects the best that one could do in a real world scenario.
While there is more to the story than just API design here, it‘s important to point out that even if the standard library’s substring search were a precise clone of this crate internally, it would still be at a disadvantage in some workloads because of its API. (The same also applies to C's standard library memmem function. There is no way to amortize construction of the searcher. You need to pay for it on every call.)
The other reason for the difference in performance is that the standard library has trouble using SIMD. In particular, substring search is implemented in the core library, where platform specific code generally can‘t exist. That’s an issue because in order to utilize SIMD beyond SSE2 while maintaining portable binaries, one needs to use dynamic CPU feature detection, and that in turn requires platform specific code. While there is an RFC for enabling target feature detection in core, it doesn't yet exist.
The bottom line here is that core's substring search implementation is limited to making use of SSE2, but not AVX.
Still though, this crate does accelerate substring search even when only SSE2 is available. The standard library could therefore adopt the techniques in this crate just for SSE2. The reason why that hasn‘t happened yet isn’t totally clear to me. It likely needs a champion to push it through. The standard library tends to be more conservative in these things. With that said, the standard library does use some SSE2 acceleration on x86-64 added in this PR. However, at the time of writing, it is only used for short needles and doesn't use the frequency based heuristics found in this crate.
NOTE: Another thing worth mentioning is that the standard library's substring search routine requires that both the needle and haystack have type &str. Unless you can assume that your data is valid UTF-8, building a &str will come with the overhead of UTF-8 validation. This may in turn result in overall slower searching depending on your workload. In contrast, the memchr crate permits both the needle and the haystack to have type &[u8], where &[u8] can be created from a &str with zero cost. Therefore, the substring search in this crate is strictly more flexible than what the standard library provides.