1use crate::arch::all::{
2 packedpair::{HeuristicFrequencyRank, Pair},
3 rabinkarp, twoway,
4};
5
6#[cfg(target_arch = "aarch64")]
7use crate::arch::aarch64::neon::packedpair as neon;
8#[cfg(target_arch = "wasm32")]
9use crate::arch::wasm32::simd128::packedpair as simd128;
10#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
11use crate::arch::x86_64::{
12 avx2::packedpair as avx2, sse2::packedpair as sse2,
13};
14
15/// A "meta" substring searcher.
16///
17/// To a first approximation, this chooses what it believes to be the "best"
18/// substring search implemnetation based on the needle at construction time.
19/// Then, every call to `find` will execute that particular implementation. To
20/// a second approximation, multiple substring search algorithms may be used,
21/// depending on the haystack. For example, for supremely short haystacks,
22/// Rabin-Karp is typically used.
23///
24/// See the documentation on `Prefilter` for an explanation of the dispatching
25/// mechanism. The quick summary is that an enum has too much overhead and
26/// we can't use dynamic dispatch via traits because we need to work in a
27/// core-only environment. (Dynamic dispatch works in core-only, but you
28/// need `&dyn Trait` and we really need a `Box<dyn Trait>` here. The latter
29/// requires `alloc`.) So instead, we use a union and an appropriately paired
30/// free function to read from the correct field on the union and execute the
31/// chosen substring search implementation.
32#[derive(Clone)]
33pub(crate) struct Searcher {
34 call: SearcherKindFn,
35 kind: SearcherKind,
36 rabinkarp: rabinkarp::Finder,
37}
38
39impl Searcher {
40 /// Creates a new "meta" substring searcher that attempts to choose the
41 /// best algorithm based on the needle, heuristics and what the current
42 /// target supports.
43 #[inline]
44 pub(crate) fn new<R: HeuristicFrequencyRank>(
45 prefilter: PrefilterConfig,
46 ranker: R,
47 needle: &[u8],
48 ) -> Searcher {
49 let rabinkarp = rabinkarp::Finder::new(needle);
50 if needle.len() <= 1 {
51 return if needle.is_empty() {
52 trace!("building empty substring searcher");
53 Searcher {
54 call: searcher_kind_empty,
55 kind: SearcherKind { empty: () },
56 rabinkarp,
57 }
58 } else {
59 trace!("building one-byte substring searcher");
60 debug_assert_eq!(1, needle.len());
61 Searcher {
62 call: searcher_kind_one_byte,
63 kind: SearcherKind { one_byte: needle[0] },
64 rabinkarp,
65 }
66 };
67 }
68 let pair = match Pair::with_ranker(needle, &ranker) {
69 Some(pair) => pair,
70 None => return Searcher::twoway(needle, rabinkarp, None),
71 };
72 debug_assert_ne!(
73 pair.index1(),
74 pair.index2(),
75 "pair offsets should not be equivalent"
76 );
77 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
78 {
79 if let Some(pp) = avx2::Finder::with_pair(needle, pair) {
80 if do_packed_search(needle) {
81 trace!("building x86_64 AVX2 substring searcher");
82 let kind = SearcherKind { avx2: pp };
83 Searcher { call: searcher_kind_avx2, kind, rabinkarp }
84 } else if prefilter.is_none() {
85 Searcher::twoway(needle, rabinkarp, None)
86 } else {
87 let prestrat = Prefilter::avx2(pp, needle);
88 Searcher::twoway(needle, rabinkarp, Some(prestrat))
89 }
90 } else if let Some(pp) = sse2::Finder::with_pair(needle, pair) {
91 if do_packed_search(needle) {
92 trace!("building x86_64 SSE2 substring searcher");
93 let kind = SearcherKind { sse2: pp };
94 Searcher { call: searcher_kind_sse2, kind, rabinkarp }
95 } else if prefilter.is_none() {
96 Searcher::twoway(needle, rabinkarp, None)
97 } else {
98 let prestrat = Prefilter::sse2(pp, needle);
99 Searcher::twoway(needle, rabinkarp, Some(prestrat))
100 }
101 } else if prefilter.is_none() {
102 Searcher::twoway(needle, rabinkarp, None)
103 } else {
104 // We're pretty unlikely to get to this point, but it is
105 // possible to be running on x86_64 without SSE2. Namely, it's
106 // really up to the OS whether it wants to support vector
107 // registers or not.
108 let prestrat = Prefilter::fallback(ranker, pair, needle);
109 Searcher::twoway(needle, rabinkarp, prestrat)
110 }
111 }
112 #[cfg(target_arch = "wasm32")]
113 {
114 if let Some(pp) = simd128::Finder::with_pair(needle, pair) {
115 if do_packed_search(needle) {
116 trace!("building wasm32 simd128 substring searcher");
117 let kind = SearcherKind { simd128: pp };
118 Searcher { call: searcher_kind_simd128, kind, rabinkarp }
119 } else if prefilter.is_none() {
120 Searcher::twoway(needle, rabinkarp, None)
121 } else {
122 let prestrat = Prefilter::simd128(pp, needle);
123 Searcher::twoway(needle, rabinkarp, Some(prestrat))
124 }
125 } else if prefilter.is_none() {
126 Searcher::twoway(needle, rabinkarp, None)
127 } else {
128 let prestrat = Prefilter::fallback(ranker, pair, needle);
129 Searcher::twoway(needle, rabinkarp, prestrat)
130 }
131 }
132 #[cfg(target_arch = "aarch64")]
133 {
134 if let Some(pp) = neon::Finder::with_pair(needle, pair) {
135 if do_packed_search(needle) {
136 trace!("building aarch64 neon substring searcher");
137 let kind = SearcherKind { neon: pp };
138 Searcher { call: searcher_kind_neon, kind, rabinkarp }
139 } else if prefilter.is_none() {
140 Searcher::twoway(needle, rabinkarp, None)
141 } else {
142 let prestrat = Prefilter::neon(pp, needle);
143 Searcher::twoway(needle, rabinkarp, Some(prestrat))
144 }
145 } else if prefilter.is_none() {
146 Searcher::twoway(needle, rabinkarp, None)
147 } else {
148 let prestrat = Prefilter::fallback(ranker, pair, needle);
149 Searcher::twoway(needle, rabinkarp, prestrat)
150 }
151 }
152 #[cfg(not(any(
153 all(target_arch = "x86_64", target_feature = "sse2"),
154 target_arch = "wasm32",
155 target_arch = "aarch64"
156 )))]
157 {
158 if prefilter.is_none() {
159 Searcher::twoway(needle, rabinkarp, None)
160 } else {
161 let prestrat = Prefilter::fallback(ranker, pair, needle);
162 Searcher::twoway(needle, rabinkarp, prestrat)
163 }
164 }
165 }
166
167 /// Creates a new searcher that always uses the Two-Way algorithm. This is
168 /// typically used when vector algorithms are unavailable or inappropriate.
169 /// (For example, when the needle is "too long.")
170 ///
171 /// If a prefilter is given, then the searcher returned will be accelerated
172 /// by the prefilter.
173 #[inline]
174 fn twoway(
175 needle: &[u8],
176 rabinkarp: rabinkarp::Finder,
177 prestrat: Option<Prefilter>,
178 ) -> Searcher {
179 let finder = twoway::Finder::new(needle);
180 match prestrat {
181 None => {
182 trace!("building scalar two-way substring searcher");
183 let kind = SearcherKind { two_way: finder };
184 Searcher { call: searcher_kind_two_way, kind, rabinkarp }
185 }
186 Some(prestrat) => {
187 trace!(
188 "building scalar two-way \
189 substring searcher with a prefilter"
190 );
191 let two_way_with_prefilter =
192 TwoWayWithPrefilter { finder, prestrat };
193 let kind = SearcherKind { two_way_with_prefilter };
194 Searcher {
195 call: searcher_kind_two_way_with_prefilter,
196 kind,
197 rabinkarp,
198 }
199 }
200 }
201 }
202
203 /// Searches the given haystack for the given needle. The needle given
204 /// should be the same as the needle that this finder was initialized
205 /// with.
206 ///
207 /// Inlining this can lead to big wins for latency, and #[inline] doesn't
208 /// seem to be enough in some cases.
209 #[inline(always)]
210 pub(crate) fn find(
211 &self,
212 prestate: &mut PrefilterState,
213 haystack: &[u8],
214 needle: &[u8],
215 ) -> Option<usize> {
216 if haystack.len() < needle.len() {
217 None
218 } else {
219 // SAFETY: By construction, we've ensured that the function
220 // in `self.call` is properly paired with the union used in
221 // `self.kind`.
222 unsafe { (self.call)(self, prestate, haystack, needle) }
223 }
224 }
225}
226
227impl core::fmt::Debug for Searcher {
228 fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
229 f.debug_struct("Searcher")
230 .field("call", &"<searcher function>")
231 .field("kind", &"<searcher kind union>")
232 .field("rabinkarp", &self.rabinkarp)
233 .finish()
234 }
235}
236
237/// A union indicating one of several possible substring search implementations
238/// that are in active use.
239///
240/// This union should only be read by one of the functions prefixed with
241/// `searcher_kind_`. Namely, the correct function is meant to be paired with
242/// the union by the caller, such that the function always reads from the
243/// designated union field.
244#[derive(Clone, Copy)]
245union SearcherKind {
246 empty: (),
247 one_byte: u8,
248 two_way: twoway::Finder,
249 two_way_with_prefilter: TwoWayWithPrefilter,
250 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
251 sse2: crate::arch::x86_64::sse2::packedpair::Finder,
252 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
253 avx2: crate::arch::x86_64::avx2::packedpair::Finder,
254 #[cfg(target_arch = "wasm32")]
255 simd128: crate::arch::wasm32::simd128::packedpair::Finder,
256 #[cfg(target_arch = "aarch64")]
257 neon: crate::arch::aarch64::neon::packedpair::Finder,
258}
259
260/// A two-way substring searcher with a prefilter.
261#[derive(Copy, Clone, Debug)]
262struct TwoWayWithPrefilter {
263 finder: twoway::Finder,
264 prestrat: Prefilter,
265}
266
267/// The type of a substring search function.
268///
269/// # Safety
270///
271/// When using a function of this type, callers must ensure that the correct
272/// function is paired with the value populated in `SearcherKind` union.
273type SearcherKindFn = unsafe fn(
274 searcher: &Searcher,
275 prestate: &mut PrefilterState,
276 haystack: &[u8],
277 needle: &[u8],
278) -> Option<usize>;
279
280/// Reads from the `empty` field of `SearcherKind` to handle the case of
281/// searching for the empty needle. Works on all platforms.
282///
283/// # Safety
284///
285/// Callers must ensure that the `searcher.kind.empty` union field is set.
286unsafe fn searcher_kind_empty(
287 _searcher: &Searcher,
288 _prestate: &mut PrefilterState,
289 _haystack: &[u8],
290 _needle: &[u8],
291) -> Option<usize> {
292 Some(0)
293}
294
295/// Reads from the `one_byte` field of `SearcherKind` to handle the case of
296/// searching for a single byte needle. Works on all platforms.
297///
298/// # Safety
299///
300/// Callers must ensure that the `searcher.kind.one_byte` union field is set.
301unsafe fn searcher_kind_one_byte(
302 searcher: &Searcher,
303 _prestate: &mut PrefilterState,
304 haystack: &[u8],
305 _needle: &[u8],
306) -> Option<usize> {
307 let needle = searcher.kind.one_byte;
308 crate::memchr(needle, haystack)
309}
310
311/// Reads from the `two_way` field of `SearcherKind` to handle the case of
312/// searching for an arbitrary needle without prefilter acceleration. Works on
313/// all platforms.
314///
315/// # Safety
316///
317/// Callers must ensure that the `searcher.kind.two_way` union field is set.
318unsafe fn searcher_kind_two_way(
319 searcher: &Searcher,
320 _prestate: &mut PrefilterState,
321 haystack: &[u8],
322 needle: &[u8],
323) -> Option<usize> {
324 if rabinkarp::is_fast(haystack, needle) {
325 searcher.rabinkarp.find(haystack, needle)
326 } else {
327 searcher.kind.two_way.find(haystack, needle)
328 }
329}
330
331/// Reads from the `two_way_with_prefilter` field of `SearcherKind` to handle
332/// the case of searching for an arbitrary needle with prefilter acceleration.
333/// Works on all platforms.
334///
335/// # Safety
336///
337/// Callers must ensure that the `searcher.kind.two_way_with_prefilter` union
338/// field is set.
339unsafe fn searcher_kind_two_way_with_prefilter(
340 searcher: &Searcher,
341 prestate: &mut PrefilterState,
342 haystack: &[u8],
343 needle: &[u8],
344) -> Option<usize> {
345 if rabinkarp::is_fast(haystack, needle) {
346 searcher.rabinkarp.find(haystack, needle)
347 } else {
348 let TwoWayWithPrefilter { ref finder, ref prestrat } =
349 searcher.kind.two_way_with_prefilter;
350 let pre = Pre { prestate, prestrat };
351 finder.find_with_prefilter(Some(pre), haystack, needle)
352 }
353}
354
355/// Reads from the `sse2` field of `SearcherKind` to execute the x86_64 SSE2
356/// vectorized substring search implementation.
357///
358/// # Safety
359///
360/// Callers must ensure that the `searcher.kind.sse2` union field is set.
361#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
362unsafe fn searcher_kind_sse2(
363 searcher: &Searcher,
364 _prestate: &mut PrefilterState,
365 haystack: &[u8],
366 needle: &[u8],
367) -> Option<usize> {
368 let finder = &searcher.kind.sse2;
369 if haystack.len() < finder.min_haystack_len() {
370 searcher.rabinkarp.find(haystack, needle)
371 } else {
372 finder.find(haystack, needle)
373 }
374}
375
376/// Reads from the `avx2` field of `SearcherKind` to execute the x86_64 AVX2
377/// vectorized substring search implementation.
378///
379/// # Safety
380///
381/// Callers must ensure that the `searcher.kind.avx2` union field is set.
382#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
383unsafe fn searcher_kind_avx2(
384 searcher: &Searcher,
385 _prestate: &mut PrefilterState,
386 haystack: &[u8],
387 needle: &[u8],
388) -> Option<usize> {
389 let finder = &searcher.kind.avx2;
390 if haystack.len() < finder.min_haystack_len() {
391 searcher.rabinkarp.find(haystack, needle)
392 } else {
393 finder.find(haystack, needle)
394 }
395}
396
397/// Reads from the `simd128` field of `SearcherKind` to execute the wasm32
398/// simd128 vectorized substring search implementation.
399///
400/// # Safety
401///
402/// Callers must ensure that the `searcher.kind.simd128` union field is set.
403#[cfg(target_arch = "wasm32")]
404unsafe fn searcher_kind_simd128(
405 searcher: &Searcher,
406 _prestate: &mut PrefilterState,
407 haystack: &[u8],
408 needle: &[u8],
409) -> Option<usize> {
410 let finder = &searcher.kind.simd128;
411 if haystack.len() < finder.min_haystack_len() {
412 searcher.rabinkarp.find(haystack, needle)
413 } else {
414 finder.find(haystack, needle)
415 }
416}
417
418/// Reads from the `neon` field of `SearcherKind` to execute the aarch64 neon
419/// vectorized substring search implementation.
420///
421/// # Safety
422///
423/// Callers must ensure that the `searcher.kind.neon` union field is set.
424#[cfg(target_arch = "aarch64")]
425unsafe fn searcher_kind_neon(
426 searcher: &Searcher,
427 _prestate: &mut PrefilterState,
428 haystack: &[u8],
429 needle: &[u8],
430) -> Option<usize> {
431 let finder = &searcher.kind.neon;
432 if haystack.len() < finder.min_haystack_len() {
433 searcher.rabinkarp.find(haystack, needle)
434 } else {
435 finder.find(haystack, needle)
436 }
437}
438
439/// A reverse substring searcher.
440#[derive(Clone, Debug)]
441pub(crate) struct SearcherRev {
442 kind: SearcherRevKind,
443 rabinkarp: rabinkarp::FinderRev,
444}
445
446/// The kind of the reverse searcher.
447///
448/// For the reverse case, we don't do any SIMD acceleration or prefilters.
449/// There is no specific technical reason why we don't, but rather don't do it
450/// because it's not clear it's worth the extra code to do so. If you have a
451/// use case for it, please file an issue.
452///
453/// We also don't do the union trick as we do with the forward case and
454/// prefilters. Basically for the same reason we don't have prefilters or
455/// vector algorithms for reverse searching: it's not clear it's worth doing.
456/// Please file an issue if you have a compelling use case for fast reverse
457/// substring search.
458#[derive(Clone, Debug)]
459enum SearcherRevKind {
460 Empty,
461 OneByte { needle: u8 },
462 TwoWay { finder: twoway::FinderRev },
463}
464
465impl SearcherRev {
466 /// Creates a new searcher for finding occurrences of the given needle in
467 /// reverse. That is, it reports the last (instead of the first) occurrence
468 /// of a needle in a haystack.
469 #[inline]
470 pub(crate) fn new(needle: &[u8]) -> SearcherRev {
471 let kind = if needle.len() <= 1 {
472 if needle.is_empty() {
473 trace!("building empty reverse substring searcher");
474 SearcherRevKind::Empty
475 } else {
476 trace!("building one-byte reverse substring searcher");
477 debug_assert_eq!(1, needle.len());
478 SearcherRevKind::OneByte { needle: needle[0] }
479 }
480 } else {
481 trace!("building scalar two-way reverse substring searcher");
482 let finder = twoway::FinderRev::new(needle);
483 SearcherRevKind::TwoWay { finder }
484 };
485 let rabinkarp = rabinkarp::FinderRev::new(needle);
486 SearcherRev { kind, rabinkarp }
487 }
488
489 /// Searches the given haystack for the last occurrence of the given
490 /// needle. The needle given should be the same as the needle that this
491 /// finder was initialized with.
492 #[inline]
493 pub(crate) fn rfind(
494 &self,
495 haystack: &[u8],
496 needle: &[u8],
497 ) -> Option<usize> {
498 if haystack.len() < needle.len() {
499 return None;
500 }
501 match self.kind {
502 SearcherRevKind::Empty => Some(haystack.len()),
503 SearcherRevKind::OneByte { needle } => {
504 crate::memrchr(needle, haystack)
505 }
506 SearcherRevKind::TwoWay { ref finder } => {
507 if rabinkarp::is_fast(haystack, needle) {
508 self.rabinkarp.rfind(haystack, needle)
509 } else {
510 finder.rfind(haystack, needle)
511 }
512 }
513 }
514 }
515}
516
517/// Prefilter controls whether heuristics are used to accelerate searching.
518///
519/// A prefilter refers to the idea of detecting candidate matches very quickly,
520/// and then confirming whether those candidates are full matches. This
521/// idea can be quite effective since it's often the case that looking for
522/// candidates can be a lot faster than running a complete substring search
523/// over the entire input. Namely, looking for candidates can be done with
524/// extremely fast vectorized code.
525///
526/// The downside of a prefilter is that it assumes false positives (which are
527/// candidates generated by a prefilter that aren't matches) are somewhat rare
528/// relative to the frequency of full matches. That is, if a lot of false
529/// positives are generated, then it's possible for search time to be worse
530/// than if the prefilter wasn't enabled in the first place.
531///
532/// Another downside of a prefilter is that it can result in highly variable
533/// performance, where some cases are extraordinarily fast and others aren't.
534/// Typically, variable performance isn't a problem, but it may be for your use
535/// case.
536///
537/// The use of prefilters in this implementation does use a heuristic to detect
538/// when a prefilter might not be carrying its weight, and will dynamically
539/// disable its use. Nevertheless, this configuration option gives callers
540/// the ability to disable prefilters if you have knowledge that they won't be
541/// useful.
542#[derive(Clone, Copy, Debug)]
543#[non_exhaustive]
544pub enum PrefilterConfig {
545 /// Never used a prefilter in substring search.
546 None,
547 /// Automatically detect whether a heuristic prefilter should be used. If
548 /// it is used, then heuristics will be used to dynamically disable the
549 /// prefilter if it is believed to not be carrying its weight.
550 Auto,
551}
552
553impl Default for PrefilterConfig {
554 fn default() -> PrefilterConfig {
555 PrefilterConfig::Auto
556 }
557}
558
559impl PrefilterConfig {
560 /// Returns true when this prefilter is set to the `None` variant.
561 fn is_none(&self) -> bool {
562 matches!(*self, PrefilterConfig::None)
563 }
564}
565
566/// The implementation of a prefilter.
567///
568/// This type encapsulates dispatch to one of several possible choices for a
569/// prefilter. Generally speaking, all prefilters have the same approximate
570/// algorithm: they choose a couple of bytes from the needle that are believed
571/// to be rare, use a fast vector algorithm to look for those bytes and return
572/// positions as candidates for some substring search algorithm (currently only
573/// Two-Way) to confirm as a match or not.
574///
575/// The differences between the algorithms are actually at the vector
576/// implementation level. Namely, we need different routines based on both
577/// which target architecture we're on and what CPU features are supported.
578///
579/// The straight-forwardly obvious approach here is to use an enum, and make
580/// `Prefilter::find` do case analysis to determine which algorithm was
581/// selected and invoke it. However, I've observed that this leads to poor
582/// codegen in some cases, especially in latency sensitive benchmarks. That is,
583/// this approach comes with overhead that I wasn't able to eliminate.
584///
585/// The second obvious approach is to use dynamic dispatch with traits. Doing
586/// that in this context where `Prefilter` owns the selection generally
587/// requires heap allocation, and this code is designed to run in core-only
588/// environments.
589///
590/// So we settle on using a union (that's `PrefilterKind`) and a function
591/// pointer (that's `PrefilterKindFn`). We select the right function pointer
592/// based on which field in the union we set, and that function in turn
593/// knows which field of the union to access. The downside of this approach
594/// is that it forces us to think about safety, but the upside is that
595/// there are some nice latency improvements to benchmarks. (Especially the
596/// `memmem/sliceslice/short` benchmark.)
597///
598/// In cases where we've selected a vector algorithm and the haystack given
599/// is too short, we fallback to the scalar version of `memchr` on the
600/// `rarest_byte`. (The scalar version of `memchr` is still better than a naive
601/// byte-at-a-time loop because it will read in `usize`-sized chunks at a
602/// time.)
603#[derive(Clone, Copy)]
604struct Prefilter {
605 call: PrefilterKindFn,
606 kind: PrefilterKind,
607 rarest_byte: u8,
608 rarest_offset: u8,
609}
610
611impl Prefilter {
612 /// Return a "fallback" prefilter, but only if it is believed to be
613 /// effective.
614 #[inline]
615 fn fallback<R: HeuristicFrequencyRank>(
616 ranker: R,
617 pair: Pair,
618 needle: &[u8],
619 ) -> Option<Prefilter> {
620 /// The maximum frequency rank permitted for the fallback prefilter.
621 /// If the rarest byte in the needle has a frequency rank above this
622 /// value, then no prefilter is used if the fallback prefilter would
623 /// otherwise be selected.
624 const MAX_FALLBACK_RANK: u8 = 250;
625
626 trace!("building fallback prefilter");
627 let rarest_offset = pair.index1();
628 let rarest_byte = needle[usize::from(rarest_offset)];
629 let rarest_rank = ranker.rank(rarest_byte);
630 if rarest_rank > MAX_FALLBACK_RANK {
631 None
632 } else {
633 let finder = crate::arch::all::packedpair::Finder::with_pair(
634 needle,
635 pair.clone(),
636 )?;
637 let call = prefilter_kind_fallback;
638 let kind = PrefilterKind { fallback: finder };
639 Some(Prefilter { call, kind, rarest_byte, rarest_offset })
640 }
641 }
642
643 /// Return a prefilter using a x86_64 SSE2 vector algorithm.
644 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
645 #[inline]
646 fn sse2(finder: sse2::Finder, needle: &[u8]) -> Prefilter {
647 trace!("building x86_64 SSE2 prefilter");
648 let rarest_offset = finder.pair().index1();
649 let rarest_byte = needle[usize::from(rarest_offset)];
650 Prefilter {
651 call: prefilter_kind_sse2,
652 kind: PrefilterKind { sse2: finder },
653 rarest_byte,
654 rarest_offset,
655 }
656 }
657
658 /// Return a prefilter using a x86_64 AVX2 vector algorithm.
659 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
660 #[inline]
661 fn avx2(finder: avx2::Finder, needle: &[u8]) -> Prefilter {
662 trace!("building x86_64 AVX2 prefilter");
663 let rarest_offset = finder.pair().index1();
664 let rarest_byte = needle[usize::from(rarest_offset)];
665 Prefilter {
666 call: prefilter_kind_avx2,
667 kind: PrefilterKind { avx2: finder },
668 rarest_byte,
669 rarest_offset,
670 }
671 }
672
673 /// Return a prefilter using a wasm32 simd128 vector algorithm.
674 #[cfg(target_arch = "wasm32")]
675 #[inline]
676 fn simd128(finder: simd128::Finder, needle: &[u8]) -> Prefilter {
677 trace!("building wasm32 simd128 prefilter");
678 let rarest_offset = finder.pair().index1();
679 let rarest_byte = needle[usize::from(rarest_offset)];
680 Prefilter {
681 call: prefilter_kind_simd128,
682 kind: PrefilterKind { simd128: finder },
683 rarest_byte,
684 rarest_offset,
685 }
686 }
687
688 /// Return a prefilter using a aarch64 neon vector algorithm.
689 #[cfg(target_arch = "aarch64")]
690 #[inline]
691 fn neon(finder: neon::Finder, needle: &[u8]) -> Prefilter {
692 trace!("building aarch64 neon prefilter");
693 let rarest_offset = finder.pair().index1();
694 let rarest_byte = needle[usize::from(rarest_offset)];
695 Prefilter {
696 call: prefilter_kind_neon,
697 kind: PrefilterKind { neon: finder },
698 rarest_byte,
699 rarest_offset,
700 }
701 }
702
703 /// Return a *candidate* position for a match.
704 ///
705 /// When this returns an offset, it implies that a match could begin at
706 /// that offset, but it may not. That is, it is possible for a false
707 /// positive to be returned.
708 ///
709 /// When `None` is returned, then it is guaranteed that there are no
710 /// matches for the needle in the given haystack. That is, it is impossible
711 /// for a false negative to be returned.
712 ///
713 /// The purpose of this routine is to look for candidate matching positions
714 /// as quickly as possible before running a (likely) slower confirmation
715 /// step.
716 #[inline]
717 fn find(&self, haystack: &[u8]) -> Option<usize> {
718 // SAFETY: By construction, we've ensured that the function in
719 // `self.call` is properly paired with the union used in `self.kind`.
720 unsafe { (self.call)(self, haystack) }
721 }
722
723 /// A "simple" prefilter that just looks for the occurrence of the rarest
724 /// byte from the needle. This is generally only used for very small
725 /// haystacks.
726 #[inline]
727 fn find_simple(&self, haystack: &[u8]) -> Option<usize> {
728 // We don't use crate::memchr here because the haystack should be small
729 // enough that memchr won't be able to use vector routines anyway. So
730 // we just skip straight to the fallback implementation which is likely
731 // faster. (A byte-at-a-time loop is only used when the haystack is
732 // smaller than `size_of::<usize>()`.)
733 crate::arch::all::memchr::One::new(self.rarest_byte)
734 .find(haystack)
735 .map(|i| i.saturating_sub(usize::from(self.rarest_offset)))
736 }
737}
738
739impl core::fmt::Debug for Prefilter {
740 fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
741 f.debug_struct("Prefilter")
742 .field("call", &"<prefilter function>")
743 .field("kind", &"<prefilter kind union>")
744 .field("rarest_byte", &self.rarest_byte)
745 .field("rarest_offset", &self.rarest_offset)
746 .finish()
747 }
748}
749
750/// A union indicating one of several possible prefilters that are in active
751/// use.
752///
753/// This union should only be read by one of the functions prefixed with
754/// `prefilter_kind_`. Namely, the correct function is meant to be paired with
755/// the union by the caller, such that the function always reads from the
756/// designated union field.
757#[derive(Clone, Copy)]
758union PrefilterKind {
759 fallback: crate::arch::all::packedpair::Finder,
760 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
761 sse2: crate::arch::x86_64::sse2::packedpair::Finder,
762 #[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
763 avx2: crate::arch::x86_64::avx2::packedpair::Finder,
764 #[cfg(target_arch = "wasm32")]
765 simd128: crate::arch::wasm32::simd128::packedpair::Finder,
766 #[cfg(target_arch = "aarch64")]
767 neon: crate::arch::aarch64::neon::packedpair::Finder,
768}
769
770/// The type of a prefilter function.
771///
772/// # Safety
773///
774/// When using a function of this type, callers must ensure that the correct
775/// function is paired with the value populated in `PrefilterKind` union.
776type PrefilterKindFn =
777 unsafe fn(strat: &Prefilter, haystack: &[u8]) -> Option<usize>;
778
779/// Reads from the `fallback` field of `PrefilterKind` to execute the fallback
780/// prefilter. Works on all platforms.
781///
782/// # Safety
783///
784/// Callers must ensure that the `strat.kind.fallback` union field is set.
785unsafe fn prefilter_kind_fallback(
786 strat: &Prefilter,
787 haystack: &[u8],
788) -> Option<usize> {
789 strat.kind.fallback.find_prefilter(haystack)
790}
791
792/// Reads from the `sse2` field of `PrefilterKind` to execute the x86_64 SSE2
793/// prefilter.
794///
795/// # Safety
796///
797/// Callers must ensure that the `strat.kind.sse2` union field is set.
798#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
799unsafe fn prefilter_kind_sse2(
800 strat: &Prefilter,
801 haystack: &[u8],
802) -> Option<usize> {
803 let finder = &strat.kind.sse2;
804 if haystack.len() < finder.min_haystack_len() {
805 strat.find_simple(haystack)
806 } else {
807 finder.find_prefilter(haystack)
808 }
809}
810
811/// Reads from the `avx2` field of `PrefilterKind` to execute the x86_64 AVX2
812/// prefilter.
813///
814/// # Safety
815///
816/// Callers must ensure that the `strat.kind.avx2` union field is set.
817#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))]
818unsafe fn prefilter_kind_avx2(
819 strat: &Prefilter,
820 haystack: &[u8],
821) -> Option<usize> {
822 let finder = &strat.kind.avx2;
823 if haystack.len() < finder.min_haystack_len() {
824 strat.find_simple(haystack)
825 } else {
826 finder.find_prefilter(haystack)
827 }
828}
829
830/// Reads from the `simd128` field of `PrefilterKind` to execute the wasm32
831/// simd128 prefilter.
832///
833/// # Safety
834///
835/// Callers must ensure that the `strat.kind.simd128` union field is set.
836#[cfg(target_arch = "wasm32")]
837unsafe fn prefilter_kind_simd128(
838 strat: &Prefilter,
839 haystack: &[u8],
840) -> Option<usize> {
841 let finder = &strat.kind.simd128;
842 if haystack.len() < finder.min_haystack_len() {
843 strat.find_simple(haystack)
844 } else {
845 finder.find_prefilter(haystack)
846 }
847}
848
849/// Reads from the `neon` field of `PrefilterKind` to execute the aarch64 neon
850/// prefilter.
851///
852/// # Safety
853///
854/// Callers must ensure that the `strat.kind.neon` union field is set.
855#[cfg(target_arch = "aarch64")]
856unsafe fn prefilter_kind_neon(
857 strat: &Prefilter,
858 haystack: &[u8],
859) -> Option<usize> {
860 let finder = &strat.kind.neon;
861 if haystack.len() < finder.min_haystack_len() {
862 strat.find_simple(haystack)
863 } else {
864 finder.find_prefilter(haystack)
865 }
866}
867
868/// PrefilterState tracks state associated with the effectiveness of a
869/// prefilter. It is used to track how many bytes, on average, are skipped by
870/// the prefilter. If this average dips below a certain threshold over time,
871/// then the state renders the prefilter inert and stops using it.
872///
873/// A prefilter state should be created for each search. (Where creating an
874/// iterator is treated as a single search.) A prefilter state should only be
875/// created from a `Freqy`. e.g., An inert `Freqy` will produce an inert
876/// `PrefilterState`.
877#[derive(Clone, Copy, Debug)]
878pub(crate) struct PrefilterState {
879 /// The number of skips that has been executed. This is always 1 greater
880 /// than the actual number of skips. The special sentinel value of 0
881 /// indicates that the prefilter is inert. This is useful to avoid
882 /// additional checks to determine whether the prefilter is still
883 /// "effective." Once a prefilter becomes inert, it should no longer be
884 /// used (according to our heuristics).
885 skips: u32,
886 /// The total number of bytes that have been skipped.
887 skipped: u32,
888}
889
890impl PrefilterState {
891 /// The minimum number of skip attempts to try before considering whether
892 /// a prefilter is effective or not.
893 const MIN_SKIPS: u32 = 50;
894
895 /// The minimum amount of bytes that skipping must average.
896 ///
897 /// This value was chosen based on varying it and checking
898 /// the microbenchmarks. In particular, this can impact the
899 /// pathological/repeated-{huge,small} benchmarks quite a bit if it's set
900 /// too low.
901 const MIN_SKIP_BYTES: u32 = 8;
902
903 /// Create a fresh prefilter state.
904 #[inline]
905 pub(crate) fn new() -> PrefilterState {
906 PrefilterState { skips: 1, skipped: 0 }
907 }
908
909 /// Update this state with the number of bytes skipped on the last
910 /// invocation of the prefilter.
911 #[inline]
912 fn update(&mut self, skipped: usize) {
913 self.skips = self.skips.saturating_add(1);
914 // We need to do this dance since it's technically possible for
915 // `skipped` to overflow a `u32`. (And we use a `u32` to reduce the
916 // size of a prefilter state.)
917 self.skipped = match u32::try_from(skipped) {
918 Err(_) => core::u32::MAX,
919 Ok(skipped) => self.skipped.saturating_add(skipped),
920 };
921 }
922
923 /// Return true if and only if this state indicates that a prefilter is
924 /// still effective.
925 #[inline]
926 fn is_effective(&mut self) -> bool {
927 if self.is_inert() {
928 return false;
929 }
930 if self.skips() < PrefilterState::MIN_SKIPS {
931 return true;
932 }
933 if self.skipped >= PrefilterState::MIN_SKIP_BYTES * self.skips() {
934 return true;
935 }
936
937 // We're inert.
938 self.skips = 0;
939 false
940 }
941
942 /// Returns true if the prefilter this state represents should no longer
943 /// be used.
944 #[inline]
945 fn is_inert(&self) -> bool {
946 self.skips == 0
947 }
948
949 /// Returns the total number of times the prefilter has been used.
950 #[inline]
951 fn skips(&self) -> u32 {
952 // Remember, `0` is a sentinel value indicating inertness, so we
953 // always need to subtract `1` to get our actual number of skips.
954 self.skips.saturating_sub(1)
955 }
956}
957
958/// A combination of prefilter effectiveness state and the prefilter itself.
959#[derive(Debug)]
960pub(crate) struct Pre<'a> {
961 /// State that tracks the effectiveness of a prefilter.
962 prestate: &'a mut PrefilterState,
963 /// The actual prefilter.
964 prestrat: &'a Prefilter,
965}
966
967impl<'a> Pre<'a> {
968 /// Call this prefilter on the given haystack with the given needle.
969 #[inline]
970 pub(crate) fn find(&mut self, haystack: &[u8]) -> Option<usize> {
971 let result = self.prestrat.find(haystack);
972 self.prestate.update(result.unwrap_or(haystack.len()));
973 result
974 }
975
976 /// Return true if and only if this prefilter should be used.
977 #[inline]
978 pub(crate) fn is_effective(&mut self) -> bool {
979 self.prestate.is_effective()
980 }
981}
982
983/// Returns true if the needle has the right characteristics for a vector
984/// algorithm to handle the entirety of substring search.
985///
986/// Vector algorithms can be used for prefilters for other substring search
987/// algorithms (like Two-Way), but they can also be used for substring search
988/// on their own. When used for substring search, vector algorithms will
989/// quickly identify candidate match positions (just like in the prefilter
990/// case), but instead of returning the candidate position they will try to
991/// confirm the match themselves. Confirmation happens via `memcmp`. This
992/// works well for short needles, but can break down when many false candidate
993/// positions are generated for large needles. Thus, we only permit vector
994/// algorithms to own substring search when the needle is of a certain length.
995#[inline]
996fn do_packed_search(needle: &[u8]) -> bool {
997 /// The minimum length of a needle required for this algorithm. The minimum
998 /// is 2 since a length of 1 should just use memchr and a length of 0 isn't
999 /// a case handled by this searcher.
1000 const MIN_LEN: usize = 2;
1001
1002 /// The maximum length of a needle required for this algorithm.
1003 ///
1004 /// In reality, there is no hard max here. The code below can handle any
1005 /// length needle. (Perhaps that suggests there are missing optimizations.)
1006 /// Instead, this is a heuristic and a bound guaranteeing our linear time
1007 /// complexity.
1008 ///
1009 /// It is a heuristic because when a candidate match is found, memcmp is
1010 /// run. For very large needles with lots of false positives, memcmp can
1011 /// make the code run quite slow.
1012 ///
1013 /// It is a bound because the worst case behavior with memcmp is
1014 /// multiplicative in the size of the needle and haystack, and we want
1015 /// to keep that additive. This bound ensures we still meet that bound
1016 /// theoretically, since it's just a constant. We aren't acting in bad
1017 /// faith here, memcmp on tiny needles is so fast that even in pathological
1018 /// cases (see pathological vector benchmarks), this is still just as fast
1019 /// or faster in practice.
1020 ///
1021 /// This specific number was chosen by tweaking a bit and running
1022 /// benchmarks. The rare-medium-needle, for example, gets about 5% faster
1023 /// by using this algorithm instead of a prefilter-accelerated Two-Way.
1024 /// There's also a theoretical desire to keep this number reasonably
1025 /// low, to mitigate the impact of pathological cases. I did try 64, and
1026 /// some benchmarks got a little better, and others (particularly the
1027 /// pathological ones), got a lot worse. So... 32 it is?
1028 const MAX_LEN: usize = 32;
1029 MIN_LEN <= needle.len() && needle.len() <= MAX_LEN
1030}
1031