]>
Commit | Line | Data |
---|---|---|
487cf647 FG |
1 | /*! |
2 | Types and routines specific to dense DFAs. | |
3 | ||
4 | This module is the home of [`dense::DFA`](DFA). | |
5 | ||
6 | This module also contains a [`dense::Builder`](Builder) and a | |
781aab86 | 7 | [`dense::Config`](Config) for building and configuring a dense DFA. |
487cf647 FG |
8 | */ |
9 | ||
781aab86 | 10 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
11 | use core::cmp; |
12 | use core::{convert::TryFrom, fmt, iter, mem::size_of, slice}; | |
13 | ||
781aab86 | 14 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
15 | use alloc::{ |
16 | collections::{BTreeMap, BTreeSet}, | |
17 | vec, | |
18 | vec::Vec, | |
19 | }; | |
20 | ||
781aab86 | 21 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
22 | use crate::{ |
23 | dfa::{ | |
781aab86 FG |
24 | accel::Accel, determinize, minimize::Minimizer, remapper::Remapper, |
25 | sparse, | |
487cf647 FG |
26 | }, |
27 | nfa::thompson, | |
781aab86 | 28 | util::{look::LookMatcher, search::MatchKind}, |
487cf647 FG |
29 | }; |
30 | use crate::{ | |
31 | dfa::{ | |
32 | accel::Accels, | |
33 | automaton::{fmt_state_indicator, Automaton}, | |
34 | special::Special, | |
781aab86 | 35 | start::StartKind, |
487cf647 FG |
36 | DEAD, |
37 | }, | |
38 | util::{ | |
781aab86 FG |
39 | alphabet::{self, ByteClasses, ByteSet}, |
40 | int::{Pointer, Usize}, | |
41 | prefilter::Prefilter, | |
42 | primitives::{PatternID, StateID}, | |
43 | search::{Anchored, Input, MatchError}, | |
44 | start::{Start, StartByteMap}, | |
45 | wire::{self, DeserializeError, Endian, SerializeError}, | |
487cf647 FG |
46 | }, |
47 | }; | |
48 | ||
49 | /// The label that is pre-pended to a serialized DFA. | |
50 | const LABEL: &str = "rust-regex-automata-dfa-dense"; | |
51 | ||
52 | /// The format version of dense regexes. This version gets incremented when a | |
53 | /// change occurs. A change may not necessarily be a breaking change, but the | |
54 | /// version does permit good error messages in the case where a breaking change | |
55 | /// is made. | |
56 | const VERSION: u32 = 2; | |
57 | ||
58 | /// The configuration used for compiling a dense DFA. | |
59 | /// | |
781aab86 FG |
60 | /// As a convenience, [`DFA::config`] is an alias for [`Config::new`]. The |
61 | /// advantage of the former is that it often lets you avoid importing the | |
62 | /// `Config` type directly. | |
63 | /// | |
487cf647 FG |
64 | /// A dense DFA configuration is a simple data object that is typically used |
65 | /// with [`dense::Builder::configure`](self::Builder::configure). | |
66 | /// | |
781aab86 FG |
67 | /// The default configuration guarantees that a search will never return |
68 | /// a "quit" error, although it is possible for a search to fail if | |
69 | /// [`Config::starts_for_each_pattern`] wasn't enabled (which it is not by | |
70 | /// default) and an [`Anchored::Pattern`] mode is requested via [`Input`]. | |
71 | #[cfg(feature = "dfa-build")] | |
72 | #[derive(Clone, Debug, Default)] | |
487cf647 FG |
73 | pub struct Config { |
74 | // As with other configuration types in this crate, we put all our knobs | |
75 | // in options so that we can distinguish between "default" and "not set." | |
76 | // This makes it possible to easily combine multiple configurations | |
77 | // without default values overwriting explicitly specified values. See the | |
78 | // 'overwrite' method. | |
79 | // | |
80 | // For docs on the fields below, see the corresponding method setters. | |
487cf647 | 81 | accelerate: Option<bool>, |
781aab86 | 82 | pre: Option<Option<Prefilter>>, |
487cf647 FG |
83 | minimize: Option<bool>, |
84 | match_kind: Option<MatchKind>, | |
781aab86 | 85 | start_kind: Option<StartKind>, |
487cf647 FG |
86 | starts_for_each_pattern: Option<bool>, |
87 | byte_classes: Option<bool>, | |
88 | unicode_word_boundary: Option<bool>, | |
781aab86 FG |
89 | quitset: Option<ByteSet>, |
90 | specialize_start_states: Option<bool>, | |
487cf647 FG |
91 | dfa_size_limit: Option<Option<usize>>, |
92 | determinize_size_limit: Option<Option<usize>>, | |
93 | } | |
94 | ||
781aab86 | 95 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
96 | impl Config { |
97 | /// Return a new default dense DFA compiler configuration. | |
98 | pub fn new() -> Config { | |
99 | Config::default() | |
100 | } | |
101 | ||
487cf647 FG |
102 | /// Enable state acceleration. |
103 | /// | |
104 | /// When enabled, DFA construction will analyze each state to determine | |
105 | /// whether it is eligible for simple acceleration. Acceleration typically | |
106 | /// occurs when most of a state's transitions loop back to itself, leaving | |
107 | /// only a select few bytes that will exit the state. When this occurs, | |
108 | /// other routines like `memchr` can be used to look for those bytes which | |
109 | /// may be much faster than traversing the DFA. | |
110 | /// | |
111 | /// Callers may elect to disable this if consistent performance is more | |
112 | /// desirable than variable performance. Namely, acceleration can sometimes | |
113 | /// make searching slower than it otherwise would be if the transitions | |
114 | /// that leave accelerated states are traversed frequently. | |
115 | /// | |
116 | /// See [`Automaton::accelerator`](crate::dfa::Automaton::accelerator) for | |
117 | /// an example. | |
118 | /// | |
119 | /// This is enabled by default. | |
120 | pub fn accelerate(mut self, yes: bool) -> Config { | |
121 | self.accelerate = Some(yes); | |
122 | self | |
123 | } | |
124 | ||
781aab86 FG |
125 | /// Set a prefilter to be used whenever a start state is entered. |
126 | /// | |
127 | /// A [`Prefilter`] in this context is meant to accelerate searches by | |
128 | /// looking for literal prefixes that every match for the corresponding | |
129 | /// pattern (or patterns) must start with. Once a prefilter produces a | |
130 | /// match, the underlying search routine continues on to try and confirm | |
131 | /// the match. | |
132 | /// | |
133 | /// Be warned that setting a prefilter does not guarantee that the search | |
134 | /// will be faster. While it's usually a good bet, if the prefilter | |
135 | /// produces a lot of false positive candidates (i.e., positions matched | |
136 | /// by the prefilter but not by the regex), then the overall result can | |
137 | /// be slower than if you had just executed the regex engine without any | |
138 | /// prefilters. | |
139 | /// | |
140 | /// Note that unless [`Config::specialize_start_states`] has been | |
141 | /// explicitly set, then setting this will also enable (when `pre` is | |
142 | /// `Some`) or disable (when `pre` is `None`) start state specialization. | |
143 | /// This occurs because without start state specialization, a prefilter | |
144 | /// is likely to be less effective. And without a prefilter, start state | |
145 | /// specialization is usually pointless. | |
146 | /// | |
147 | /// **WARNING:** Note that prefilters are not preserved as part of | |
148 | /// serialization. Serializing a DFA will drop its prefilter. | |
149 | /// | |
150 | /// By default no prefilter is set. | |
151 | /// | |
152 | /// # Example | |
153 | /// | |
154 | /// ``` | |
155 | /// use regex_automata::{ | |
156 | /// dfa::{dense::DFA, Automaton}, | |
157 | /// util::prefilter::Prefilter, | |
158 | /// Input, HalfMatch, MatchKind, | |
159 | /// }; | |
160 | /// | |
161 | /// let pre = Prefilter::new(MatchKind::LeftmostFirst, &["foo", "bar"]); | |
162 | /// let re = DFA::builder() | |
163 | /// .configure(DFA::config().prefilter(pre)) | |
164 | /// .build(r"(foo|bar)[a-z]+")?; | |
165 | /// let input = Input::new("foo1 barfox bar"); | |
166 | /// assert_eq!( | |
167 | /// Some(HalfMatch::must(0, 11)), | |
168 | /// re.try_search_fwd(&input)?, | |
169 | /// ); | |
170 | /// | |
171 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
172 | /// ``` | |
173 | /// | |
174 | /// Be warned though that an incorrect prefilter can lead to incorrect | |
175 | /// results! | |
176 | /// | |
177 | /// ``` | |
178 | /// use regex_automata::{ | |
179 | /// dfa::{dense::DFA, Automaton}, | |
180 | /// util::prefilter::Prefilter, | |
181 | /// Input, HalfMatch, MatchKind, | |
182 | /// }; | |
183 | /// | |
184 | /// let pre = Prefilter::new(MatchKind::LeftmostFirst, &["foo", "car"]); | |
185 | /// let re = DFA::builder() | |
186 | /// .configure(DFA::config().prefilter(pre)) | |
187 | /// .build(r"(foo|bar)[a-z]+")?; | |
188 | /// let input = Input::new("foo1 barfox bar"); | |
189 | /// assert_eq!( | |
190 | /// // No match reported even though there clearly is one! | |
191 | /// None, | |
192 | /// re.try_search_fwd(&input)?, | |
193 | /// ); | |
194 | /// | |
195 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
196 | /// ``` | |
197 | pub fn prefilter(mut self, pre: Option<Prefilter>) -> Config { | |
198 | self.pre = Some(pre); | |
199 | if self.specialize_start_states.is_none() { | |
200 | self.specialize_start_states = | |
201 | Some(self.get_prefilter().is_some()); | |
202 | } | |
203 | self | |
204 | } | |
205 | ||
487cf647 FG |
206 | /// Minimize the DFA. |
207 | /// | |
208 | /// When enabled, the DFA built will be minimized such that it is as small | |
209 | /// as possible. | |
210 | /// | |
211 | /// Whether one enables minimization or not depends on the types of costs | |
212 | /// you're willing to pay and how much you care about its benefits. In | |
213 | /// particular, minimization has worst case `O(n*k*logn)` time and `O(k*n)` | |
214 | /// space, where `n` is the number of DFA states and `k` is the alphabet | |
215 | /// size. In practice, minimization can be quite costly in terms of both | |
216 | /// space and time, so it should only be done if you're willing to wait | |
217 | /// longer to produce a DFA. In general, you might want a minimal DFA in | |
218 | /// the following circumstances: | |
219 | /// | |
220 | /// 1. You would like to optimize for the size of the automaton. This can | |
221 | /// manifest in one of two ways. Firstly, if you're converting the | |
222 | /// DFA into Rust code (or a table embedded in the code), then a minimal | |
223 | /// DFA will translate into a corresponding reduction in code size, and | |
224 | /// thus, also the final compiled binary size. Secondly, if you are | |
225 | /// building many DFAs and putting them on the heap, you'll be able to | |
226 | /// fit more if they are smaller. Note though that building a minimal | |
227 | /// DFA itself requires additional space; you only realize the space | |
228 | /// savings once the minimal DFA is constructed (at which point, the | |
229 | /// space used for minimization is freed). | |
230 | /// 2. You've observed that a smaller DFA results in faster match | |
231 | /// performance. Naively, this isn't guaranteed since there is no | |
232 | /// inherent difference between matching with a bigger-than-minimal | |
233 | /// DFA and a minimal DFA. However, a smaller DFA may make use of your | |
234 | /// CPU's cache more efficiently. | |
235 | /// 3. You are trying to establish an equivalence between regular | |
236 | /// languages. The standard method for this is to build a minimal DFA | |
237 | /// for each language and then compare them. If the DFAs are equivalent | |
238 | /// (up to state renaming), then the languages are equivalent. | |
239 | /// | |
240 | /// Typically, minimization only makes sense as an offline process. That | |
241 | /// is, one might minimize a DFA before serializing it to persistent | |
242 | /// storage. In practical terms, minimization can take around an order of | |
243 | /// magnitude more time than compiling the initial DFA via determinization. | |
244 | /// | |
245 | /// This option is disabled by default. | |
246 | pub fn minimize(mut self, yes: bool) -> Config { | |
247 | self.minimize = Some(yes); | |
248 | self | |
249 | } | |
250 | ||
251 | /// Set the desired match semantics. | |
252 | /// | |
253 | /// The default is [`MatchKind::LeftmostFirst`], which corresponds to the | |
254 | /// match semantics of Perl-like regex engines. That is, when multiple | |
255 | /// patterns would match at the same leftmost position, the pattern that | |
256 | /// appears first in the concrete syntax is chosen. | |
257 | /// | |
258 | /// Currently, the only other kind of match semantics supported is | |
259 | /// [`MatchKind::All`]. This corresponds to classical DFA construction | |
260 | /// where all possible matches are added to the DFA. | |
261 | /// | |
262 | /// Typically, `All` is used when one wants to execute an overlapping | |
263 | /// search and `LeftmostFirst` otherwise. In particular, it rarely makes | |
264 | /// sense to use `All` with the various "leftmost" find routines, since the | |
265 | /// leftmost routines depend on the `LeftmostFirst` automata construction | |
266 | /// strategy. Specifically, `LeftmostFirst` adds dead states to the DFA | |
267 | /// as a way to terminate the search and report a match. `LeftmostFirst` | |
268 | /// also supports non-greedy matches using this strategy where as `All` | |
269 | /// does not. | |
270 | /// | |
271 | /// # Example: overlapping search | |
272 | /// | |
273 | /// This example shows the typical use of `MatchKind::All`, which is to | |
274 | /// report overlapping matches. | |
275 | /// | |
276 | /// ``` | |
781aab86 | 277 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
487cf647 FG |
278 | /// use regex_automata::{ |
279 | /// dfa::{Automaton, OverlappingState, dense}, | |
781aab86 | 280 | /// HalfMatch, Input, MatchKind, |
487cf647 FG |
281 | /// }; |
282 | /// | |
283 | /// let dfa = dense::Builder::new() | |
284 | /// .configure(dense::Config::new().match_kind(MatchKind::All)) | |
285 | /// .build_many(&[r"\w+$", r"\S+$"])?; | |
781aab86 | 286 | /// let input = Input::new("@foo"); |
487cf647 FG |
287 | /// let mut state = OverlappingState::start(); |
288 | /// | |
289 | /// let expected = Some(HalfMatch::must(1, 4)); | |
781aab86 FG |
290 | /// dfa.try_search_overlapping_fwd(&input, &mut state)?; |
291 | /// assert_eq!(expected, state.get_match()); | |
487cf647 FG |
292 | /// |
293 | /// // The first pattern also matches at the same position, so re-running | |
294 | /// // the search will yield another match. Notice also that the first | |
295 | /// // pattern is returned after the second. This is because the second | |
296 | /// // pattern begins its match before the first, is therefore an earlier | |
297 | /// // match and is thus reported first. | |
298 | /// let expected = Some(HalfMatch::must(0, 4)); | |
781aab86 FG |
299 | /// dfa.try_search_overlapping_fwd(&input, &mut state)?; |
300 | /// assert_eq!(expected, state.get_match()); | |
487cf647 FG |
301 | /// |
302 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
303 | /// ``` | |
304 | /// | |
305 | /// # Example: reverse automaton to find start of match | |
306 | /// | |
307 | /// Another example for using `MatchKind::All` is for constructing a | |
308 | /// reverse automaton to find the start of a match. `All` semantics are | |
309 | /// used for this in order to find the longest possible match, which | |
310 | /// corresponds to the leftmost starting position. | |
311 | /// | |
312 | /// Note that if you need the starting position then | |
313 | /// [`dfa::regex::Regex`](crate::dfa::regex::Regex) will handle this for | |
314 | /// you, so it's usually not necessary to do this yourself. | |
315 | /// | |
316 | /// ``` | |
781aab86 FG |
317 | /// use regex_automata::{ |
318 | /// dfa::{dense, Automaton, StartKind}, | |
319 | /// nfa::thompson::NFA, | |
320 | /// Anchored, HalfMatch, Input, MatchKind, | |
321 | /// }; | |
487cf647 FG |
322 | /// |
323 | /// let haystack = "123foobar456".as_bytes(); | |
781aab86 | 324 | /// let pattern = r"[a-z]+r"; |
487cf647 FG |
325 | /// |
326 | /// let dfa_fwd = dense::DFA::new(pattern)?; | |
327 | /// let dfa_rev = dense::Builder::new() | |
781aab86 | 328 | /// .thompson(NFA::config().reverse(true)) |
487cf647 | 329 | /// .configure(dense::Config::new() |
781aab86 FG |
330 | /// // This isn't strictly necessary since both anchored and |
331 | /// // unanchored searches are supported by default. But since | |
332 | /// // finding the start-of-match only requires anchored searches, | |
333 | /// // we can get rid of the unanchored configuration and possibly | |
334 | /// // slim down our DFA considerably. | |
335 | /// .start_kind(StartKind::Anchored) | |
487cf647 FG |
336 | /// .match_kind(MatchKind::All) |
337 | /// ) | |
338 | /// .build(pattern)?; | |
339 | /// let expected_fwd = HalfMatch::must(0, 9); | |
340 | /// let expected_rev = HalfMatch::must(0, 3); | |
781aab86 | 341 | /// let got_fwd = dfa_fwd.try_search_fwd(&Input::new(haystack))?.unwrap(); |
487cf647 FG |
342 | /// // Here we don't specify the pattern to search for since there's only |
343 | /// // one pattern and we're doing a leftmost search. But if this were an | |
344 | /// // overlapping search, you'd need to specify the pattern that matched | |
345 | /// // in the forward direction. (Otherwise, you might wind up finding the | |
346 | /// // starting position of a match of some other pattern.) That in turn | |
347 | /// // requires building the reverse automaton with starts_for_each_pattern | |
348 | /// // enabled. Indeed, this is what Regex does internally. | |
781aab86 FG |
349 | /// let input = Input::new(haystack) |
350 | /// .range(..got_fwd.offset()) | |
351 | /// .anchored(Anchored::Yes); | |
352 | /// let got_rev = dfa_rev.try_search_rev(&input)?.unwrap(); | |
487cf647 FG |
353 | /// assert_eq!(expected_fwd, got_fwd); |
354 | /// assert_eq!(expected_rev, got_rev); | |
355 | /// | |
356 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
357 | /// ``` | |
358 | pub fn match_kind(mut self, kind: MatchKind) -> Config { | |
359 | self.match_kind = Some(kind); | |
360 | self | |
361 | } | |
362 | ||
781aab86 FG |
363 | /// The type of starting state configuration to use for a DFA. |
364 | /// | |
365 | /// By default, the starting state configuration is [`StartKind::Both`]. | |
366 | /// | |
367 | /// # Example | |
368 | /// | |
369 | /// ``` | |
370 | /// use regex_automata::{ | |
371 | /// dfa::{dense::DFA, Automaton, StartKind}, | |
372 | /// Anchored, HalfMatch, Input, | |
373 | /// }; | |
374 | /// | |
375 | /// let haystack = "quux foo123"; | |
376 | /// let expected = HalfMatch::must(0, 11); | |
377 | /// | |
378 | /// // By default, DFAs support both anchored and unanchored searches. | |
379 | /// let dfa = DFA::new(r"[0-9]+")?; | |
380 | /// let input = Input::new(haystack); | |
381 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&input)?); | |
382 | /// | |
383 | /// // But if we only need anchored searches, then we can build a DFA | |
384 | /// // that only supports anchored searches. This leads to a smaller DFA | |
385 | /// // (potentially significantly smaller in some cases), but a DFA that | |
386 | /// // will panic if you try to use it with an unanchored search. | |
387 | /// let dfa = DFA::builder() | |
388 | /// .configure(DFA::config().start_kind(StartKind::Anchored)) | |
389 | /// .build(r"[0-9]+")?; | |
390 | /// let input = Input::new(haystack) | |
391 | /// .range(8..) | |
392 | /// .anchored(Anchored::Yes); | |
393 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&input)?); | |
394 | /// | |
395 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
396 | /// ``` | |
397 | pub fn start_kind(mut self, kind: StartKind) -> Config { | |
398 | self.start_kind = Some(kind); | |
399 | self | |
400 | } | |
401 | ||
487cf647 FG |
402 | /// Whether to compile a separate start state for each pattern in the |
403 | /// automaton. | |
404 | /// | |
405 | /// When enabled, a separate **anchored** start state is added for each | |
406 | /// pattern in the DFA. When this start state is used, then the DFA will | |
407 | /// only search for matches for the pattern specified, even if there are | |
408 | /// other patterns in the DFA. | |
409 | /// | |
410 | /// The main downside of this option is that it can potentially increase | |
411 | /// the size of the DFA and/or increase the time it takes to build the DFA. | |
412 | /// | |
413 | /// There are a few reasons one might want to enable this (it's disabled | |
414 | /// by default): | |
415 | /// | |
416 | /// 1. When looking for the start of an overlapping match (using a | |
417 | /// reverse DFA), doing it correctly requires starting the reverse search | |
418 | /// using the starting state of the pattern that matched in the forward | |
419 | /// direction. Indeed, when building a [`Regex`](crate::dfa::regex::Regex), | |
420 | /// it will automatically enable this option when building the reverse DFA | |
421 | /// internally. | |
422 | /// 2. When you want to use a DFA with multiple patterns to both search | |
423 | /// for matches of any pattern or to search for anchored matches of one | |
424 | /// particular pattern while using the same DFA. (Otherwise, you would need | |
425 | /// to compile a new DFA for each pattern.) | |
426 | /// 3. Since the start states added for each pattern are anchored, if you | |
427 | /// compile an unanchored DFA with one pattern while also enabling this | |
428 | /// option, then you can use the same DFA to perform anchored or unanchored | |
429 | /// searches. The latter you get with the standard search APIs. The former | |
430 | /// you get from the various `_at` search methods that allow you specify a | |
431 | /// pattern ID to search for. | |
432 | /// | |
433 | /// By default this is disabled. | |
434 | /// | |
435 | /// # Example | |
436 | /// | |
437 | /// This example shows how to use this option to permit the same DFA to | |
438 | /// run both anchored and unanchored searches for a single pattern. | |
439 | /// | |
440 | /// ``` | |
441 | /// use regex_automata::{ | |
781aab86 FG |
442 | /// dfa::{dense, Automaton}, |
443 | /// Anchored, HalfMatch, PatternID, Input, | |
487cf647 FG |
444 | /// }; |
445 | /// | |
446 | /// let dfa = dense::Builder::new() | |
447 | /// .configure(dense::Config::new().starts_for_each_pattern(true)) | |
448 | /// .build(r"foo[0-9]+")?; | |
781aab86 | 449 | /// let haystack = "quux foo123"; |
487cf647 FG |
450 | /// |
451 | /// // Here's a normal unanchored search. Notice that we use 'None' for the | |
452 | /// // pattern ID. Since the DFA was built as an unanchored machine, it | |
453 | /// // use its default unanchored starting state. | |
454 | /// let expected = HalfMatch::must(0, 11); | |
781aab86 FG |
455 | /// let input = Input::new(haystack); |
456 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&input)?); | |
487cf647 FG |
457 | /// // But now if we explicitly specify the pattern to search ('0' being |
458 | /// // the only pattern in the DFA), then it will use the starting state | |
459 | /// // for that specific pattern which is always anchored. Since the | |
460 | /// // pattern doesn't have a match at the beginning of the haystack, we | |
461 | /// // find nothing. | |
781aab86 FG |
462 | /// let input = Input::new(haystack) |
463 | /// .anchored(Anchored::Pattern(PatternID::must(0))); | |
464 | /// assert_eq!(None, dfa.try_search_fwd(&input)?); | |
487cf647 FG |
465 | /// // And finally, an anchored search is not the same as putting a '^' at |
466 | /// // beginning of the pattern. An anchored search can only match at the | |
467 | /// // beginning of the *search*, which we can change: | |
781aab86 FG |
468 | /// let input = Input::new(haystack) |
469 | /// .anchored(Anchored::Pattern(PatternID::must(0))) | |
470 | /// .range(5..); | |
471 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&input)?); | |
487cf647 FG |
472 | /// |
473 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
474 | /// ``` | |
475 | pub fn starts_for_each_pattern(mut self, yes: bool) -> Config { | |
476 | self.starts_for_each_pattern = Some(yes); | |
477 | self | |
478 | } | |
479 | ||
480 | /// Whether to attempt to shrink the size of the DFA's alphabet or not. | |
481 | /// | |
482 | /// This option is enabled by default and should never be disabled unless | |
483 | /// one is debugging a generated DFA. | |
484 | /// | |
485 | /// When enabled, the DFA will use a map from all possible bytes to their | |
486 | /// corresponding equivalence class. Each equivalence class represents a | |
487 | /// set of bytes that does not discriminate between a match and a non-match | |
488 | /// in the DFA. For example, the pattern `[ab]+` has at least two | |
489 | /// equivalence classes: a set containing `a` and `b` and a set containing | |
490 | /// every byte except for `a` and `b`. `a` and `b` are in the same | |
781aab86 FG |
491 | /// equivalence class because they never discriminate between a match and a |
492 | /// non-match. | |
487cf647 FG |
493 | /// |
494 | /// The advantage of this map is that the size of the transition table | |
495 | /// can be reduced drastically from `#states * 256 * sizeof(StateID)` to | |
496 | /// `#states * k * sizeof(StateID)` where `k` is the number of equivalence | |
497 | /// classes (rounded up to the nearest power of 2). As a result, total | |
498 | /// space usage can decrease substantially. Moreover, since a smaller | |
499 | /// alphabet is used, DFA compilation becomes faster as well. | |
500 | /// | |
501 | /// **WARNING:** This is only useful for debugging DFAs. Disabling this | |
502 | /// does not yield any speed advantages. Namely, even when this is | |
503 | /// disabled, a byte class map is still used while searching. The only | |
504 | /// difference is that every byte will be forced into its own distinct | |
505 | /// equivalence class. This is useful for debugging the actual generated | |
506 | /// transitions because it lets one see the transitions defined on actual | |
507 | /// bytes instead of the equivalence classes. | |
508 | pub fn byte_classes(mut self, yes: bool) -> Config { | |
509 | self.byte_classes = Some(yes); | |
510 | self | |
511 | } | |
512 | ||
513 | /// Heuristically enable Unicode word boundaries. | |
514 | /// | |
515 | /// When set, this will attempt to implement Unicode word boundaries as if | |
516 | /// they were ASCII word boundaries. This only works when the search input | |
517 | /// is ASCII only. If a non-ASCII byte is observed while searching, then a | |
781aab86 | 518 | /// [`MatchError::quit`](crate::MatchError::quit) error is returned. |
487cf647 FG |
519 | /// |
520 | /// A possible alternative to enabling this option is to simply use an | |
521 | /// ASCII word boundary, e.g., via `(?-u:\b)`. The main reason to use this | |
522 | /// option is if you absolutely need Unicode support. This option lets one | |
523 | /// use a fast search implementation (a DFA) for some potentially very | |
524 | /// common cases, while providing the option to fall back to some other | |
525 | /// regex engine to handle the general case when an error is returned. | |
526 | /// | |
527 | /// If the pattern provided has no Unicode word boundary in it, then this | |
528 | /// option has no effect. (That is, quitting on a non-ASCII byte only | |
529 | /// occurs when this option is enabled _and_ a Unicode word boundary is | |
530 | /// present in the pattern.) | |
531 | /// | |
532 | /// This is almost equivalent to setting all non-ASCII bytes to be quit | |
533 | /// bytes. The only difference is that this will cause non-ASCII bytes to | |
534 | /// be quit bytes _only_ when a Unicode word boundary is present in the | |
535 | /// pattern. | |
536 | /// | |
537 | /// When enabling this option, callers _must_ be prepared to handle | |
538 | /// a [`MatchError`](crate::MatchError) error during search. | |
539 | /// When using a [`Regex`](crate::dfa::regex::Regex), this corresponds | |
540 | /// to using the `try_` suite of methods. Alternatively, if | |
541 | /// callers can guarantee that their input is ASCII only, then a | |
781aab86 | 542 | /// [`MatchError::quit`](crate::MatchError::quit) error will never be |
487cf647 FG |
543 | /// returned while searching. |
544 | /// | |
545 | /// This is disabled by default. | |
546 | /// | |
547 | /// # Example | |
548 | /// | |
549 | /// This example shows how to heuristically enable Unicode word boundaries | |
550 | /// in a pattern. It also shows what happens when a search comes across a | |
551 | /// non-ASCII byte. | |
552 | /// | |
553 | /// ``` | |
554 | /// use regex_automata::{ | |
555 | /// dfa::{Automaton, dense}, | |
781aab86 | 556 | /// HalfMatch, Input, MatchError, |
487cf647 FG |
557 | /// }; |
558 | /// | |
559 | /// let dfa = dense::Builder::new() | |
560 | /// .configure(dense::Config::new().unicode_word_boundary(true)) | |
561 | /// .build(r"\b[0-9]+\b")?; | |
562 | /// | |
563 | /// // The match occurs before the search ever observes the snowman | |
564 | /// // character, so no error occurs. | |
781aab86 | 565 | /// let haystack = "foo 123 ☃".as_bytes(); |
487cf647 | 566 | /// let expected = Some(HalfMatch::must(0, 7)); |
781aab86 | 567 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
487cf647 FG |
568 | /// assert_eq!(expected, got); |
569 | /// | |
570 | /// // Notice that this search fails, even though the snowman character | |
571 | /// // occurs after the ending match offset. This is because search | |
572 | /// // routines read one byte past the end of the search to account for | |
573 | /// // look-around, and indeed, this is required here to determine whether | |
574 | /// // the trailing \b matches. | |
781aab86 FG |
575 | /// let haystack = "foo 123 ☃".as_bytes(); |
576 | /// let expected = MatchError::quit(0xE2, 8); | |
577 | /// let got = dfa.try_search_fwd(&Input::new(haystack)); | |
578 | /// assert_eq!(Err(expected), got); | |
579 | /// | |
580 | /// // Another example is executing a search where the span of the haystack | |
581 | /// // we specify is all ASCII, but there is non-ASCII just before it. This | |
582 | /// // correctly also reports an error. | |
583 | /// let input = Input::new("β123").range(2..); | |
584 | /// let expected = MatchError::quit(0xB2, 1); | |
585 | /// let got = dfa.try_search_fwd(&input); | |
586 | /// assert_eq!(Err(expected), got); | |
587 | /// | |
588 | /// // And similarly for the trailing word boundary. | |
589 | /// let input = Input::new("123β").range(..3); | |
590 | /// let expected = MatchError::quit(0xCE, 3); | |
591 | /// let got = dfa.try_search_fwd(&input); | |
487cf647 FG |
592 | /// assert_eq!(Err(expected), got); |
593 | /// | |
594 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
595 | /// ``` | |
596 | pub fn unicode_word_boundary(mut self, yes: bool) -> Config { | |
597 | // We have a separate option for this instead of just setting the | |
598 | // appropriate quit bytes here because we don't want to set quit bytes | |
599 | // for every regex. We only want to set them when the regex contains a | |
600 | // Unicode word boundary. | |
601 | self.unicode_word_boundary = Some(yes); | |
602 | self | |
603 | } | |
604 | ||
605 | /// Add a "quit" byte to the DFA. | |
606 | /// | |
607 | /// When a quit byte is seen during search time, then search will return | |
781aab86 | 608 | /// a [`MatchError::quit`](crate::MatchError::quit) error indicating the |
487cf647 FG |
609 | /// offset at which the search stopped. |
610 | /// | |
611 | /// A quit byte will always overrule any other aspects of a regex. For | |
612 | /// example, if the `x` byte is added as a quit byte and the regex `\w` is | |
613 | /// used, then observing `x` will cause the search to quit immediately | |
614 | /// despite the fact that `x` is in the `\w` class. | |
615 | /// | |
616 | /// This mechanism is primarily useful for heuristically enabling certain | |
617 | /// features like Unicode word boundaries in a DFA. Namely, if the input | |
618 | /// to search is ASCII, then a Unicode word boundary can be implemented | |
619 | /// via an ASCII word boundary with no change in semantics. Thus, a DFA | |
620 | /// can attempt to match a Unicode word boundary but give up as soon as it | |
621 | /// observes a non-ASCII byte. Indeed, if callers set all non-ASCII bytes | |
622 | /// to be quit bytes, then Unicode word boundaries will be permitted when | |
623 | /// building DFAs. Of course, callers should enable | |
624 | /// [`Config::unicode_word_boundary`] if they want this behavior instead. | |
625 | /// (The advantage being that non-ASCII quit bytes will only be added if a | |
626 | /// Unicode word boundary is in the pattern.) | |
627 | /// | |
628 | /// When enabling this option, callers _must_ be prepared to handle a | |
629 | /// [`MatchError`](crate::MatchError) error during search. When using a | |
630 | /// [`Regex`](crate::dfa::regex::Regex), this corresponds to using the | |
631 | /// `try_` suite of methods. | |
632 | /// | |
633 | /// By default, there are no quit bytes set. | |
634 | /// | |
635 | /// # Panics | |
636 | /// | |
637 | /// This panics if heuristic Unicode word boundaries are enabled and any | |
638 | /// non-ASCII byte is removed from the set of quit bytes. Namely, enabling | |
639 | /// Unicode word boundaries requires setting every non-ASCII byte to a quit | |
640 | /// byte. So if the caller attempts to undo any of that, then this will | |
641 | /// panic. | |
642 | /// | |
643 | /// # Example | |
644 | /// | |
645 | /// This example shows how to cause a search to terminate if it sees a | |
646 | /// `\n` byte. This could be useful if, for example, you wanted to prevent | |
647 | /// a user supplied pattern from matching across a line boundary. | |
648 | /// | |
649 | /// ``` | |
781aab86 FG |
650 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
651 | /// use regex_automata::{dfa::{Automaton, dense}, Input, MatchError}; | |
487cf647 FG |
652 | /// |
653 | /// let dfa = dense::Builder::new() | |
654 | /// .configure(dense::Config::new().quit(b'\n', true)) | |
655 | /// .build(r"foo\p{any}+bar")?; | |
656 | /// | |
657 | /// let haystack = "foo\nbar".as_bytes(); | |
658 | /// // Normally this would produce a match, since \p{any} contains '\n'. | |
659 | /// // But since we instructed the automaton to enter a quit state if a | |
660 | /// // '\n' is observed, this produces a match error instead. | |
781aab86 FG |
661 | /// let expected = MatchError::quit(b'\n', 3); |
662 | /// let got = dfa.try_search_fwd(&Input::new(haystack)).unwrap_err(); | |
487cf647 FG |
663 | /// assert_eq!(expected, got); |
664 | /// | |
665 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
666 | /// ``` | |
667 | pub fn quit(mut self, byte: u8, yes: bool) -> Config { | |
668 | if self.get_unicode_word_boundary() && !byte.is_ascii() && !yes { | |
669 | panic!( | |
670 | "cannot set non-ASCII byte to be non-quit when \ | |
671 | Unicode word boundaries are enabled" | |
672 | ); | |
673 | } | |
781aab86 FG |
674 | if self.quitset.is_none() { |
675 | self.quitset = Some(ByteSet::empty()); | |
487cf647 FG |
676 | } |
677 | if yes { | |
781aab86 | 678 | self.quitset.as_mut().unwrap().add(byte); |
487cf647 | 679 | } else { |
781aab86 | 680 | self.quitset.as_mut().unwrap().remove(byte); |
487cf647 FG |
681 | } |
682 | self | |
683 | } | |
684 | ||
781aab86 FG |
685 | /// Enable specializing start states in the DFA. |
686 | /// | |
687 | /// When start states are specialized, an implementor of a search routine | |
688 | /// using a lazy DFA can tell when the search has entered a starting state. | |
689 | /// When start states aren't specialized, then it is impossible to know | |
690 | /// whether the search has entered a start state. | |
691 | /// | |
692 | /// Ideally, this option wouldn't need to exist and we could always | |
693 | /// specialize start states. The problem is that start states can be quite | |
694 | /// active. This in turn means that an efficient search routine is likely | |
695 | /// to ping-pong between a heavily optimized hot loop that handles most | |
696 | /// states and to a less optimized specialized handling of start states. | |
697 | /// This causes branches to get heavily mispredicted and overall can | |
698 | /// materially decrease throughput. Therefore, specializing start states | |
699 | /// should only be enabled when it is needed. | |
700 | /// | |
701 | /// Knowing whether a search is in a start state is typically useful when a | |
702 | /// prefilter is active for the search. A prefilter is typically only run | |
703 | /// when in a start state and a prefilter can greatly accelerate a search. | |
704 | /// Therefore, the possible cost of specializing start states is worth it | |
705 | /// in this case. Otherwise, if you have no prefilter, there is likely no | |
706 | /// reason to specialize start states. | |
707 | /// | |
708 | /// This is disabled by default, but note that it is automatically | |
709 | /// enabled (or disabled) if [`Config::prefilter`] is set. Namely, unless | |
710 | /// `specialize_start_states` has already been set, [`Config::prefilter`] | |
711 | /// will automatically enable or disable it based on whether a prefilter | |
712 | /// is present or not, respectively. This is done because a prefilter's | |
713 | /// effectiveness is rooted in being executed whenever the DFA is in a | |
714 | /// start state, and that's only possible to do when they are specialized. | |
715 | /// | |
716 | /// Note that it is plausibly reasonable to _disable_ this option | |
717 | /// explicitly while _enabling_ a prefilter. In that case, a prefilter | |
718 | /// will still be run at the beginning of a search, but never again. This | |
719 | /// in theory could strike a good balance if you're in a situation where a | |
720 | /// prefilter is likely to produce many false positive candidates. | |
721 | /// | |
722 | /// # Example | |
723 | /// | |
724 | /// This example shows how to enable start state specialization and then | |
725 | /// shows how to check whether a state is a start state or not. | |
726 | /// | |
727 | /// ``` | |
728 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, Input}; | |
729 | /// | |
730 | /// let dfa = DFA::builder() | |
731 | /// .configure(DFA::config().specialize_start_states(true)) | |
732 | /// .build(r"[a-z]+")?; | |
733 | /// | |
734 | /// let haystack = "123 foobar 4567".as_bytes(); | |
735 | /// let sid = dfa.start_state_forward(&Input::new(haystack))?; | |
736 | /// // The ID returned by 'start_state_forward' will always be tagged as | |
737 | /// // a start state when start state specialization is enabled. | |
738 | /// assert!(dfa.is_special_state(sid)); | |
739 | /// assert!(dfa.is_start_state(sid)); | |
740 | /// | |
741 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
742 | /// ``` | |
743 | /// | |
744 | /// Compare the above with the default DFA configuration where start states | |
745 | /// are _not_ specialized. In this case, the start state is not tagged at | |
746 | /// all: | |
747 | /// | |
748 | /// ``` | |
749 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, Input}; | |
750 | /// | |
751 | /// let dfa = DFA::new(r"[a-z]+")?; | |
752 | /// | |
753 | /// let haystack = "123 foobar 4567"; | |
754 | /// let sid = dfa.start_state_forward(&Input::new(haystack))?; | |
755 | /// // Start states are not special in the default configuration! | |
756 | /// assert!(!dfa.is_special_state(sid)); | |
757 | /// assert!(!dfa.is_start_state(sid)); | |
758 | /// | |
759 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
760 | /// ``` | |
761 | pub fn specialize_start_states(mut self, yes: bool) -> Config { | |
762 | self.specialize_start_states = Some(yes); | |
763 | self | |
764 | } | |
765 | ||
487cf647 FG |
766 | /// Set a size limit on the total heap used by a DFA. |
767 | /// | |
768 | /// This size limit is expressed in bytes and is applied during | |
769 | /// determinization of an NFA into a DFA. If the DFA's heap usage, and only | |
770 | /// the DFA, exceeds this configured limit, then determinization is stopped | |
771 | /// and an error is returned. | |
772 | /// | |
773 | /// This limit does not apply to auxiliary storage used during | |
774 | /// determinization that isn't part of the generated DFA. | |
775 | /// | |
776 | /// This limit is only applied during determinization. Currently, there is | |
777 | /// no way to post-pone this check to after minimization if minimization | |
778 | /// was enabled. | |
779 | /// | |
780 | /// The total limit on heap used during determinization is the sum of the | |
781 | /// DFA and determinization size limits. | |
782 | /// | |
783 | /// The default is no limit. | |
784 | /// | |
785 | /// # Example | |
786 | /// | |
787 | /// This example shows a DFA that fails to build because of a configured | |
788 | /// size limit. This particular example also serves as a cautionary tale | |
789 | /// demonstrating just how big DFAs with large Unicode character classes | |
790 | /// can get. | |
791 | /// | |
792 | /// ``` | |
781aab86 FG |
793 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
794 | /// use regex_automata::{dfa::{dense, Automaton}, Input}; | |
487cf647 | 795 | /// |
781aab86 | 796 | /// // 6MB isn't enough! |
487cf647 | 797 | /// dense::Builder::new() |
781aab86 | 798 | /// .configure(dense::Config::new().dfa_size_limit(Some(6_000_000))) |
487cf647 FG |
799 | /// .build(r"\w{20}") |
800 | /// .unwrap_err(); | |
801 | /// | |
781aab86 | 802 | /// // ... but 7MB probably is! |
487cf647 FG |
803 | /// // (Note that DFA sizes aren't necessarily stable between releases.) |
804 | /// let dfa = dense::Builder::new() | |
781aab86 | 805 | /// .configure(dense::Config::new().dfa_size_limit(Some(7_000_000))) |
487cf647 FG |
806 | /// .build(r"\w{20}")?; |
807 | /// let haystack = "A".repeat(20).into_bytes(); | |
781aab86 | 808 | /// assert!(dfa.try_search_fwd(&Input::new(&haystack))?.is_some()); |
487cf647 FG |
809 | /// |
810 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
811 | /// ``` | |
812 | /// | |
781aab86 FG |
813 | /// While one needs a little more than 6MB to represent `\w{20}`, it |
814 | /// turns out that you only need a little more than 6KB to represent | |
487cf647 | 815 | /// `(?-u:\w{20})`. So only use Unicode if you need it! |
781aab86 FG |
816 | /// |
817 | /// As with [`Config::determinize_size_limit`], the size of a DFA is | |
818 | /// influenced by other factors, such as what start state configurations | |
819 | /// to support. For example, if you only need unanchored searches and not | |
820 | /// anchored searches, then configuring the DFA to only support unanchored | |
821 | /// searches can reduce its size. By default, DFAs support both unanchored | |
822 | /// and anchored searches. | |
823 | /// | |
824 | /// ``` | |
825 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long | |
826 | /// use regex_automata::{dfa::{dense, Automaton, StartKind}, Input}; | |
827 | /// | |
828 | /// // 3MB isn't enough! | |
829 | /// dense::Builder::new() | |
830 | /// .configure(dense::Config::new() | |
831 | /// .dfa_size_limit(Some(3_000_000)) | |
832 | /// .start_kind(StartKind::Unanchored) | |
833 | /// ) | |
834 | /// .build(r"\w{20}") | |
835 | /// .unwrap_err(); | |
836 | /// | |
837 | /// // ... but 4MB probably is! | |
838 | /// // (Note that DFA sizes aren't necessarily stable between releases.) | |
839 | /// let dfa = dense::Builder::new() | |
840 | /// .configure(dense::Config::new() | |
841 | /// .dfa_size_limit(Some(4_000_000)) | |
842 | /// .start_kind(StartKind::Unanchored) | |
843 | /// ) | |
844 | /// .build(r"\w{20}")?; | |
845 | /// let haystack = "A".repeat(20).into_bytes(); | |
846 | /// assert!(dfa.try_search_fwd(&Input::new(&haystack))?.is_some()); | |
847 | /// | |
848 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
849 | /// ``` | |
487cf647 FG |
850 | pub fn dfa_size_limit(mut self, bytes: Option<usize>) -> Config { |
851 | self.dfa_size_limit = Some(bytes); | |
852 | self | |
853 | } | |
854 | ||
855 | /// Set a size limit on the total heap used by determinization. | |
856 | /// | |
857 | /// This size limit is expressed in bytes and is applied during | |
858 | /// determinization of an NFA into a DFA. If the heap used for auxiliary | |
859 | /// storage during determinization (memory that is not in the DFA but | |
860 | /// necessary for building the DFA) exceeds this configured limit, then | |
861 | /// determinization is stopped and an error is returned. | |
862 | /// | |
863 | /// This limit does not apply to heap used by the DFA itself. | |
864 | /// | |
865 | /// The total limit on heap used during determinization is the sum of the | |
866 | /// DFA and determinization size limits. | |
867 | /// | |
868 | /// The default is no limit. | |
869 | /// | |
870 | /// # Example | |
871 | /// | |
872 | /// This example shows a DFA that fails to build because of a | |
873 | /// configured size limit on the amount of heap space used by | |
874 | /// determinization. This particular example complements the example for | |
875 | /// [`Config::dfa_size_limit`] by demonstrating that not only does Unicode | |
876 | /// potentially make DFAs themselves big, but it also results in more | |
877 | /// auxiliary storage during determinization. (Although, auxiliary storage | |
878 | /// is still not as much as the DFA itself.) | |
879 | /// | |
880 | /// ``` | |
781aab86 FG |
881 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
882 | /// # if !cfg!(target_pointer_width = "64") { return Ok(()); } // see #1039 | |
883 | /// use regex_automata::{dfa::{dense, Automaton}, Input}; | |
487cf647 | 884 | /// |
781aab86 | 885 | /// // 600KB isn't enough! |
487cf647 FG |
886 | /// dense::Builder::new() |
887 | /// .configure(dense::Config::new() | |
781aab86 FG |
888 | /// .determinize_size_limit(Some(600_000)) |
889 | /// ) | |
890 | /// .build(r"\w{20}") | |
891 | /// .unwrap_err(); | |
892 | /// | |
893 | /// // ... but 700KB probably is! | |
894 | /// // (Note that auxiliary storage sizes aren't necessarily stable between | |
895 | /// // releases.) | |
896 | /// let dfa = dense::Builder::new() | |
897 | /// .configure(dense::Config::new() | |
898 | /// .determinize_size_limit(Some(700_000)) | |
899 | /// ) | |
900 | /// .build(r"\w{20}")?; | |
901 | /// let haystack = "A".repeat(20).into_bytes(); | |
902 | /// assert!(dfa.try_search_fwd(&Input::new(&haystack))?.is_some()); | |
903 | /// | |
904 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
905 | /// ``` | |
906 | /// | |
907 | /// Note that some parts of the configuration on a DFA can have a | |
908 | /// big impact on how big the DFA is, and thus, how much memory is | |
909 | /// used. For example, the default setting for [`Config::start_kind`] is | |
910 | /// [`StartKind::Both`]. But if you only need an anchored search, for | |
911 | /// example, then it can be much cheaper to build a DFA that only supports | |
912 | /// anchored searches. (Running an unanchored search with it would panic.) | |
913 | /// | |
914 | /// ``` | |
915 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long | |
916 | /// # if !cfg!(target_pointer_width = "64") { return Ok(()); } // see #1039 | |
917 | /// use regex_automata::{ | |
918 | /// dfa::{dense, Automaton, StartKind}, | |
919 | /// Anchored, Input, | |
920 | /// }; | |
921 | /// | |
922 | /// // 200KB isn't enough! | |
923 | /// dense::Builder::new() | |
924 | /// .configure(dense::Config::new() | |
925 | /// .determinize_size_limit(Some(200_000)) | |
926 | /// .start_kind(StartKind::Anchored) | |
487cf647 FG |
927 | /// ) |
928 | /// .build(r"\w{20}") | |
929 | /// .unwrap_err(); | |
930 | /// | |
781aab86 | 931 | /// // ... but 300KB probably is! |
487cf647 FG |
932 | /// // (Note that auxiliary storage sizes aren't necessarily stable between |
933 | /// // releases.) | |
934 | /// let dfa = dense::Builder::new() | |
935 | /// .configure(dense::Config::new() | |
781aab86 FG |
936 | /// .determinize_size_limit(Some(300_000)) |
937 | /// .start_kind(StartKind::Anchored) | |
487cf647 FG |
938 | /// ) |
939 | /// .build(r"\w{20}")?; | |
940 | /// let haystack = "A".repeat(20).into_bytes(); | |
781aab86 FG |
941 | /// let input = Input::new(&haystack).anchored(Anchored::Yes); |
942 | /// assert!(dfa.try_search_fwd(&input)?.is_some()); | |
487cf647 FG |
943 | /// |
944 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
945 | /// ``` | |
946 | pub fn determinize_size_limit(mut self, bytes: Option<usize>) -> Config { | |
947 | self.determinize_size_limit = Some(bytes); | |
948 | self | |
949 | } | |
950 | ||
487cf647 FG |
951 | /// Returns whether this configuration has enabled simple state |
952 | /// acceleration. | |
953 | pub fn get_accelerate(&self) -> bool { | |
954 | self.accelerate.unwrap_or(true) | |
955 | } | |
956 | ||
781aab86 FG |
957 | /// Returns the prefilter attached to this configuration, if any. |
958 | pub fn get_prefilter(&self) -> Option<&Prefilter> { | |
959 | self.pre.as_ref().unwrap_or(&None).as_ref() | |
960 | } | |
961 | ||
487cf647 FG |
962 | /// Returns whether this configuration has enabled the expensive process |
963 | /// of minimizing a DFA. | |
964 | pub fn get_minimize(&self) -> bool { | |
965 | self.minimize.unwrap_or(false) | |
966 | } | |
967 | ||
968 | /// Returns the match semantics set in this configuration. | |
969 | pub fn get_match_kind(&self) -> MatchKind { | |
970 | self.match_kind.unwrap_or(MatchKind::LeftmostFirst) | |
971 | } | |
972 | ||
781aab86 FG |
973 | /// Returns the starting state configuration for a DFA. |
974 | pub fn get_starts(&self) -> StartKind { | |
975 | self.start_kind.unwrap_or(StartKind::Both) | |
976 | } | |
977 | ||
487cf647 FG |
978 | /// Returns whether this configuration has enabled anchored starting states |
979 | /// for every pattern in the DFA. | |
980 | pub fn get_starts_for_each_pattern(&self) -> bool { | |
981 | self.starts_for_each_pattern.unwrap_or(false) | |
982 | } | |
983 | ||
984 | /// Returns whether this configuration has enabled byte classes or not. | |
985 | /// This is typically a debugging oriented option, as disabling it confers | |
986 | /// no speed benefit. | |
987 | pub fn get_byte_classes(&self) -> bool { | |
988 | self.byte_classes.unwrap_or(true) | |
989 | } | |
990 | ||
991 | /// Returns whether this configuration has enabled heuristic Unicode word | |
992 | /// boundary support. When enabled, it is possible for a search to return | |
993 | /// an error. | |
994 | pub fn get_unicode_word_boundary(&self) -> bool { | |
995 | self.unicode_word_boundary.unwrap_or(false) | |
996 | } | |
997 | ||
998 | /// Returns whether this configuration will instruct the DFA to enter a | |
999 | /// quit state whenever the given byte is seen during a search. When at | |
1000 | /// least one byte has this enabled, it is possible for a search to return | |
1001 | /// an error. | |
1002 | pub fn get_quit(&self, byte: u8) -> bool { | |
781aab86 FG |
1003 | self.quitset.map_or(false, |q| q.contains(byte)) |
1004 | } | |
1005 | ||
1006 | /// Returns whether this configuration will instruct the DFA to | |
1007 | /// "specialize" start states. When enabled, the DFA will mark start states | |
1008 | /// as "special" so that search routines using the DFA can detect when | |
1009 | /// it's in a start state and do some kind of optimization (like run a | |
1010 | /// prefilter). | |
1011 | pub fn get_specialize_start_states(&self) -> bool { | |
1012 | self.specialize_start_states.unwrap_or(false) | |
487cf647 FG |
1013 | } |
1014 | ||
1015 | /// Returns the DFA size limit of this configuration if one was set. | |
1016 | /// The size limit is total number of bytes on the heap that a DFA is | |
1017 | /// permitted to use. If the DFA exceeds this limit during construction, | |
1018 | /// then construction is stopped and an error is returned. | |
1019 | pub fn get_dfa_size_limit(&self) -> Option<usize> { | |
1020 | self.dfa_size_limit.unwrap_or(None) | |
1021 | } | |
1022 | ||
1023 | /// Returns the determinization size limit of this configuration if one | |
1024 | /// was set. The size limit is total number of bytes on the heap that | |
1025 | /// determinization is permitted to use. If determinization exceeds this | |
1026 | /// limit during construction, then construction is stopped and an error is | |
1027 | /// returned. | |
1028 | /// | |
1029 | /// This is different from the DFA size limit in that this only applies to | |
1030 | /// the auxiliary storage used during determinization. Once determinization | |
1031 | /// is complete, this memory is freed. | |
1032 | /// | |
1033 | /// The limit on the total heap memory used is the sum of the DFA and | |
1034 | /// determinization size limits. | |
1035 | pub fn get_determinize_size_limit(&self) -> Option<usize> { | |
1036 | self.determinize_size_limit.unwrap_or(None) | |
1037 | } | |
1038 | ||
1039 | /// Overwrite the default configuration such that the options in `o` are | |
1040 | /// always used. If an option in `o` is not set, then the corresponding | |
1041 | /// option in `self` is used. If it's not set in `self` either, then it | |
1042 | /// remains not set. | |
781aab86 | 1043 | pub(crate) fn overwrite(&self, o: Config) -> Config { |
487cf647 | 1044 | Config { |
487cf647 | 1045 | accelerate: o.accelerate.or(self.accelerate), |
781aab86 | 1046 | pre: o.pre.or_else(|| self.pre.clone()), |
487cf647 FG |
1047 | minimize: o.minimize.or(self.minimize), |
1048 | match_kind: o.match_kind.or(self.match_kind), | |
781aab86 | 1049 | start_kind: o.start_kind.or(self.start_kind), |
487cf647 FG |
1050 | starts_for_each_pattern: o |
1051 | .starts_for_each_pattern | |
1052 | .or(self.starts_for_each_pattern), | |
1053 | byte_classes: o.byte_classes.or(self.byte_classes), | |
1054 | unicode_word_boundary: o | |
1055 | .unicode_word_boundary | |
1056 | .or(self.unicode_word_boundary), | |
781aab86 FG |
1057 | quitset: o.quitset.or(self.quitset), |
1058 | specialize_start_states: o | |
1059 | .specialize_start_states | |
1060 | .or(self.specialize_start_states), | |
487cf647 FG |
1061 | dfa_size_limit: o.dfa_size_limit.or(self.dfa_size_limit), |
1062 | determinize_size_limit: o | |
1063 | .determinize_size_limit | |
1064 | .or(self.determinize_size_limit), | |
1065 | } | |
1066 | } | |
1067 | } | |
1068 | ||
1069 | /// A builder for constructing a deterministic finite automaton from regular | |
1070 | /// expressions. | |
1071 | /// | |
1072 | /// This builder provides two main things: | |
1073 | /// | |
1074 | /// 1. It provides a few different `build` routines for actually constructing | |
1075 | /// a DFA from different kinds of inputs. The most convenient is | |
1076 | /// [`Builder::build`], which builds a DFA directly from a pattern string. The | |
1077 | /// most flexible is [`Builder::build_from_nfa`], which builds a DFA straight | |
1078 | /// from an NFA. | |
1079 | /// 2. The builder permits configuring a number of things. | |
1080 | /// [`Builder::configure`] is used with [`Config`] to configure aspects of | |
1081 | /// the DFA and the construction process itself. [`Builder::syntax`] and | |
1082 | /// [`Builder::thompson`] permit configuring the regex parser and Thompson NFA | |
1083 | /// construction, respectively. The syntax and thompson configurations only | |
1084 | /// apply when building from a pattern string. | |
1085 | /// | |
1086 | /// This builder always constructs a *single* DFA. As such, this builder | |
1087 | /// can only be used to construct regexes that either detect the presence | |
1088 | /// of a match or find the end location of a match. A single DFA cannot | |
1089 | /// produce both the start and end of a match. For that information, use a | |
1090 | /// [`Regex`](crate::dfa::regex::Regex), which can be similarly configured | |
1091 | /// using [`regex::Builder`](crate::dfa::regex::Builder). The main reason to | |
1092 | /// use a DFA directly is if the end location of a match is enough for your use | |
1093 | /// case. Namely, a `Regex` will construct two DFAs instead of one, since a | |
1094 | /// second reverse DFA is needed to find the start of a match. | |
1095 | /// | |
1096 | /// Note that if one wants to build a sparse DFA, you must first build a dense | |
1097 | /// DFA and convert that to a sparse DFA. There is no way to build a sparse | |
1098 | /// DFA without first building a dense DFA. | |
1099 | /// | |
1100 | /// # Example | |
1101 | /// | |
1102 | /// This example shows how to build a minimized DFA that completely disables | |
1103 | /// Unicode. That is: | |
1104 | /// | |
1105 | /// * Things such as `\w`, `.` and `\b` are no longer Unicode-aware. `\w` | |
1106 | /// and `\b` are ASCII-only while `.` matches any byte except for `\n` | |
1107 | /// (instead of any UTF-8 encoding of a Unicode scalar value except for | |
1108 | /// `\n`). Things that are Unicode only, such as `\pL`, are not allowed. | |
1109 | /// * The pattern itself is permitted to match invalid UTF-8. For example, | |
1110 | /// things like `[^a]` that match any byte except for `a` are permitted. | |
487cf647 FG |
1111 | /// |
1112 | /// ``` | |
1113 | /// use regex_automata::{ | |
1114 | /// dfa::{Automaton, dense}, | |
781aab86 FG |
1115 | /// util::syntax, |
1116 | /// HalfMatch, Input, | |
487cf647 FG |
1117 | /// }; |
1118 | /// | |
1119 | /// let dfa = dense::Builder::new() | |
1120 | /// .configure(dense::Config::new().minimize(false)) | |
781aab86 | 1121 | /// .syntax(syntax::Config::new().unicode(false).utf8(false)) |
487cf647 FG |
1122 | /// .build(r"foo[^b]ar.*")?; |
1123 | /// | |
1124 | /// let haystack = b"\xFEfoo\xFFar\xE2\x98\xFF\n"; | |
1125 | /// let expected = Some(HalfMatch::must(0, 10)); | |
781aab86 | 1126 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
487cf647 FG |
1127 | /// assert_eq!(expected, got); |
1128 | /// | |
1129 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
1130 | /// ``` | |
781aab86 | 1131 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1132 | #[derive(Clone, Debug)] |
1133 | pub struct Builder { | |
1134 | config: Config, | |
781aab86 FG |
1135 | #[cfg(feature = "syntax")] |
1136 | thompson: thompson::Compiler, | |
487cf647 FG |
1137 | } |
1138 | ||
781aab86 | 1139 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1140 | impl Builder { |
1141 | /// Create a new dense DFA builder with the default configuration. | |
1142 | pub fn new() -> Builder { | |
1143 | Builder { | |
1144 | config: Config::default(), | |
781aab86 FG |
1145 | #[cfg(feature = "syntax")] |
1146 | thompson: thompson::Compiler::new(), | |
487cf647 FG |
1147 | } |
1148 | } | |
1149 | ||
1150 | /// Build a DFA from the given pattern. | |
1151 | /// | |
1152 | /// If there was a problem parsing or compiling the pattern, then an error | |
1153 | /// is returned. | |
781aab86 FG |
1154 | #[cfg(feature = "syntax")] |
1155 | pub fn build(&self, pattern: &str) -> Result<OwnedDFA, BuildError> { | |
487cf647 FG |
1156 | self.build_many(&[pattern]) |
1157 | } | |
1158 | ||
1159 | /// Build a DFA from the given patterns. | |
1160 | /// | |
1161 | /// When matches are returned, the pattern ID corresponds to the index of | |
1162 | /// the pattern in the slice given. | |
781aab86 | 1163 | #[cfg(feature = "syntax")] |
487cf647 FG |
1164 | pub fn build_many<P: AsRef<str>>( |
1165 | &self, | |
1166 | patterns: &[P], | |
781aab86 FG |
1167 | ) -> Result<OwnedDFA, BuildError> { |
1168 | let nfa = self | |
1169 | .thompson | |
1170 | .clone() | |
1171 | // We can always forcefully disable captures because DFAs do not | |
1172 | // support them. | |
1173 | .configure( | |
1174 | thompson::Config::new() | |
1175 | .which_captures(thompson::WhichCaptures::None), | |
1176 | ) | |
1177 | .build_many(patterns) | |
1178 | .map_err(BuildError::nfa)?; | |
487cf647 FG |
1179 | self.build_from_nfa(&nfa) |
1180 | } | |
1181 | ||
1182 | /// Build a DFA from the given NFA. | |
1183 | /// | |
1184 | /// # Example | |
1185 | /// | |
1186 | /// This example shows how to build a DFA if you already have an NFA in | |
1187 | /// hand. | |
1188 | /// | |
1189 | /// ``` | |
1190 | /// use regex_automata::{ | |
1191 | /// dfa::{Automaton, dense}, | |
781aab86 FG |
1192 | /// nfa::thompson::NFA, |
1193 | /// HalfMatch, Input, | |
487cf647 FG |
1194 | /// }; |
1195 | /// | |
1196 | /// let haystack = "foo123bar".as_bytes(); | |
1197 | /// | |
1198 | /// // This shows how to set non-default options for building an NFA. | |
781aab86 FG |
1199 | /// let nfa = NFA::compiler() |
1200 | /// .configure(NFA::config().shrink(true)) | |
487cf647 FG |
1201 | /// .build(r"[0-9]+")?; |
1202 | /// let dfa = dense::Builder::new().build_from_nfa(&nfa)?; | |
1203 | /// let expected = Some(HalfMatch::must(0, 6)); | |
781aab86 | 1204 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
487cf647 FG |
1205 | /// assert_eq!(expected, got); |
1206 | /// | |
1207 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
1208 | /// ``` | |
1209 | pub fn build_from_nfa( | |
1210 | &self, | |
1211 | nfa: &thompson::NFA, | |
781aab86 FG |
1212 | ) -> Result<OwnedDFA, BuildError> { |
1213 | let mut quitset = self.config.quitset.unwrap_or(ByteSet::empty()); | |
487cf647 | 1214 | if self.config.get_unicode_word_boundary() |
781aab86 | 1215 | && nfa.look_set_any().contains_word_unicode() |
487cf647 FG |
1216 | { |
1217 | for b in 0x80..=0xFF { | |
781aab86 | 1218 | quitset.add(b); |
487cf647 FG |
1219 | } |
1220 | } | |
1221 | let classes = if !self.config.get_byte_classes() { | |
1222 | // DFAs will always use the equivalence class map, but enabling | |
1223 | // this option is useful for debugging. Namely, this will cause all | |
1224 | // transitions to be defined over their actual bytes instead of an | |
1225 | // opaque equivalence class identifier. The former is much easier | |
1226 | // to grok as a human. | |
1227 | ByteClasses::singletons() | |
1228 | } else { | |
1229 | let mut set = nfa.byte_class_set().clone(); | |
1230 | // It is important to distinguish any "quit" bytes from all other | |
1231 | // bytes. Otherwise, a non-quit byte may end up in the same class | |
1232 | // as a quit byte, and thus cause the DFA stop when it shouldn't. | |
781aab86 FG |
1233 | // |
1234 | // Test case: | |
1235 | // | |
1236 | // regex-cli find hybrid regex -w @conn.json.1000x.log \ | |
1237 | // '^#' '\b10\.55\.182\.100\b' | |
1238 | if !quitset.is_empty() { | |
1239 | set.add_set(&quitset); | |
487cf647 FG |
1240 | } |
1241 | set.byte_classes() | |
1242 | }; | |
1243 | ||
1244 | let mut dfa = DFA::initial( | |
1245 | classes, | |
1246 | nfa.pattern_len(), | |
781aab86 FG |
1247 | self.config.get_starts(), |
1248 | nfa.look_matcher(), | |
487cf647 | 1249 | self.config.get_starts_for_each_pattern(), |
781aab86 FG |
1250 | self.config.get_prefilter().map(|p| p.clone()), |
1251 | quitset, | |
1252 | Flags::from_nfa(&nfa), | |
487cf647 FG |
1253 | )?; |
1254 | determinize::Config::new() | |
487cf647 | 1255 | .match_kind(self.config.get_match_kind()) |
781aab86 | 1256 | .quit(quitset) |
487cf647 FG |
1257 | .dfa_size_limit(self.config.get_dfa_size_limit()) |
1258 | .determinize_size_limit(self.config.get_determinize_size_limit()) | |
1259 | .run(nfa, &mut dfa)?; | |
1260 | if self.config.get_minimize() { | |
1261 | dfa.minimize(); | |
1262 | } | |
1263 | if self.config.get_accelerate() { | |
1264 | dfa.accelerate(); | |
1265 | } | |
781aab86 FG |
1266 | // The state shuffling done before this point always assumes that start |
1267 | // states should be marked as "special," even though it isn't the | |
1268 | // default configuration. State shuffling is complex enough as it is, | |
1269 | // so it's simpler to just "fix" our special state ID ranges to not | |
1270 | // include starting states after-the-fact. | |
1271 | if !self.config.get_specialize_start_states() { | |
1272 | dfa.special.set_no_special_start_states(); | |
1273 | } | |
1274 | // Look for and set the universal starting states. | |
1275 | dfa.set_universal_starts(); | |
487cf647 FG |
1276 | Ok(dfa) |
1277 | } | |
1278 | ||
1279 | /// Apply the given dense DFA configuration options to this builder. | |
1280 | pub fn configure(&mut self, config: Config) -> &mut Builder { | |
1281 | self.config = self.config.overwrite(config); | |
1282 | self | |
1283 | } | |
1284 | ||
1285 | /// Set the syntax configuration for this builder using | |
781aab86 | 1286 | /// [`syntax::Config`](crate::util::syntax::Config). |
487cf647 FG |
1287 | /// |
1288 | /// This permits setting things like case insensitivity, Unicode and multi | |
1289 | /// line mode. | |
1290 | /// | |
1291 | /// These settings only apply when constructing a DFA directly from a | |
1292 | /// pattern. | |
781aab86 | 1293 | #[cfg(feature = "syntax")] |
487cf647 FG |
1294 | pub fn syntax( |
1295 | &mut self, | |
781aab86 | 1296 | config: crate::util::syntax::Config, |
487cf647 FG |
1297 | ) -> &mut Builder { |
1298 | self.thompson.syntax(config); | |
1299 | self | |
1300 | } | |
1301 | ||
1302 | /// Set the Thompson NFA configuration for this builder using | |
1303 | /// [`nfa::thompson::Config`](crate::nfa::thompson::Config). | |
1304 | /// | |
1305 | /// This permits setting things like whether the DFA should match the regex | |
1306 | /// in reverse or if additional time should be spent shrinking the size of | |
1307 | /// the NFA. | |
1308 | /// | |
1309 | /// These settings only apply when constructing a DFA directly from a | |
1310 | /// pattern. | |
781aab86 | 1311 | #[cfg(feature = "syntax")] |
487cf647 FG |
1312 | pub fn thompson(&mut self, config: thompson::Config) -> &mut Builder { |
1313 | self.thompson.configure(config); | |
1314 | self | |
1315 | } | |
1316 | } | |
1317 | ||
781aab86 | 1318 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1319 | impl Default for Builder { |
1320 | fn default() -> Builder { | |
1321 | Builder::new() | |
1322 | } | |
1323 | } | |
1324 | ||
1325 | /// A convenience alias for an owned DFA. We use this particular instantiation | |
1326 | /// a lot in this crate, so it's worth giving it a name. This instantiation | |
1327 | /// is commonly used for mutable APIs on the DFA while building it. The main | |
1328 | /// reason for making DFAs generic is no_std support, and more generally, | |
1329 | /// making it possible to load a DFA from an arbitrary slice of bytes. | |
1330 | #[cfg(feature = "alloc")] | |
781aab86 | 1331 | pub(crate) type OwnedDFA = DFA<alloc::vec::Vec<u32>>; |
487cf647 FG |
1332 | |
1333 | /// A dense table-based deterministic finite automaton (DFA). | |
1334 | /// | |
1335 | /// All dense DFAs have one or more start states, zero or more match states | |
1336 | /// and a transition table that maps the current state and the current byte | |
1337 | /// of input to the next state. A DFA can use this information to implement | |
1338 | /// fast searching. In particular, the use of a dense DFA generally makes the | |
1339 | /// trade off that match speed is the most valuable characteristic, even if | |
1340 | /// building the DFA may take significant time *and* space. (More concretely, | |
1341 | /// building a DFA takes time and space that is exponential in the size of the | |
1342 | /// pattern in the worst case.) As such, the processing of every byte of input | |
1343 | /// is done with a small constant number of operations that does not vary with | |
1344 | /// the pattern, its size or the size of the alphabet. If your needs don't line | |
1345 | /// up with this trade off, then a dense DFA may not be an adequate solution to | |
1346 | /// your problem. | |
1347 | /// | |
1348 | /// In contrast, a [`sparse::DFA`] makes the opposite | |
1349 | /// trade off: it uses less space but will execute a variable number of | |
1350 | /// instructions per byte at match time, which makes it slower for matching. | |
1351 | /// (Note that space usage is still exponential in the size of the pattern in | |
1352 | /// the worst case.) | |
1353 | /// | |
1354 | /// A DFA can be built using the default configuration via the | |
1355 | /// [`DFA::new`] constructor. Otherwise, one can | |
1356 | /// configure various aspects via [`dense::Builder`](Builder). | |
1357 | /// | |
1358 | /// A single DFA fundamentally supports the following operations: | |
1359 | /// | |
1360 | /// 1. Detection of a match. | |
1361 | /// 2. Location of the end of a match. | |
1362 | /// 3. In the case of a DFA with multiple patterns, which pattern matched is | |
1363 | /// reported as well. | |
1364 | /// | |
1365 | /// A notable absence from the above list of capabilities is the location of | |
1366 | /// the *start* of a match. In order to provide both the start and end of | |
1367 | /// a match, *two* DFAs are required. This functionality is provided by a | |
1368 | /// [`Regex`](crate::dfa::regex::Regex). | |
1369 | /// | |
1370 | /// # Type parameters | |
1371 | /// | |
1372 | /// A `DFA` has one type parameter, `T`, which is used to represent state IDs, | |
1373 | /// pattern IDs and accelerators. `T` is typically a `Vec<u32>` or a `&[u32]`. | |
1374 | /// | |
1375 | /// # The `Automaton` trait | |
1376 | /// | |
1377 | /// This type implements the [`Automaton`] trait, which means it can be used | |
1378 | /// for searching. For example: | |
1379 | /// | |
1380 | /// ``` | |
781aab86 | 1381 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1382 | /// |
1383 | /// let dfa = DFA::new("foo[0-9]+")?; | |
1384 | /// let expected = HalfMatch::must(0, 8); | |
781aab86 | 1385 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&Input::new("foo12345"))?); |
487cf647 FG |
1386 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1387 | /// ``` | |
1388 | #[derive(Clone)] | |
1389 | pub struct DFA<T> { | |
1390 | /// The transition table for this DFA. This includes the transitions | |
1391 | /// themselves, along with the stride, number of states and the equivalence | |
1392 | /// class mapping. | |
1393 | tt: TransitionTable<T>, | |
1394 | /// The set of starting state identifiers for this DFA. The starting state | |
1395 | /// IDs act as pointers into the transition table. The specific starting | |
1396 | /// state chosen for each search is dependent on the context at which the | |
1397 | /// search begins. | |
1398 | st: StartTable<T>, | |
1399 | /// The set of match states and the patterns that match for each | |
1400 | /// corresponding match state. | |
1401 | /// | |
1402 | /// This structure is technically only needed because of support for | |
1403 | /// multi-regexes. Namely, multi-regexes require answering not just whether | |
1404 | /// a match exists, but _which_ patterns match. So we need to store the | |
1405 | /// matching pattern IDs for each match state. We do this even when there | |
1406 | /// is only one pattern for the sake of simplicity. In practice, this uses | |
781aab86 | 1407 | /// up very little space for the case of one pattern. |
487cf647 FG |
1408 | ms: MatchStates<T>, |
1409 | /// Information about which states are "special." Special states are states | |
1410 | /// that are dead, quit, matching, starting or accelerated. For more info, | |
1411 | /// see the docs for `Special`. | |
1412 | special: Special, | |
1413 | /// The accelerators for this DFA. | |
1414 | /// | |
1415 | /// If a state is accelerated, then there exist only a small number of | |
1416 | /// bytes that can cause the DFA to leave the state. This permits searching | |
1417 | /// to use optimized routines to find those specific bytes instead of using | |
1418 | /// the transition table. | |
1419 | /// | |
1420 | /// All accelerated states exist in a contiguous range in the DFA's | |
1421 | /// transition table. See dfa/special.rs for more details on how states are | |
1422 | /// arranged. | |
1423 | accels: Accels<T>, | |
781aab86 FG |
1424 | /// Any prefilter attached to this DFA. |
1425 | /// | |
1426 | /// Note that currently prefilters are not serialized. When deserializing | |
1427 | /// a DFA from bytes, this is always set to `None`. | |
1428 | pre: Option<Prefilter>, | |
1429 | /// The set of "quit" bytes for this DFA. | |
1430 | /// | |
1431 | /// This is only used when computing the start state for a particular | |
1432 | /// position in a haystack. Namely, in the case where there is a quit | |
1433 | /// byte immediately before the start of the search, this set needs to be | |
1434 | /// explicitly consulted. In all other cases, quit bytes are detected by | |
1435 | /// the DFA itself, by transitioning all quit bytes to a special "quit | |
1436 | /// state." | |
1437 | quitset: ByteSet, | |
1438 | /// Various flags describing the behavior of this DFA. | |
1439 | flags: Flags, | |
487cf647 FG |
1440 | } |
1441 | ||
781aab86 | 1442 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1443 | impl OwnedDFA { |
1444 | /// Parse the given regular expression using a default configuration and | |
1445 | /// return the corresponding DFA. | |
1446 | /// | |
1447 | /// If you want a non-default configuration, then use the | |
1448 | /// [`dense::Builder`](Builder) to set your own configuration. | |
1449 | /// | |
1450 | /// # Example | |
1451 | /// | |
1452 | /// ``` | |
781aab86 | 1453 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
487cf647 FG |
1454 | /// |
1455 | /// let dfa = dense::DFA::new("foo[0-9]+bar")?; | |
781aab86 FG |
1456 | /// let expected = Some(HalfMatch::must(0, 11)); |
1457 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345bar"))?); | |
487cf647 FG |
1458 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1459 | /// ``` | |
781aab86 FG |
1460 | #[cfg(feature = "syntax")] |
1461 | pub fn new(pattern: &str) -> Result<OwnedDFA, BuildError> { | |
487cf647 FG |
1462 | Builder::new().build(pattern) |
1463 | } | |
1464 | ||
1465 | /// Parse the given regular expressions using a default configuration and | |
1466 | /// return the corresponding multi-DFA. | |
1467 | /// | |
1468 | /// If you want a non-default configuration, then use the | |
1469 | /// [`dense::Builder`](Builder) to set your own configuration. | |
1470 | /// | |
1471 | /// # Example | |
1472 | /// | |
1473 | /// ``` | |
781aab86 | 1474 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
487cf647 FG |
1475 | /// |
1476 | /// let dfa = dense::DFA::new_many(&["[0-9]+", "[a-z]+"])?; | |
781aab86 FG |
1477 | /// let expected = Some(HalfMatch::must(1, 3)); |
1478 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345bar"))?); | |
487cf647 FG |
1479 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1480 | /// ``` | |
781aab86 FG |
1481 | #[cfg(feature = "syntax")] |
1482 | pub fn new_many<P: AsRef<str>>( | |
1483 | patterns: &[P], | |
1484 | ) -> Result<OwnedDFA, BuildError> { | |
487cf647 FG |
1485 | Builder::new().build_many(patterns) |
1486 | } | |
1487 | } | |
1488 | ||
781aab86 | 1489 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1490 | impl OwnedDFA { |
1491 | /// Create a new DFA that matches every input. | |
1492 | /// | |
1493 | /// # Example | |
1494 | /// | |
1495 | /// ``` | |
781aab86 | 1496 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
487cf647 FG |
1497 | /// |
1498 | /// let dfa = dense::DFA::always_match()?; | |
1499 | /// | |
781aab86 FG |
1500 | /// let expected = Some(HalfMatch::must(0, 0)); |
1501 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new(""))?); | |
1502 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo"))?); | |
487cf647 FG |
1503 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1504 | /// ``` | |
781aab86 | 1505 | pub fn always_match() -> Result<OwnedDFA, BuildError> { |
487cf647 FG |
1506 | let nfa = thompson::NFA::always_match(); |
1507 | Builder::new().build_from_nfa(&nfa) | |
1508 | } | |
1509 | ||
1510 | /// Create a new DFA that never matches any input. | |
1511 | /// | |
1512 | /// # Example | |
1513 | /// | |
1514 | /// ``` | |
781aab86 | 1515 | /// use regex_automata::{dfa::{Automaton, dense}, Input}; |
487cf647 FG |
1516 | /// |
1517 | /// let dfa = dense::DFA::never_match()?; | |
781aab86 FG |
1518 | /// assert_eq!(None, dfa.try_search_fwd(&Input::new(""))?); |
1519 | /// assert_eq!(None, dfa.try_search_fwd(&Input::new("foo"))?); | |
487cf647 FG |
1520 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1521 | /// ``` | |
781aab86 | 1522 | pub fn never_match() -> Result<OwnedDFA, BuildError> { |
487cf647 FG |
1523 | let nfa = thompson::NFA::never_match(); |
1524 | Builder::new().build_from_nfa(&nfa) | |
1525 | } | |
1526 | ||
781aab86 FG |
1527 | /// Create an initial DFA with the given equivalence classes, pattern |
1528 | /// length and whether anchored starting states are enabled for each | |
1529 | /// pattern. An initial DFA can be further mutated via determinization. | |
487cf647 FG |
1530 | fn initial( |
1531 | classes: ByteClasses, | |
781aab86 FG |
1532 | pattern_len: usize, |
1533 | starts: StartKind, | |
1534 | lookm: &LookMatcher, | |
487cf647 | 1535 | starts_for_each_pattern: bool, |
781aab86 FG |
1536 | pre: Option<Prefilter>, |
1537 | quitset: ByteSet, | |
1538 | flags: Flags, | |
1539 | ) -> Result<OwnedDFA, BuildError> { | |
1540 | let start_pattern_len = | |
1541 | if starts_for_each_pattern { Some(pattern_len) } else { None }; | |
487cf647 FG |
1542 | Ok(DFA { |
1543 | tt: TransitionTable::minimal(classes), | |
781aab86 FG |
1544 | st: StartTable::dead(starts, lookm, start_pattern_len)?, |
1545 | ms: MatchStates::empty(pattern_len), | |
487cf647 FG |
1546 | special: Special::new(), |
1547 | accels: Accels::empty(), | |
781aab86 FG |
1548 | pre, |
1549 | quitset, | |
1550 | flags, | |
487cf647 FG |
1551 | }) |
1552 | } | |
1553 | } | |
1554 | ||
781aab86 FG |
1555 | #[cfg(feature = "dfa-build")] |
1556 | impl DFA<&[u32]> { | |
1557 | /// Return a new default dense DFA compiler configuration. | |
1558 | /// | |
1559 | /// This is a convenience routine to avoid needing to import the [`Config`] | |
1560 | /// type when customizing the construction of a dense DFA. | |
1561 | pub fn config() -> Config { | |
1562 | Config::new() | |
1563 | } | |
1564 | ||
1565 | /// Create a new dense DFA builder with the default configuration. | |
1566 | /// | |
1567 | /// This is a convenience routine to avoid needing to import the | |
1568 | /// [`Builder`] type in common cases. | |
1569 | pub fn builder() -> Builder { | |
1570 | Builder::new() | |
1571 | } | |
1572 | } | |
1573 | ||
487cf647 FG |
1574 | impl<T: AsRef<[u32]>> DFA<T> { |
1575 | /// Cheaply return a borrowed version of this dense DFA. Specifically, | |
1576 | /// the DFA returned always uses `&[u32]` for its transition table. | |
1577 | pub fn as_ref(&self) -> DFA<&'_ [u32]> { | |
1578 | DFA { | |
1579 | tt: self.tt.as_ref(), | |
1580 | st: self.st.as_ref(), | |
1581 | ms: self.ms.as_ref(), | |
1582 | special: self.special, | |
1583 | accels: self.accels(), | |
781aab86 FG |
1584 | pre: self.pre.clone(), |
1585 | quitset: self.quitset, | |
1586 | flags: self.flags, | |
487cf647 FG |
1587 | } |
1588 | } | |
1589 | ||
1590 | /// Return an owned version of this sparse DFA. Specifically, the DFA | |
1591 | /// returned always uses `Vec<u32>` for its transition table. | |
1592 | /// | |
1593 | /// Effectively, this returns a dense DFA whose transition table lives on | |
1594 | /// the heap. | |
1595 | #[cfg(feature = "alloc")] | |
1596 | pub fn to_owned(&self) -> OwnedDFA { | |
1597 | DFA { | |
1598 | tt: self.tt.to_owned(), | |
1599 | st: self.st.to_owned(), | |
1600 | ms: self.ms.to_owned(), | |
1601 | special: self.special, | |
1602 | accels: self.accels().to_owned(), | |
781aab86 FG |
1603 | pre: self.pre.clone(), |
1604 | quitset: self.quitset, | |
1605 | flags: self.flags, | |
487cf647 FG |
1606 | } |
1607 | } | |
1608 | ||
781aab86 FG |
1609 | /// Returns the starting state configuration for this DFA. |
1610 | /// | |
1611 | /// The default is [`StartKind::Both`], which means the DFA supports both | |
1612 | /// unanchored and anchored searches. However, this can generally lead to | |
1613 | /// bigger DFAs. Therefore, a DFA might be compiled with support for just | |
1614 | /// unanchored or anchored searches. In that case, running a search with | |
1615 | /// an unsupported configuration will panic. | |
1616 | pub fn start_kind(&self) -> StartKind { | |
1617 | self.st.kind | |
1618 | } | |
1619 | ||
1620 | /// Returns the start byte map used for computing the `Start` configuration | |
1621 | /// at the beginning of a search. | |
1622 | pub(crate) fn start_map(&self) -> &StartByteMap { | |
1623 | &self.st.start_map | |
1624 | } | |
1625 | ||
487cf647 FG |
1626 | /// Returns true only if this DFA has starting states for each pattern. |
1627 | /// | |
1628 | /// When a DFA has starting states for each pattern, then a search with the | |
1629 | /// DFA can be configured to only look for anchored matches of a specific | |
781aab86 FG |
1630 | /// pattern. Specifically, APIs like [`Automaton::try_search_fwd`] can |
1631 | /// accept a non-None `pattern_id` if and only if this method returns true. | |
1632 | /// Otherwise, calling `try_search_fwd` will panic. | |
487cf647 FG |
1633 | /// |
1634 | /// Note that if the DFA has no patterns, this always returns false. | |
781aab86 FG |
1635 | pub fn starts_for_each_pattern(&self) -> bool { |
1636 | self.st.pattern_len.is_some() | |
1637 | } | |
1638 | ||
1639 | /// Returns the equivalence classes that make up the alphabet for this DFA. | |
1640 | /// | |
1641 | /// Unless [`Config::byte_classes`] was disabled, it is possible that | |
1642 | /// multiple distinct bytes are grouped into the same equivalence class | |
1643 | /// if it is impossible for them to discriminate between a match and a | |
1644 | /// non-match. This has the effect of reducing the overall alphabet size | |
1645 | /// and in turn potentially substantially reducing the size of the DFA's | |
1646 | /// transition table. | |
1647 | /// | |
1648 | /// The downside of using equivalence classes like this is that every state | |
1649 | /// transition will automatically use this map to convert an arbitrary | |
1650 | /// byte to its corresponding equivalence class. In practice this has a | |
1651 | /// negligible impact on performance. | |
1652 | pub fn byte_classes(&self) -> &ByteClasses { | |
1653 | &self.tt.classes | |
487cf647 FG |
1654 | } |
1655 | ||
1656 | /// Returns the total number of elements in the alphabet for this DFA. | |
1657 | /// | |
1658 | /// That is, this returns the total number of transitions that each state | |
1659 | /// in this DFA must have. Typically, a normal byte oriented DFA would | |
1660 | /// always have an alphabet size of 256, corresponding to the number of | |
1661 | /// unique values in a single byte. However, this implementation has two | |
1662 | /// peculiarities that impact the alphabet length: | |
1663 | /// | |
1664 | /// * Every state has a special "EOI" transition that is only followed | |
1665 | /// after the end of some haystack is reached. This EOI transition is | |
1666 | /// necessary to account for one byte of look-ahead when implementing | |
1667 | /// things like `\b` and `$`. | |
1668 | /// * Bytes are grouped into equivalence classes such that no two bytes in | |
1669 | /// the same class can distinguish a match from a non-match. For example, | |
1670 | /// in the regex `^[a-z]+$`, the ASCII bytes `a-z` could all be in the | |
1671 | /// same equivalence class. This leads to a massive space savings. | |
1672 | /// | |
1673 | /// Note though that the alphabet length does _not_ necessarily equal the | |
1674 | /// total stride space taken up by a single DFA state in the transition | |
1675 | /// table. Namely, for performance reasons, the stride is always the | |
1676 | /// smallest power of two that is greater than or equal to the alphabet | |
1677 | /// length. For this reason, [`DFA::stride`] or [`DFA::stride2`] are | |
1678 | /// often more useful. The alphabet length is typically useful only for | |
1679 | /// informational purposes. | |
1680 | pub fn alphabet_len(&self) -> usize { | |
1681 | self.tt.alphabet_len() | |
1682 | } | |
1683 | ||
1684 | /// Returns the total stride for every state in this DFA, expressed as the | |
1685 | /// exponent of a power of 2. The stride is the amount of space each state | |
1686 | /// takes up in the transition table, expressed as a number of transitions. | |
1687 | /// (Unused transitions map to dead states.) | |
1688 | /// | |
1689 | /// The stride of a DFA is always equivalent to the smallest power of 2 | |
1690 | /// that is greater than or equal to the DFA's alphabet length. This | |
1691 | /// definition uses extra space, but permits faster translation between | |
1692 | /// premultiplied state identifiers and contiguous indices (by using shifts | |
1693 | /// instead of relying on integer division). | |
1694 | /// | |
1695 | /// For example, if the DFA's stride is 16 transitions, then its `stride2` | |
1696 | /// is `4` since `2^4 = 16`. | |
1697 | /// | |
1698 | /// The minimum `stride2` value is `1` (corresponding to a stride of `2`) | |
1699 | /// while the maximum `stride2` value is `9` (corresponding to a stride of | |
1700 | /// `512`). The maximum is not `8` since the maximum alphabet size is `257` | |
1701 | /// when accounting for the special EOI transition. However, an alphabet | |
1702 | /// length of that size is exceptionally rare since the alphabet is shrunk | |
1703 | /// into equivalence classes. | |
1704 | pub fn stride2(&self) -> usize { | |
1705 | self.tt.stride2 | |
1706 | } | |
1707 | ||
1708 | /// Returns the total stride for every state in this DFA. This corresponds | |
1709 | /// to the total number of transitions used by each state in this DFA's | |
1710 | /// transition table. | |
1711 | /// | |
1712 | /// Please see [`DFA::stride2`] for more information. In particular, this | |
1713 | /// returns the stride as the number of transitions, where as `stride2` | |
1714 | /// returns it as the exponent of a power of 2. | |
1715 | pub fn stride(&self) -> usize { | |
1716 | self.tt.stride() | |
1717 | } | |
1718 | ||
487cf647 FG |
1719 | /// Returns the memory usage, in bytes, of this DFA. |
1720 | /// | |
1721 | /// The memory usage is computed based on the number of bytes used to | |
1722 | /// represent this DFA. | |
1723 | /// | |
1724 | /// This does **not** include the stack size used up by this DFA. To | |
1725 | /// compute that, use `std::mem::size_of::<dense::DFA>()`. | |
1726 | pub fn memory_usage(&self) -> usize { | |
1727 | self.tt.memory_usage() | |
1728 | + self.st.memory_usage() | |
1729 | + self.ms.memory_usage() | |
1730 | + self.accels.memory_usage() | |
1731 | } | |
1732 | } | |
1733 | ||
1734 | /// Routines for converting a dense DFA to other representations, such as | |
1735 | /// sparse DFAs or raw bytes suitable for persistent storage. | |
1736 | impl<T: AsRef<[u32]>> DFA<T> { | |
1737 | /// Convert this dense DFA to a sparse DFA. | |
1738 | /// | |
1739 | /// If a `StateID` is too small to represent all states in the sparse | |
1740 | /// DFA, then this returns an error. In most cases, if a dense DFA is | |
1741 | /// constructable with `StateID` then a sparse DFA will be as well. | |
1742 | /// However, it is not guaranteed. | |
1743 | /// | |
1744 | /// # Example | |
1745 | /// | |
1746 | /// ``` | |
781aab86 | 1747 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
487cf647 FG |
1748 | /// |
1749 | /// let dense = dense::DFA::new("foo[0-9]+")?; | |
1750 | /// let sparse = dense.to_sparse()?; | |
1751 | /// | |
781aab86 FG |
1752 | /// let expected = Some(HalfMatch::must(0, 8)); |
1753 | /// assert_eq!(expected, sparse.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
1754 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1755 | /// ``` | |
781aab86 FG |
1756 | #[cfg(feature = "dfa-build")] |
1757 | pub fn to_sparse(&self) -> Result<sparse::DFA<Vec<u8>>, BuildError> { | |
487cf647 FG |
1758 | sparse::DFA::from_dense(self) |
1759 | } | |
1760 | ||
1761 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in little endian | |
1762 | /// format. Upon success, the `Vec<u8>` and the initial padding length are | |
1763 | /// returned. | |
1764 | /// | |
1765 | /// The written bytes are guaranteed to be deserialized correctly and | |
1766 | /// without errors in a semver compatible release of this crate by a | |
1767 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
1768 | /// deserialization APIs has been satisfied): | |
1769 | /// | |
1770 | /// * [`DFA::from_bytes`] | |
1771 | /// * [`DFA::from_bytes_unchecked`] | |
1772 | /// | |
1773 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at | |
1774 | /// an address that does not have the same alignment as `u32`. The padding | |
1775 | /// corresponds to the number of leading bytes written to the returned | |
1776 | /// `Vec<u8>`. | |
1777 | /// | |
1778 | /// # Example | |
1779 | /// | |
1780 | /// This example shows how to serialize and deserialize a DFA: | |
1781 | /// | |
1782 | /// ``` | |
781aab86 | 1783 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1784 | /// |
1785 | /// // Compile our original DFA. | |
1786 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
1787 | /// | |
1788 | /// // N.B. We use native endianness here to make the example work, but | |
1789 | /// // using to_bytes_little_endian would work on a little endian target. | |
1790 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); | |
1791 | /// // Even if buf has initial padding, DFA::from_bytes will automatically | |
1792 | /// // ignore it. | |
1793 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; | |
1794 | /// | |
781aab86 FG |
1795 | /// let expected = Some(HalfMatch::must(0, 8)); |
1796 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
1797 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1798 | /// ``` | |
781aab86 | 1799 | #[cfg(feature = "dfa-build")] |
487cf647 | 1800 | pub fn to_bytes_little_endian(&self) -> (Vec<u8>, usize) { |
781aab86 | 1801 | self.to_bytes::<wire::LE>() |
487cf647 FG |
1802 | } |
1803 | ||
1804 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in big endian | |
1805 | /// format. Upon success, the `Vec<u8>` and the initial padding length are | |
1806 | /// returned. | |
1807 | /// | |
1808 | /// The written bytes are guaranteed to be deserialized correctly and | |
1809 | /// without errors in a semver compatible release of this crate by a | |
1810 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
1811 | /// deserialization APIs has been satisfied): | |
1812 | /// | |
1813 | /// * [`DFA::from_bytes`] | |
1814 | /// * [`DFA::from_bytes_unchecked`] | |
1815 | /// | |
1816 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at | |
1817 | /// an address that does not have the same alignment as `u32`. The padding | |
1818 | /// corresponds to the number of leading bytes written to the returned | |
1819 | /// `Vec<u8>`. | |
1820 | /// | |
1821 | /// # Example | |
1822 | /// | |
1823 | /// This example shows how to serialize and deserialize a DFA: | |
1824 | /// | |
1825 | /// ``` | |
781aab86 | 1826 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1827 | /// |
1828 | /// // Compile our original DFA. | |
1829 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
1830 | /// | |
1831 | /// // N.B. We use native endianness here to make the example work, but | |
1832 | /// // using to_bytes_big_endian would work on a big endian target. | |
1833 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); | |
1834 | /// // Even if buf has initial padding, DFA::from_bytes will automatically | |
1835 | /// // ignore it. | |
1836 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; | |
1837 | /// | |
781aab86 FG |
1838 | /// let expected = Some(HalfMatch::must(0, 8)); |
1839 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
1840 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1841 | /// ``` | |
781aab86 | 1842 | #[cfg(feature = "dfa-build")] |
487cf647 | 1843 | pub fn to_bytes_big_endian(&self) -> (Vec<u8>, usize) { |
781aab86 | 1844 | self.to_bytes::<wire::BE>() |
487cf647 FG |
1845 | } |
1846 | ||
1847 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in native endian | |
1848 | /// format. Upon success, the `Vec<u8>` and the initial padding length are | |
1849 | /// returned. | |
1850 | /// | |
1851 | /// The written bytes are guaranteed to be deserialized correctly and | |
1852 | /// without errors in a semver compatible release of this crate by a | |
1853 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
1854 | /// deserialization APIs has been satisfied): | |
1855 | /// | |
1856 | /// * [`DFA::from_bytes`] | |
1857 | /// * [`DFA::from_bytes_unchecked`] | |
1858 | /// | |
1859 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at | |
1860 | /// an address that does not have the same alignment as `u32`. The padding | |
1861 | /// corresponds to the number of leading bytes written to the returned | |
1862 | /// `Vec<u8>`. | |
1863 | /// | |
1864 | /// Generally speaking, native endian format should only be used when | |
1865 | /// you know that the target you're compiling the DFA for matches the | |
1866 | /// endianness of the target on which you're compiling DFA. For example, | |
1867 | /// if serialization and deserialization happen in the same process or on | |
1868 | /// the same machine. Otherwise, when serializing a DFA for use in a | |
1869 | /// portable environment, you'll almost certainly want to serialize _both_ | |
1870 | /// a little endian and a big endian version and then load the correct one | |
1871 | /// based on the target's configuration. | |
1872 | /// | |
1873 | /// # Example | |
1874 | /// | |
1875 | /// This example shows how to serialize and deserialize a DFA: | |
1876 | /// | |
1877 | /// ``` | |
781aab86 | 1878 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1879 | /// |
1880 | /// // Compile our original DFA. | |
1881 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
1882 | /// | |
1883 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); | |
1884 | /// // Even if buf has initial padding, DFA::from_bytes will automatically | |
1885 | /// // ignore it. | |
1886 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; | |
1887 | /// | |
781aab86 FG |
1888 | /// let expected = Some(HalfMatch::must(0, 8)); |
1889 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
1890 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1891 | /// ``` | |
781aab86 | 1892 | #[cfg(feature = "dfa-build")] |
487cf647 | 1893 | pub fn to_bytes_native_endian(&self) -> (Vec<u8>, usize) { |
781aab86 | 1894 | self.to_bytes::<wire::NE>() |
487cf647 FG |
1895 | } |
1896 | ||
1897 | /// The implementation of the public `to_bytes` serialization methods, | |
1898 | /// which is generic over endianness. | |
781aab86 | 1899 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
1900 | fn to_bytes<E: Endian>(&self) -> (Vec<u8>, usize) { |
1901 | let len = self.write_to_len(); | |
781aab86 | 1902 | let (mut buf, padding) = wire::alloc_aligned_buffer::<u32>(len); |
487cf647 FG |
1903 | // This should always succeed since the only possible serialization |
1904 | // error is providing a buffer that's too small, but we've ensured that | |
1905 | // `buf` is big enough here. | |
1906 | self.as_ref().write_to::<E>(&mut buf[padding..]).unwrap(); | |
1907 | (buf, padding) | |
1908 | } | |
1909 | ||
1910 | /// Serialize this DFA as raw bytes to the given slice, in little endian | |
1911 | /// format. Upon success, the total number of bytes written to `dst` is | |
1912 | /// returned. | |
1913 | /// | |
1914 | /// The written bytes are guaranteed to be deserialized correctly and | |
1915 | /// without errors in a semver compatible release of this crate by a | |
1916 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
1917 | /// deserialization APIs has been satisfied): | |
1918 | /// | |
1919 | /// * [`DFA::from_bytes`] | |
1920 | /// * [`DFA::from_bytes_unchecked`] | |
1921 | /// | |
1922 | /// Note that unlike the various `to_byte_*` routines, this does not write | |
1923 | /// any padding. Callers are responsible for handling alignment correctly. | |
1924 | /// | |
1925 | /// # Errors | |
1926 | /// | |
1927 | /// This returns an error if the given destination slice is not big enough | |
1928 | /// to contain the full serialized DFA. If an error occurs, then nothing | |
1929 | /// is written to `dst`. | |
1930 | /// | |
1931 | /// # Example | |
1932 | /// | |
1933 | /// This example shows how to serialize and deserialize a DFA without | |
1934 | /// dynamic memory allocation. | |
1935 | /// | |
1936 | /// ``` | |
781aab86 | 1937 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1938 | /// |
1939 | /// // Compile our original DFA. | |
1940 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
1941 | /// | |
781aab86 FG |
1942 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
1943 | /// // need to use a special type to force the alignment of our [u8; N] | |
1944 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing | |
1945 | /// // the DFA may fail because of an alignment mismatch. | |
1946 | /// #[repr(C)] | |
1947 | /// struct Aligned<B: ?Sized> { | |
1948 | /// _align: [u32; 0], | |
1949 | /// bytes: B, | |
1950 | /// } | |
1951 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; | |
487cf647 FG |
1952 | /// // N.B. We use native endianness here to make the example work, but |
1953 | /// // using write_to_little_endian would work on a little endian target. | |
781aab86 FG |
1954 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; |
1955 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; | |
487cf647 | 1956 | /// |
781aab86 FG |
1957 | /// let expected = Some(HalfMatch::must(0, 8)); |
1958 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
1959 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1960 | /// ``` | |
1961 | pub fn write_to_little_endian( | |
1962 | &self, | |
1963 | dst: &mut [u8], | |
1964 | ) -> Result<usize, SerializeError> { | |
781aab86 | 1965 | self.as_ref().write_to::<wire::LE>(dst) |
487cf647 FG |
1966 | } |
1967 | ||
1968 | /// Serialize this DFA as raw bytes to the given slice, in big endian | |
1969 | /// format. Upon success, the total number of bytes written to `dst` is | |
1970 | /// returned. | |
1971 | /// | |
1972 | /// The written bytes are guaranteed to be deserialized correctly and | |
1973 | /// without errors in a semver compatible release of this crate by a | |
1974 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
1975 | /// deserialization APIs has been satisfied): | |
1976 | /// | |
1977 | /// * [`DFA::from_bytes`] | |
1978 | /// * [`DFA::from_bytes_unchecked`] | |
1979 | /// | |
1980 | /// Note that unlike the various `to_byte_*` routines, this does not write | |
1981 | /// any padding. Callers are responsible for handling alignment correctly. | |
1982 | /// | |
1983 | /// # Errors | |
1984 | /// | |
1985 | /// This returns an error if the given destination slice is not big enough | |
1986 | /// to contain the full serialized DFA. If an error occurs, then nothing | |
1987 | /// is written to `dst`. | |
1988 | /// | |
1989 | /// # Example | |
1990 | /// | |
1991 | /// This example shows how to serialize and deserialize a DFA without | |
1992 | /// dynamic memory allocation. | |
1993 | /// | |
1994 | /// ``` | |
781aab86 | 1995 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
1996 | /// |
1997 | /// // Compile our original DFA. | |
1998 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
1999 | /// | |
781aab86 FG |
2000 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
2001 | /// // need to use a special type to force the alignment of our [u8; N] | |
2002 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing | |
2003 | /// // the DFA may fail because of an alignment mismatch. | |
2004 | /// #[repr(C)] | |
2005 | /// struct Aligned<B: ?Sized> { | |
2006 | /// _align: [u32; 0], | |
2007 | /// bytes: B, | |
2008 | /// } | |
2009 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; | |
487cf647 FG |
2010 | /// // N.B. We use native endianness here to make the example work, but |
2011 | /// // using write_to_big_endian would work on a big endian target. | |
781aab86 FG |
2012 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; |
2013 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; | |
487cf647 | 2014 | /// |
781aab86 FG |
2015 | /// let expected = Some(HalfMatch::must(0, 8)); |
2016 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2017 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2018 | /// ``` | |
2019 | pub fn write_to_big_endian( | |
2020 | &self, | |
2021 | dst: &mut [u8], | |
2022 | ) -> Result<usize, SerializeError> { | |
781aab86 | 2023 | self.as_ref().write_to::<wire::BE>(dst) |
487cf647 FG |
2024 | } |
2025 | ||
2026 | /// Serialize this DFA as raw bytes to the given slice, in native endian | |
2027 | /// format. Upon success, the total number of bytes written to `dst` is | |
2028 | /// returned. | |
2029 | /// | |
2030 | /// The written bytes are guaranteed to be deserialized correctly and | |
2031 | /// without errors in a semver compatible release of this crate by a | |
2032 | /// `DFA`'s deserialization APIs (assuming all other criteria for the | |
2033 | /// deserialization APIs has been satisfied): | |
2034 | /// | |
2035 | /// * [`DFA::from_bytes`] | |
2036 | /// * [`DFA::from_bytes_unchecked`] | |
2037 | /// | |
2038 | /// Generally speaking, native endian format should only be used when | |
2039 | /// you know that the target you're compiling the DFA for matches the | |
2040 | /// endianness of the target on which you're compiling DFA. For example, | |
2041 | /// if serialization and deserialization happen in the same process or on | |
2042 | /// the same machine. Otherwise, when serializing a DFA for use in a | |
2043 | /// portable environment, you'll almost certainly want to serialize _both_ | |
2044 | /// a little endian and a big endian version and then load the correct one | |
2045 | /// based on the target's configuration. | |
2046 | /// | |
2047 | /// Note that unlike the various `to_byte_*` routines, this does not write | |
2048 | /// any padding. Callers are responsible for handling alignment correctly. | |
2049 | /// | |
2050 | /// # Errors | |
2051 | /// | |
2052 | /// This returns an error if the given destination slice is not big enough | |
2053 | /// to contain the full serialized DFA. If an error occurs, then nothing | |
2054 | /// is written to `dst`. | |
2055 | /// | |
2056 | /// # Example | |
2057 | /// | |
2058 | /// This example shows how to serialize and deserialize a DFA without | |
2059 | /// dynamic memory allocation. | |
2060 | /// | |
2061 | /// ``` | |
781aab86 | 2062 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
2063 | /// |
2064 | /// // Compile our original DFA. | |
2065 | /// let original_dfa = DFA::new("foo[0-9]+")?; | |
2066 | /// | |
781aab86 FG |
2067 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
2068 | /// // need to use a special type to force the alignment of our [u8; N] | |
2069 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing | |
2070 | /// // the DFA may fail because of an alignment mismatch. | |
2071 | /// #[repr(C)] | |
2072 | /// struct Aligned<B: ?Sized> { | |
2073 | /// _align: [u32; 0], | |
2074 | /// bytes: B, | |
2075 | /// } | |
2076 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; | |
2077 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; | |
2078 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; | |
487cf647 | 2079 | /// |
781aab86 FG |
2080 | /// let expected = Some(HalfMatch::must(0, 8)); |
2081 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2082 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2083 | /// ``` | |
2084 | pub fn write_to_native_endian( | |
2085 | &self, | |
2086 | dst: &mut [u8], | |
2087 | ) -> Result<usize, SerializeError> { | |
781aab86 | 2088 | self.as_ref().write_to::<wire::NE>(dst) |
487cf647 FG |
2089 | } |
2090 | ||
2091 | /// Return the total number of bytes required to serialize this DFA. | |
2092 | /// | |
2093 | /// This is useful for determining the size of the buffer required to pass | |
2094 | /// to one of the serialization routines: | |
2095 | /// | |
2096 | /// * [`DFA::write_to_little_endian`] | |
2097 | /// * [`DFA::write_to_big_endian`] | |
2098 | /// * [`DFA::write_to_native_endian`] | |
2099 | /// | |
2100 | /// Passing a buffer smaller than the size returned by this method will | |
2101 | /// result in a serialization error. Serialization routines are guaranteed | |
2102 | /// to succeed when the buffer is big enough. | |
2103 | /// | |
2104 | /// # Example | |
2105 | /// | |
2106 | /// This example shows how to dynamically allocate enough room to serialize | |
2107 | /// a DFA. | |
2108 | /// | |
2109 | /// ``` | |
781aab86 | 2110 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 | 2111 | /// |
487cf647 FG |
2112 | /// let original_dfa = DFA::new("foo[0-9]+")?; |
2113 | /// | |
2114 | /// let mut buf = vec![0; original_dfa.write_to_len()]; | |
781aab86 FG |
2115 | /// // This is guaranteed to succeed, because the only serialization error |
2116 | /// // that can occur is when the provided buffer is too small. But | |
2117 | /// // write_to_len guarantees a correct size. | |
2118 | /// let written = original_dfa.write_to_native_endian(&mut buf).unwrap(); | |
2119 | /// // But this is not guaranteed to succeed! In particular, | |
2120 | /// // deserialization requires proper alignment for &[u32], but our buffer | |
2121 | /// // was allocated as a &[u8] whose required alignment is smaller than | |
2122 | /// // &[u32]. However, it's likely to work in practice because of how most | |
2123 | /// // allocators work. So if you write code like this, make sure to either | |
2124 | /// // handle the error correctly and/or run it under Miri since Miri will | |
2125 | /// // likely provoke the error by returning Vec<u8> buffers with alignment | |
2126 | /// // less than &[u32]. | |
2127 | /// let dfa: DFA<&[u32]> = match DFA::from_bytes(&buf[..written]) { | |
2128 | /// // As mentioned above, it is legal for an error to be returned | |
2129 | /// // here. It is quite difficult to get a Vec<u8> with a guaranteed | |
2130 | /// // alignment equivalent to Vec<u32>. | |
2131 | /// Err(_) => return Ok(()), | |
2132 | /// Ok((dfa, _)) => dfa, | |
2133 | /// }; | |
487cf647 | 2134 | /// |
781aab86 FG |
2135 | /// let expected = Some(HalfMatch::must(0, 8)); |
2136 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2137 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2138 | /// ``` | |
2139 | /// | |
2140 | /// Note that this example isn't actually guaranteed to work! In | |
2141 | /// particular, if `buf` is not aligned to a 4-byte boundary, then the | |
2142 | /// `DFA::from_bytes` call will fail. If you need this to work, then you | |
2143 | /// either need to deal with adding some initial padding yourself, or use | |
2144 | /// one of the `to_bytes` methods, which will do it for you. | |
2145 | pub fn write_to_len(&self) -> usize { | |
781aab86 FG |
2146 | wire::write_label_len(LABEL) |
2147 | + wire::write_endianness_check_len() | |
2148 | + wire::write_version_len() | |
487cf647 | 2149 | + size_of::<u32>() // unused, intended for future flexibility |
781aab86 | 2150 | + self.flags.write_to_len() |
487cf647 FG |
2151 | + self.tt.write_to_len() |
2152 | + self.st.write_to_len() | |
2153 | + self.ms.write_to_len() | |
2154 | + self.special.write_to_len() | |
2155 | + self.accels.write_to_len() | |
781aab86 | 2156 | + self.quitset.write_to_len() |
487cf647 FG |
2157 | } |
2158 | } | |
2159 | ||
2160 | impl<'a> DFA<&'a [u32]> { | |
2161 | /// Safely deserialize a DFA with a specific state identifier | |
2162 | /// representation. Upon success, this returns both the deserialized DFA | |
2163 | /// and the number of bytes read from the given slice. Namely, the contents | |
2164 | /// of the slice beyond the DFA are not read. | |
2165 | /// | |
2166 | /// Deserializing a DFA using this routine will never allocate heap memory. | |
2167 | /// For safety purposes, the DFA's transition table will be verified such | |
2168 | /// that every transition points to a valid state. If this verification is | |
2169 | /// too costly, then a [`DFA::from_bytes_unchecked`] API is provided, which | |
2170 | /// will always execute in constant time. | |
2171 | /// | |
2172 | /// The bytes given must be generated by one of the serialization APIs | |
2173 | /// of a `DFA` using a semver compatible release of this crate. Those | |
2174 | /// include: | |
2175 | /// | |
2176 | /// * [`DFA::to_bytes_little_endian`] | |
2177 | /// * [`DFA::to_bytes_big_endian`] | |
2178 | /// * [`DFA::to_bytes_native_endian`] | |
2179 | /// * [`DFA::write_to_little_endian`] | |
2180 | /// * [`DFA::write_to_big_endian`] | |
2181 | /// * [`DFA::write_to_native_endian`] | |
2182 | /// | |
2183 | /// The `to_bytes` methods allocate and return a `Vec<u8>` for you, along | |
2184 | /// with handling alignment correctly. The `write_to` methods do not | |
2185 | /// allocate and write to an existing slice (which may be on the stack). | |
2186 | /// Since deserialization always uses the native endianness of the target | |
2187 | /// platform, the serialization API you use should match the endianness of | |
2188 | /// the target platform. (It's often a good idea to generate serialized | |
2189 | /// DFAs for both forms of endianness and then load the correct one based | |
2190 | /// on endianness.) | |
2191 | /// | |
2192 | /// # Errors | |
2193 | /// | |
2194 | /// Generally speaking, it's easier to state the conditions in which an | |
2195 | /// error is _not_ returned. All of the following must be true: | |
2196 | /// | |
2197 | /// * The bytes given must be produced by one of the serialization APIs | |
2198 | /// on this DFA, as mentioned above. | |
2199 | /// * The endianness of the target platform matches the endianness used to | |
2200 | /// serialized the provided DFA. | |
2201 | /// * The slice given must have the same alignment as `u32`. | |
2202 | /// | |
2203 | /// If any of the above are not true, then an error will be returned. | |
2204 | /// | |
2205 | /// # Panics | |
2206 | /// | |
2207 | /// This routine will never panic for any input. | |
2208 | /// | |
2209 | /// # Example | |
2210 | /// | |
2211 | /// This example shows how to serialize a DFA to raw bytes, deserialize it | |
2212 | /// and then use it for searching. | |
2213 | /// | |
2214 | /// ``` | |
781aab86 | 2215 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
2216 | /// |
2217 | /// let initial = DFA::new("foo[0-9]+")?; | |
2218 | /// let (bytes, _) = initial.to_bytes_native_endian(); | |
2219 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&bytes)?.0; | |
2220 | /// | |
781aab86 FG |
2221 | /// let expected = Some(HalfMatch::must(0, 8)); |
2222 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2223 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2224 | /// ``` | |
2225 | /// | |
2226 | /// # Example: dealing with alignment and padding | |
2227 | /// | |
2228 | /// In the above example, we used the `to_bytes_native_endian` method to | |
2229 | /// serialize a DFA, but we ignored part of its return value corresponding | |
2230 | /// to padding added to the beginning of the serialized DFA. This is OK | |
2231 | /// because deserialization will skip this initial padding. What matters | |
2232 | /// is that the address immediately following the padding has an alignment | |
2233 | /// that matches `u32`. That is, the following is an equivalent but | |
2234 | /// alternative way to write the above example: | |
2235 | /// | |
2236 | /// ``` | |
781aab86 | 2237 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
2238 | /// |
2239 | /// let initial = DFA::new("foo[0-9]+")?; | |
2240 | /// // Serialization returns the number of leading padding bytes added to | |
2241 | /// // the returned Vec<u8>. | |
2242 | /// let (bytes, pad) = initial.to_bytes_native_endian(); | |
2243 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&bytes[pad..])?.0; | |
2244 | /// | |
781aab86 FG |
2245 | /// let expected = Some(HalfMatch::must(0, 8)); |
2246 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2247 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2248 | /// ``` | |
2249 | /// | |
2250 | /// This padding is necessary because Rust's standard library does | |
2251 | /// not expose any safe and robust way of creating a `Vec<u8>` with a | |
2252 | /// guaranteed alignment other than 1. Now, in practice, the underlying | |
2253 | /// allocator is likely to provide a `Vec<u8>` that meets our alignment | |
2254 | /// requirements, which means `pad` is zero in practice most of the time. | |
2255 | /// | |
2256 | /// The purpose of exposing the padding like this is flexibility for the | |
2257 | /// caller. For example, if one wants to embed a serialized DFA into a | |
2258 | /// compiled program, then it's important to guarantee that it starts at a | |
2259 | /// `u32`-aligned address. The simplest way to do this is to discard the | |
2260 | /// padding bytes and set it up so that the serialized DFA itself begins at | |
2261 | /// a properly aligned address. We can show this in two parts. The first | |
2262 | /// part is serializing the DFA to a file: | |
2263 | /// | |
2264 | /// ```no_run | |
781aab86 | 2265 | /// use regex_automata::dfa::dense::DFA; |
487cf647 FG |
2266 | /// |
2267 | /// let dfa = DFA::new("foo[0-9]+")?; | |
2268 | /// | |
2269 | /// let (bytes, pad) = dfa.to_bytes_big_endian(); | |
2270 | /// // Write the contents of the DFA *without* the initial padding. | |
2271 | /// std::fs::write("foo.bigendian.dfa", &bytes[pad..])?; | |
2272 | /// | |
2273 | /// // Do it again, but this time for little endian. | |
2274 | /// let (bytes, pad) = dfa.to_bytes_little_endian(); | |
2275 | /// std::fs::write("foo.littleendian.dfa", &bytes[pad..])?; | |
2276 | /// # Ok::<(), Box<dyn std::error::Error>>(()) | |
2277 | /// ``` | |
2278 | /// | |
2279 | /// And now the second part is embedding the DFA into the compiled program | |
2280 | /// and deserializing it at runtime on first use. We use conditional | |
2281 | /// compilation to choose the correct endianness. | |
2282 | /// | |
2283 | /// ```no_run | |
781aab86 FG |
2284 | /// use regex_automata::{ |
2285 | /// dfa::{Automaton, dense::DFA}, | |
2286 | /// util::{lazy::Lazy, wire::AlignAs}, | |
2287 | /// HalfMatch, Input, | |
2288 | /// }; | |
487cf647 | 2289 | /// |
781aab86 FG |
2290 | /// // This crate provides its own "lazy" type, kind of like |
2291 | /// // lazy_static! or once_cell::sync::Lazy. But it works in no-alloc | |
2292 | /// // no-std environments and let's us write this using completely | |
2293 | /// // safe code. | |
2294 | /// static RE: Lazy<DFA<&'static [u32]>> = Lazy::new(|| { | |
487cf647 FG |
2295 | /// # const _: &str = stringify! { |
2296 | /// // This assignment is made possible (implicitly) via the | |
781aab86 FG |
2297 | /// // CoerceUnsized trait. This is what guarantees that our |
2298 | /// // bytes are stored in memory on a 4 byte boundary. You | |
2299 | /// // *must* do this or something equivalent for correct | |
2300 | /// // deserialization. | |
2301 | /// static ALIGNED: &AlignAs<[u8], u32> = &AlignAs { | |
487cf647 FG |
2302 | /// _align: [], |
2303 | /// #[cfg(target_endian = "big")] | |
2304 | /// bytes: *include_bytes!("foo.bigendian.dfa"), | |
2305 | /// #[cfg(target_endian = "little")] | |
2306 | /// bytes: *include_bytes!("foo.littleendian.dfa"), | |
2307 | /// }; | |
2308 | /// # }; | |
781aab86 | 2309 | /// # static ALIGNED: &AlignAs<[u8], u32> = &AlignAs { |
487cf647 FG |
2310 | /// # _align: [], |
2311 | /// # bytes: [], | |
2312 | /// # }; | |
2313 | /// | |
781aab86 FG |
2314 | /// let (dfa, _) = DFA::from_bytes(&ALIGNED.bytes) |
2315 | /// .expect("serialized DFA should be valid"); | |
2316 | /// dfa | |
2317 | /// }); | |
487cf647 | 2318 | /// |
781aab86 FG |
2319 | /// let expected = Ok(Some(HalfMatch::must(0, 8))); |
2320 | /// assert_eq!(expected, RE.try_search_fwd(&Input::new("foo12345"))); | |
487cf647 FG |
2321 | /// ``` |
2322 | /// | |
781aab86 FG |
2323 | /// An alternative to [`util::lazy::Lazy`](crate::util::lazy::Lazy) |
2324 | /// is [`lazy_static`](https://crates.io/crates/lazy_static) or | |
2325 | /// [`once_cell`](https://crates.io/crates/once_cell), which provide | |
2326 | /// stronger guarantees (like the initialization function only being | |
2327 | /// executed once). And `once_cell` in particular provides a more | |
2328 | /// expressive API. But a `Lazy` value from this crate is likely just fine | |
2329 | /// in most circumstances. | |
2330 | /// | |
2331 | /// Note that regardless of which initialization method you use, you | |
2332 | /// will still need to use the [`AlignAs`](crate::util::wire::AlignAs) | |
2333 | /// trick above to force correct alignment, but this is safe to do and | |
2334 | /// `from_bytes` will return an error if you get it wrong. | |
487cf647 FG |
2335 | pub fn from_bytes( |
2336 | slice: &'a [u8], | |
2337 | ) -> Result<(DFA<&'a [u32]>, usize), DeserializeError> { | |
781aab86 FG |
2338 | // SAFETY: This is safe because we validate the transition table, start |
2339 | // table, match states and accelerators below. If any validation fails, | |
2340 | // then we return an error. | |
487cf647 | 2341 | let (dfa, nread) = unsafe { DFA::from_bytes_unchecked(slice)? }; |
781aab86 | 2342 | dfa.tt.validate(&dfa.special)?; |
487cf647 FG |
2343 | dfa.st.validate(&dfa.tt)?; |
2344 | dfa.ms.validate(&dfa)?; | |
2345 | dfa.accels.validate()?; | |
2346 | // N.B. dfa.special doesn't have a way to do unchecked deserialization, | |
2347 | // so it has already been validated. | |
2348 | Ok((dfa, nread)) | |
2349 | } | |
2350 | ||
2351 | /// Deserialize a DFA with a specific state identifier representation in | |
2352 | /// constant time by omitting the verification of the validity of the | |
2353 | /// transition table and other data inside the DFA. | |
2354 | /// | |
2355 | /// This is just like [`DFA::from_bytes`], except it can potentially return | |
2356 | /// a DFA that exhibits undefined behavior if its transition table contains | |
2357 | /// invalid state identifiers. | |
2358 | /// | |
2359 | /// This routine is useful if you need to deserialize a DFA cheaply | |
2360 | /// and cannot afford the transition table validation performed by | |
2361 | /// `from_bytes`. | |
2362 | /// | |
2363 | /// # Example | |
2364 | /// | |
2365 | /// ``` | |
781aab86 | 2366 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
487cf647 FG |
2367 | /// |
2368 | /// let initial = DFA::new("foo[0-9]+")?; | |
2369 | /// let (bytes, _) = initial.to_bytes_native_endian(); | |
2370 | /// // SAFETY: This is guaranteed to be safe since the bytes given come | |
2371 | /// // directly from a compatible serialization routine. | |
2372 | /// let dfa: DFA<&[u32]> = unsafe { DFA::from_bytes_unchecked(&bytes)?.0 }; | |
2373 | /// | |
781aab86 FG |
2374 | /// let expected = Some(HalfMatch::must(0, 8)); |
2375 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345"))?); | |
487cf647 FG |
2376 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2377 | /// ``` | |
2378 | pub unsafe fn from_bytes_unchecked( | |
2379 | slice: &'a [u8], | |
2380 | ) -> Result<(DFA<&'a [u32]>, usize), DeserializeError> { | |
2381 | let mut nr = 0; | |
2382 | ||
781aab86 FG |
2383 | nr += wire::skip_initial_padding(slice); |
2384 | wire::check_alignment::<StateID>(&slice[nr..])?; | |
2385 | nr += wire::read_label(&slice[nr..], LABEL)?; | |
2386 | nr += wire::read_endianness_check(&slice[nr..])?; | |
2387 | nr += wire::read_version(&slice[nr..], VERSION)?; | |
487cf647 | 2388 | |
781aab86 | 2389 | let _unused = wire::try_read_u32(&slice[nr..], "unused space")?; |
487cf647 FG |
2390 | nr += size_of::<u32>(); |
2391 | ||
781aab86 FG |
2392 | let (flags, nread) = Flags::from_bytes(&slice[nr..])?; |
2393 | nr += nread; | |
2394 | ||
487cf647 FG |
2395 | let (tt, nread) = TransitionTable::from_bytes_unchecked(&slice[nr..])?; |
2396 | nr += nread; | |
2397 | ||
2398 | let (st, nread) = StartTable::from_bytes_unchecked(&slice[nr..])?; | |
2399 | nr += nread; | |
2400 | ||
2401 | let (ms, nread) = MatchStates::from_bytes_unchecked(&slice[nr..])?; | |
2402 | nr += nread; | |
2403 | ||
2404 | let (special, nread) = Special::from_bytes(&slice[nr..])?; | |
2405 | nr += nread; | |
781aab86 | 2406 | special.validate_state_len(tt.len(), tt.stride2)?; |
487cf647 FG |
2407 | |
2408 | let (accels, nread) = Accels::from_bytes_unchecked(&slice[nr..])?; | |
2409 | nr += nread; | |
2410 | ||
781aab86 FG |
2411 | let (quitset, nread) = ByteSet::from_bytes(&slice[nr..])?; |
2412 | nr += nread; | |
2413 | ||
2414 | // Prefilters don't support serialization, so they're always absent. | |
2415 | let pre = None; | |
2416 | Ok((DFA { tt, st, ms, special, accels, pre, quitset, flags }, nr)) | |
487cf647 FG |
2417 | } |
2418 | ||
2419 | /// The implementation of the public `write_to` serialization methods, | |
2420 | /// which is generic over endianness. | |
2421 | /// | |
2422 | /// This is defined only for &[u32] to reduce binary size/compilation time. | |
2423 | fn write_to<E: Endian>( | |
2424 | &self, | |
2425 | mut dst: &mut [u8], | |
2426 | ) -> Result<usize, SerializeError> { | |
2427 | let nwrite = self.write_to_len(); | |
2428 | if dst.len() < nwrite { | |
2429 | return Err(SerializeError::buffer_too_small("dense DFA")); | |
2430 | } | |
2431 | dst = &mut dst[..nwrite]; | |
2432 | ||
2433 | let mut nw = 0; | |
781aab86 FG |
2434 | nw += wire::write_label(LABEL, &mut dst[nw..])?; |
2435 | nw += wire::write_endianness_check::<E>(&mut dst[nw..])?; | |
2436 | nw += wire::write_version::<E>(VERSION, &mut dst[nw..])?; | |
487cf647 FG |
2437 | nw += { |
2438 | // Currently unused, intended for future flexibility | |
2439 | E::write_u32(0, &mut dst[nw..]); | |
2440 | size_of::<u32>() | |
2441 | }; | |
781aab86 | 2442 | nw += self.flags.write_to::<E>(&mut dst[nw..])?; |
487cf647 FG |
2443 | nw += self.tt.write_to::<E>(&mut dst[nw..])?; |
2444 | nw += self.st.write_to::<E>(&mut dst[nw..])?; | |
2445 | nw += self.ms.write_to::<E>(&mut dst[nw..])?; | |
2446 | nw += self.special.write_to::<E>(&mut dst[nw..])?; | |
2447 | nw += self.accels.write_to::<E>(&mut dst[nw..])?; | |
781aab86 | 2448 | nw += self.quitset.write_to::<E>(&mut dst[nw..])?; |
487cf647 FG |
2449 | Ok(nw) |
2450 | } | |
2451 | } | |
2452 | ||
781aab86 FG |
2453 | // The following methods implement mutable routines on the internal |
2454 | // representation of a DFA. As such, we must fix the first type parameter to a | |
2455 | // `Vec<u32>` since a generic `T: AsRef<[u32]>` does not permit mutation. We | |
2456 | // can get away with this because these methods are internal to the crate and | |
2457 | // are exclusively used during construction of the DFA. | |
2458 | #[cfg(feature = "dfa-build")] | |
487cf647 FG |
2459 | impl OwnedDFA { |
2460 | /// Add a start state of this DFA. | |
2461 | pub(crate) fn set_start_state( | |
2462 | &mut self, | |
781aab86 FG |
2463 | anchored: Anchored, |
2464 | start: Start, | |
487cf647 FG |
2465 | id: StateID, |
2466 | ) { | |
2467 | assert!(self.tt.is_valid(id), "invalid start state"); | |
781aab86 | 2468 | self.st.set_start(anchored, start, id); |
487cf647 FG |
2469 | } |
2470 | ||
2471 | /// Set the given transition to this DFA. Both the `from` and `to` states | |
2472 | /// must already exist. | |
2473 | pub(crate) fn set_transition( | |
2474 | &mut self, | |
2475 | from: StateID, | |
2476 | byte: alphabet::Unit, | |
2477 | to: StateID, | |
2478 | ) { | |
2479 | self.tt.set(from, byte, to); | |
2480 | } | |
2481 | ||
2482 | /// An an empty state (a state where all transitions lead to a dead state) | |
2483 | /// and return its identifier. The identifier returned is guaranteed to | |
2484 | /// not point to any other existing state. | |
2485 | /// | |
2486 | /// If adding a state would exceed `StateID::LIMIT`, then this returns an | |
2487 | /// error. | |
781aab86 | 2488 | pub(crate) fn add_empty_state(&mut self) -> Result<StateID, BuildError> { |
487cf647 FG |
2489 | self.tt.add_empty_state() |
2490 | } | |
2491 | ||
2492 | /// Swap the two states given in the transition table. | |
2493 | /// | |
2494 | /// This routine does not do anything to check the correctness of this | |
2495 | /// swap. Callers must ensure that other states pointing to id1 and id2 are | |
2496 | /// updated appropriately. | |
2497 | pub(crate) fn swap_states(&mut self, id1: StateID, id2: StateID) { | |
2498 | self.tt.swap(id1, id2); | |
2499 | } | |
2500 | ||
781aab86 FG |
2501 | /// Remap all of the state identifiers in this DFA according to the map |
2502 | /// function given. This includes all transitions and all starting state | |
2503 | /// identifiers. | |
2504 | pub(crate) fn remap(&mut self, map: impl Fn(StateID) -> StateID) { | |
2505 | // We could loop over each state ID and call 'remap_state' here, but | |
2506 | // this is more direct: just map every transition directly. This | |
2507 | // technically might do a little extra work since the alphabet length | |
2508 | // is likely less than the stride, but if that is indeed an issue we | |
2509 | // should benchmark it and fix it. | |
2510 | for sid in self.tt.table_mut().iter_mut() { | |
2511 | *sid = map(*sid); | |
2512 | } | |
2513 | for sid in self.st.table_mut().iter_mut() { | |
2514 | *sid = map(*sid); | |
2515 | } | |
2516 | } | |
2517 | ||
2518 | /// Remap the transitions for the state given according to the function | |
2519 | /// given. This applies the given map function to every transition in the | |
2520 | /// given state and changes the transition in place to the result of the | |
2521 | /// map function for that transition. | |
2522 | pub(crate) fn remap_state( | |
2523 | &mut self, | |
2524 | id: StateID, | |
2525 | map: impl Fn(StateID) -> StateID, | |
2526 | ) { | |
2527 | self.tt.remap(id, map); | |
2528 | } | |
2529 | ||
2530 | /// Truncate the states in this DFA to the given length. | |
487cf647 FG |
2531 | /// |
2532 | /// This routine does not do anything to check the correctness of this | |
2533 | /// truncation. Callers must ensure that other states pointing to truncated | |
2534 | /// states are updated appropriately. | |
781aab86 FG |
2535 | pub(crate) fn truncate_states(&mut self, len: usize) { |
2536 | self.tt.truncate(len); | |
487cf647 FG |
2537 | } |
2538 | ||
2539 | /// Minimize this DFA in place using Hopcroft's algorithm. | |
2540 | pub(crate) fn minimize(&mut self) { | |
2541 | Minimizer::new(self).run(); | |
2542 | } | |
2543 | ||
2544 | /// Updates the match state pattern ID map to use the one provided. | |
2545 | /// | |
2546 | /// This is useful when it's convenient to manipulate matching states | |
2547 | /// (and their corresponding pattern IDs) as a map. In particular, the | |
2548 | /// representation used by a DFA for this map is not amenable to mutation, | |
2549 | /// so if things need to be changed (like when shuffling states), it's | |
2550 | /// often easier to work with the map form. | |
2551 | pub(crate) fn set_pattern_map( | |
2552 | &mut self, | |
2553 | map: &BTreeMap<StateID, Vec<PatternID>>, | |
781aab86 | 2554 | ) -> Result<(), BuildError> { |
487cf647 FG |
2555 | self.ms = self.ms.new_with_map(map)?; |
2556 | Ok(()) | |
2557 | } | |
2558 | ||
2559 | /// Find states that have a small number of non-loop transitions and mark | |
2560 | /// them as candidates for acceleration during search. | |
2561 | pub(crate) fn accelerate(&mut self) { | |
2562 | // dead and quit states can never be accelerated. | |
781aab86 | 2563 | if self.state_len() <= 2 { |
487cf647 FG |
2564 | return; |
2565 | } | |
2566 | ||
2567 | // Go through every state and record their accelerator, if possible. | |
2568 | let mut accels = BTreeMap::new(); | |
2569 | // Count the number of accelerated match, start and non-match/start | |
2570 | // states. | |
2571 | let (mut cmatch, mut cstart, mut cnormal) = (0, 0, 0); | |
2572 | for state in self.states() { | |
2573 | if let Some(accel) = state.accelerate(self.byte_classes()) { | |
781aab86 FG |
2574 | debug!( |
2575 | "accelerating full DFA state {}: {:?}", | |
2576 | state.id().as_usize(), | |
2577 | accel, | |
2578 | ); | |
487cf647 FG |
2579 | accels.insert(state.id(), accel); |
2580 | if self.is_match_state(state.id()) { | |
2581 | cmatch += 1; | |
2582 | } else if self.is_start_state(state.id()) { | |
2583 | cstart += 1; | |
2584 | } else { | |
2585 | assert!(!self.is_dead_state(state.id())); | |
2586 | assert!(!self.is_quit_state(state.id())); | |
2587 | cnormal += 1; | |
2588 | } | |
2589 | } | |
2590 | } | |
2591 | // If no states were able to be accelerated, then we're done. | |
2592 | if accels.is_empty() { | |
2593 | return; | |
2594 | } | |
2595 | let original_accels_len = accels.len(); | |
2596 | ||
2597 | // A remapper keeps track of state ID changes. Once we're done | |
2598 | // shuffling, the remapper is used to rewrite all transitions in the | |
2599 | // DFA based on the new positions of states. | |
781aab86 | 2600 | let mut remapper = Remapper::new(self); |
487cf647 FG |
2601 | |
2602 | // As we swap states, if they are match states, we need to swap their | |
2603 | // pattern ID lists too (for multi-regexes). We do this by converting | |
2604 | // the lists to an easily swappable map, and then convert back to | |
2605 | // MatchStates once we're done. | |
2606 | let mut new_matches = self.ms.to_map(self); | |
2607 | ||
2608 | // There is at least one state that gets accelerated, so these are | |
2609 | // guaranteed to get set to sensible values below. | |
2610 | self.special.min_accel = StateID::MAX; | |
2611 | self.special.max_accel = StateID::ZERO; | |
2612 | let update_special_accel = | |
2613 | |special: &mut Special, accel_id: StateID| { | |
2614 | special.min_accel = cmp::min(special.min_accel, accel_id); | |
2615 | special.max_accel = cmp::max(special.max_accel, accel_id); | |
2616 | }; | |
2617 | ||
2618 | // Start by shuffling match states. Any match states that are | |
2619 | // accelerated get moved to the end of the match state range. | |
2620 | if cmatch > 0 && self.special.matches() { | |
2621 | // N.B. special.{min,max}_match do not need updating, since the | |
2622 | // range/number of match states does not change. Only the ordering | |
2623 | // of match states may change. | |
2624 | let mut next_id = self.special.max_match; | |
2625 | let mut cur_id = next_id; | |
2626 | while cur_id >= self.special.min_match { | |
2627 | if let Some(accel) = accels.remove(&cur_id) { | |
2628 | accels.insert(next_id, accel); | |
2629 | update_special_accel(&mut self.special, next_id); | |
2630 | ||
2631 | // No need to do any actual swapping for equivalent IDs. | |
2632 | if cur_id != next_id { | |
2633 | remapper.swap(self, cur_id, next_id); | |
2634 | ||
2635 | // Swap pattern IDs for match states. | |
2636 | let cur_pids = new_matches.remove(&cur_id).unwrap(); | |
2637 | let next_pids = new_matches.remove(&next_id).unwrap(); | |
2638 | new_matches.insert(cur_id, next_pids); | |
2639 | new_matches.insert(next_id, cur_pids); | |
2640 | } | |
2641 | next_id = self.tt.prev_state_id(next_id); | |
2642 | } | |
2643 | cur_id = self.tt.prev_state_id(cur_id); | |
2644 | } | |
2645 | } | |
2646 | ||
2647 | // This is where it gets tricky. Without acceleration, start states | |
2648 | // normally come right after match states. But we want accelerated | |
2649 | // states to be a single contiguous range (to make it very fast | |
2650 | // to determine whether a state *is* accelerated), while also keeping | |
2651 | // match and starting states as contiguous ranges for the same reason. | |
2652 | // So what we do here is shuffle states such that it looks like this: | |
2653 | // | |
2654 | // DQMMMMAAAAASSSSSSNNNNNNN | |
2655 | // | | | |
2656 | // |---------| | |
2657 | // accelerated states | |
2658 | // | |
2659 | // Where: | |
2660 | // D - dead state | |
2661 | // Q - quit state | |
2662 | // M - match state (may be accelerated) | |
2663 | // A - normal state that is accelerated | |
2664 | // S - start state (may be accelerated) | |
2665 | // N - normal state that is NOT accelerated | |
2666 | // | |
2667 | // We implement this by shuffling states, which is done by a sequence | |
2668 | // of pairwise swaps. We start by looking at all normal states to be | |
2669 | // accelerated. When we find one, we swap it with the earliest starting | |
2670 | // state, and then swap that with the earliest normal state. This | |
2671 | // preserves the contiguous property. | |
2672 | // | |
2673 | // Once we're done looking for accelerated normal states, now we look | |
2674 | // for accelerated starting states by moving them to the beginning | |
2675 | // of the starting state range (just like we moved accelerated match | |
2676 | // states to the end of the matching state range). | |
2677 | // | |
2678 | // For a more detailed/different perspective on this, see the docs | |
2679 | // in dfa/special.rs. | |
2680 | if cnormal > 0 { | |
2681 | // our next available starting and normal states for swapping. | |
2682 | let mut next_start_id = self.special.min_start; | |
781aab86 | 2683 | let mut cur_id = self.to_state_id(self.state_len() - 1); |
487cf647 FG |
2684 | // This is guaranteed to exist since cnormal > 0. |
2685 | let mut next_norm_id = | |
2686 | self.tt.next_state_id(self.special.max_start); | |
2687 | while cur_id >= next_norm_id { | |
2688 | if let Some(accel) = accels.remove(&cur_id) { | |
2689 | remapper.swap(self, next_start_id, cur_id); | |
2690 | remapper.swap(self, next_norm_id, cur_id); | |
2691 | // Keep our accelerator map updated with new IDs if the | |
2692 | // states we swapped were also accelerated. | |
2693 | if let Some(accel2) = accels.remove(&next_norm_id) { | |
2694 | accels.insert(cur_id, accel2); | |
2695 | } | |
2696 | if let Some(accel2) = accels.remove(&next_start_id) { | |
2697 | accels.insert(next_norm_id, accel2); | |
2698 | } | |
2699 | accels.insert(next_start_id, accel); | |
2700 | update_special_accel(&mut self.special, next_start_id); | |
2701 | // Our start range shifts one to the right now. | |
2702 | self.special.min_start = | |
2703 | self.tt.next_state_id(self.special.min_start); | |
2704 | self.special.max_start = | |
2705 | self.tt.next_state_id(self.special.max_start); | |
2706 | next_start_id = self.tt.next_state_id(next_start_id); | |
2707 | next_norm_id = self.tt.next_state_id(next_norm_id); | |
2708 | } | |
2709 | // This is pretty tricky, but if our 'next_norm_id' state also | |
2710 | // happened to be accelerated, then the result is that it is | |
2711 | // now in the position of cur_id, so we need to consider it | |
2712 | // again. This loop is still guaranteed to terminate though, | |
2713 | // because when accels contains cur_id, we're guaranteed to | |
2714 | // increment next_norm_id even if cur_id remains unchanged. | |
2715 | if !accels.contains_key(&cur_id) { | |
2716 | cur_id = self.tt.prev_state_id(cur_id); | |
2717 | } | |
2718 | } | |
2719 | } | |
2720 | // Just like we did for match states, but we want to move accelerated | |
2721 | // start states to the beginning of the range instead of the end. | |
2722 | if cstart > 0 { | |
2723 | // N.B. special.{min,max}_start do not need updating, since the | |
2724 | // range/number of start states does not change at this point. Only | |
2725 | // the ordering of start states may change. | |
2726 | let mut next_id = self.special.min_start; | |
2727 | let mut cur_id = next_id; | |
2728 | while cur_id <= self.special.max_start { | |
2729 | if let Some(accel) = accels.remove(&cur_id) { | |
2730 | remapper.swap(self, cur_id, next_id); | |
2731 | accels.insert(next_id, accel); | |
2732 | update_special_accel(&mut self.special, next_id); | |
2733 | next_id = self.tt.next_state_id(next_id); | |
2734 | } | |
2735 | cur_id = self.tt.next_state_id(cur_id); | |
2736 | } | |
2737 | } | |
2738 | ||
2739 | // Remap all transitions in our DFA and assert some things. | |
2740 | remapper.remap(self); | |
2741 | // This unwrap is OK because acceleration never changes the number of | |
2742 | // match states or patterns in those match states. Since acceleration | |
2743 | // runs after the pattern map has been set at least once, we know that | |
2744 | // our match states cannot error. | |
2745 | self.set_pattern_map(&new_matches).unwrap(); | |
2746 | self.special.set_max(); | |
2747 | self.special.validate().expect("special state ranges should validate"); | |
2748 | self.special | |
781aab86 | 2749 | .validate_state_len(self.state_len(), self.stride2()) |
487cf647 | 2750 | .expect( |
781aab86 | 2751 | "special state ranges should be consistent with state length", |
487cf647 FG |
2752 | ); |
2753 | assert_eq!( | |
2754 | self.special.accel_len(self.stride()), | |
2755 | // We record the number of accelerated states initially detected | |
2756 | // since the accels map is itself mutated in the process above. | |
2757 | // If mutated incorrectly, its size may change, and thus can't be | |
2758 | // trusted as a source of truth of how many accelerated states we | |
2759 | // expected there to be. | |
2760 | original_accels_len, | |
2761 | "mismatch with expected number of accelerated states", | |
2762 | ); | |
2763 | ||
2764 | // And finally record our accelerators. We kept our accels map updated | |
2765 | // as we shuffled states above, so the accelerators should now | |
2766 | // correspond to a contiguous range in the state ID space. (Which we | |
2767 | // assert.) | |
2768 | let mut prev: Option<StateID> = None; | |
2769 | for (id, accel) in accels { | |
2770 | assert!(prev.map_or(true, |p| self.tt.next_state_id(p) == id)); | |
2771 | prev = Some(id); | |
2772 | self.accels.add(accel); | |
2773 | } | |
2774 | } | |
2775 | ||
2776 | /// Shuffle the states in this DFA so that starting states, match | |
2777 | /// states and accelerated states are all contiguous. | |
2778 | /// | |
2779 | /// See dfa/special.rs for more details. | |
2780 | pub(crate) fn shuffle( | |
2781 | &mut self, | |
2782 | mut matches: BTreeMap<StateID, Vec<PatternID>>, | |
781aab86 | 2783 | ) -> Result<(), BuildError> { |
487cf647 | 2784 | // The determinizer always adds a quit state and it is always second. |
781aab86 | 2785 | self.special.quit_id = self.to_state_id(1); |
487cf647 FG |
2786 | // If all we have are the dead and quit states, then we're done and |
2787 | // the DFA will never produce a match. | |
781aab86 | 2788 | if self.state_len() <= 2 { |
487cf647 FG |
2789 | self.special.set_max(); |
2790 | return Ok(()); | |
2791 | } | |
2792 | ||
781aab86 FG |
2793 | // Collect all our non-DEAD start states into a convenient set and |
2794 | // confirm there is no overlap with match states. In the classicl DFA | |
2795 | // construction, start states can be match states. But because of | |
2796 | // look-around, we delay all matches by a byte, which prevents start | |
2797 | // states from being match states. | |
487cf647 FG |
2798 | let mut is_start: BTreeSet<StateID> = BTreeSet::new(); |
2799 | for (start_id, _, _) in self.starts() { | |
781aab86 FG |
2800 | // If a starting configuration points to a DEAD state, then we |
2801 | // don't want to shuffle it. The DEAD state is always the first | |
2802 | // state with ID=0. So we can just leave it be. | |
2803 | if start_id == DEAD { | |
2804 | continue; | |
2805 | } | |
487cf647 FG |
2806 | assert!( |
2807 | !matches.contains_key(&start_id), | |
2808 | "{:?} is both a start and a match state, which is not allowed", | |
2809 | start_id, | |
2810 | ); | |
2811 | is_start.insert(start_id); | |
2812 | } | |
2813 | ||
2814 | // We implement shuffling by a sequence of pairwise swaps of states. | |
2815 | // Since we have a number of things referencing states via their | |
2816 | // IDs and swapping them changes their IDs, we need to record every | |
2817 | // swap we make so that we can remap IDs. The remapper handles this | |
2818 | // book-keeping for us. | |
781aab86 | 2819 | let mut remapper = Remapper::new(self); |
487cf647 FG |
2820 | |
2821 | // Shuffle matching states. | |
2822 | if matches.is_empty() { | |
2823 | self.special.min_match = DEAD; | |
2824 | self.special.max_match = DEAD; | |
2825 | } else { | |
2826 | // The determinizer guarantees that the first two states are the | |
2827 | // dead and quit states, respectively. We want our match states to | |
2828 | // come right after quit. | |
781aab86 | 2829 | let mut next_id = self.to_state_id(2); |
487cf647 FG |
2830 | let mut new_matches = BTreeMap::new(); |
2831 | self.special.min_match = next_id; | |
2832 | for (id, pids) in matches { | |
2833 | remapper.swap(self, next_id, id); | |
2834 | new_matches.insert(next_id, pids); | |
2835 | // If we swapped a start state, then update our set. | |
2836 | if is_start.contains(&next_id) { | |
2837 | is_start.remove(&next_id); | |
2838 | is_start.insert(id); | |
2839 | } | |
2840 | next_id = self.tt.next_state_id(next_id); | |
2841 | } | |
2842 | matches = new_matches; | |
2843 | self.special.max_match = cmp::max( | |
2844 | self.special.min_match, | |
2845 | self.tt.prev_state_id(next_id), | |
2846 | ); | |
2847 | } | |
2848 | ||
2849 | // Shuffle starting states. | |
2850 | { | |
781aab86 | 2851 | let mut next_id = self.to_state_id(2); |
487cf647 FG |
2852 | if self.special.matches() { |
2853 | next_id = self.tt.next_state_id(self.special.max_match); | |
2854 | } | |
2855 | self.special.min_start = next_id; | |
2856 | for id in is_start { | |
2857 | remapper.swap(self, next_id, id); | |
2858 | next_id = self.tt.next_state_id(next_id); | |
2859 | } | |
2860 | self.special.max_start = cmp::max( | |
2861 | self.special.min_start, | |
2862 | self.tt.prev_state_id(next_id), | |
2863 | ); | |
2864 | } | |
2865 | ||
2866 | // Finally remap all transitions in our DFA. | |
2867 | remapper.remap(self); | |
2868 | self.set_pattern_map(&matches)?; | |
2869 | self.special.set_max(); | |
2870 | self.special.validate().expect("special state ranges should validate"); | |
2871 | self.special | |
781aab86 | 2872 | .validate_state_len(self.state_len(), self.stride2()) |
487cf647 | 2873 | .expect( |
781aab86 | 2874 | "special state ranges should be consistent with state length", |
487cf647 FG |
2875 | ); |
2876 | Ok(()) | |
2877 | } | |
487cf647 | 2878 | |
781aab86 FG |
2879 | /// Checks whether there are universal start states (both anchored and |
2880 | /// unanchored), and if so, sets the relevant fields to the start state | |
2881 | /// IDs. | |
2882 | /// | |
2883 | /// Universal start states occur precisely when the all patterns in the | |
2884 | /// DFA have no look-around assertions in their prefix. | |
2885 | fn set_universal_starts(&mut self) { | |
2886 | assert_eq!(6, Start::len(), "expected 6 start configurations"); | |
2887 | ||
2888 | let start_id = |dfa: &mut OwnedDFA, inp: &Input<'_>, start: Start| { | |
2889 | // This OK because we only call 'start' under conditions | |
2890 | // in which we know it will succeed. | |
2891 | dfa.st.start(inp, start).expect("valid Input configuration") | |
2892 | }; | |
2893 | if self.start_kind().has_unanchored() { | |
2894 | let inp = Input::new("").anchored(Anchored::No); | |
2895 | let sid = start_id(self, &inp, Start::NonWordByte); | |
2896 | if sid == start_id(self, &inp, Start::WordByte) | |
2897 | && sid == start_id(self, &inp, Start::Text) | |
2898 | && sid == start_id(self, &inp, Start::LineLF) | |
2899 | && sid == start_id(self, &inp, Start::LineCR) | |
2900 | && sid == start_id(self, &inp, Start::CustomLineTerminator) | |
2901 | { | |
2902 | self.st.universal_start_unanchored = Some(sid); | |
2903 | } | |
2904 | } | |
2905 | if self.start_kind().has_anchored() { | |
2906 | let inp = Input::new("").anchored(Anchored::Yes); | |
2907 | let sid = start_id(self, &inp, Start::NonWordByte); | |
2908 | if sid == start_id(self, &inp, Start::WordByte) | |
2909 | && sid == start_id(self, &inp, Start::Text) | |
2910 | && sid == start_id(self, &inp, Start::LineLF) | |
2911 | && sid == start_id(self, &inp, Start::LineCR) | |
2912 | && sid == start_id(self, &inp, Start::CustomLineTerminator) | |
2913 | { | |
2914 | self.st.universal_start_anchored = Some(sid); | |
2915 | } | |
2916 | } | |
487cf647 | 2917 | } |
781aab86 | 2918 | } |
487cf647 | 2919 | |
781aab86 FG |
2920 | // A variety of generic internal methods for accessing DFA internals. |
2921 | impl<T: AsRef<[u32]>> DFA<T> { | |
487cf647 FG |
2922 | /// Return the info about special states. |
2923 | pub(crate) fn special(&self) -> &Special { | |
2924 | &self.special | |
2925 | } | |
2926 | ||
2927 | /// Return the info about special states as a mutable borrow. | |
781aab86 | 2928 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
2929 | pub(crate) fn special_mut(&mut self) -> &mut Special { |
2930 | &mut self.special | |
2931 | } | |
2932 | ||
781aab86 FG |
2933 | /// Returns the quit set (may be empty) used by this DFA. |
2934 | pub(crate) fn quitset(&self) -> &ByteSet { | |
2935 | &self.quitset | |
2936 | } | |
2937 | ||
2938 | /// Returns the flags for this DFA. | |
2939 | pub(crate) fn flags(&self) -> &Flags { | |
2940 | &self.flags | |
2941 | } | |
2942 | ||
487cf647 FG |
2943 | /// Returns an iterator over all states in this DFA. |
2944 | /// | |
2945 | /// This iterator yields a tuple for each state. The first element of the | |
2946 | /// tuple corresponds to a state's identifier, and the second element | |
2947 | /// corresponds to the state itself (comprised of its transitions). | |
2948 | pub(crate) fn states(&self) -> StateIter<'_, T> { | |
2949 | self.tt.states() | |
2950 | } | |
2951 | ||
2952 | /// Return the total number of states in this DFA. Every DFA has at least | |
2953 | /// 1 state, even the empty DFA. | |
781aab86 FG |
2954 | pub(crate) fn state_len(&self) -> usize { |
2955 | self.tt.len() | |
487cf647 FG |
2956 | } |
2957 | ||
2958 | /// Return an iterator over all pattern IDs for the given match state. | |
2959 | /// | |
2960 | /// If the given state is not a match state, then this panics. | |
781aab86 | 2961 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
2962 | pub(crate) fn pattern_id_slice(&self, id: StateID) -> &[PatternID] { |
2963 | assert!(self.is_match_state(id)); | |
2964 | self.ms.pattern_id_slice(self.match_state_index(id)) | |
2965 | } | |
2966 | ||
2967 | /// Return the total number of pattern IDs for the given match state. | |
2968 | /// | |
2969 | /// If the given state is not a match state, then this panics. | |
2970 | pub(crate) fn match_pattern_len(&self, id: StateID) -> usize { | |
2971 | assert!(self.is_match_state(id)); | |
2972 | self.ms.pattern_len(self.match_state_index(id)) | |
2973 | } | |
2974 | ||
2975 | /// Returns the total number of patterns matched by this DFA. | |
781aab86 FG |
2976 | pub(crate) fn pattern_len(&self) -> usize { |
2977 | self.ms.pattern_len | |
487cf647 FG |
2978 | } |
2979 | ||
2980 | /// Returns a map from match state ID to a list of pattern IDs that match | |
2981 | /// in that state. | |
781aab86 | 2982 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
2983 | pub(crate) fn pattern_map(&self) -> BTreeMap<StateID, Vec<PatternID>> { |
2984 | self.ms.to_map(self) | |
2985 | } | |
2986 | ||
2987 | /// Returns the ID of the quit state for this DFA. | |
781aab86 | 2988 | #[cfg(feature = "dfa-build")] |
487cf647 | 2989 | pub(crate) fn quit_id(&self) -> StateID { |
781aab86 | 2990 | self.to_state_id(1) |
487cf647 FG |
2991 | } |
2992 | ||
2993 | /// Convert the given state identifier to the state's index. The state's | |
2994 | /// index corresponds to the position in which it appears in the transition | |
2995 | /// table. When a DFA is NOT premultiplied, then a state's identifier is | |
2996 | /// also its index. When a DFA is premultiplied, then a state's identifier | |
2997 | /// is equal to `index * alphabet_len`. This routine reverses that. | |
2998 | pub(crate) fn to_index(&self, id: StateID) -> usize { | |
2999 | self.tt.to_index(id) | |
3000 | } | |
3001 | ||
781aab86 | 3002 | /// Convert an index to a state (in the range 0..self.state_len()) to an |
487cf647 FG |
3003 | /// actual state identifier. |
3004 | /// | |
3005 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state | |
3006 | /// to some other information (such as a remapped state ID). | |
781aab86 FG |
3007 | #[cfg(feature = "dfa-build")] |
3008 | pub(crate) fn to_state_id(&self, index: usize) -> StateID { | |
3009 | self.tt.to_state_id(index) | |
487cf647 FG |
3010 | } |
3011 | ||
3012 | /// Return the table of state IDs for this DFA's start states. | |
3013 | pub(crate) fn starts(&self) -> StartStateIter<'_> { | |
3014 | self.st.iter() | |
3015 | } | |
3016 | ||
3017 | /// Returns the index of the match state for the given ID. If the | |
3018 | /// given ID does not correspond to a match state, then this may | |
3019 | /// panic or produce an incorrect result. | |
781aab86 | 3020 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3021 | fn match_state_index(&self, id: StateID) -> usize { |
3022 | debug_assert!(self.is_match_state(id)); | |
3023 | // This is one of the places where we rely on the fact that match | |
3024 | // states are contiguous in the transition table. Namely, that the | |
781aab86 | 3025 | // first match state ID always corresponds to dfa.special.min_match. |
487cf647 FG |
3026 | // From there, since we know the stride, we can compute the overall |
3027 | // index of any match state given the match state's ID. | |
3028 | let min = self.special().min_match.as_usize(); | |
3029 | // CORRECTNESS: We're allowed to produce an incorrect result or panic, | |
3030 | // so both the subtraction and the unchecked StateID construction is | |
3031 | // OK. | |
3032 | self.to_index(StateID::new_unchecked(id.as_usize() - min)) | |
3033 | } | |
3034 | ||
3035 | /// Returns the index of the accelerator state for the given ID. If the | |
3036 | /// given ID does not correspond to an accelerator state, then this may | |
3037 | /// panic or produce an incorrect result. | |
3038 | fn accelerator_index(&self, id: StateID) -> usize { | |
3039 | let min = self.special().min_accel.as_usize(); | |
3040 | // CORRECTNESS: We're allowed to produce an incorrect result or panic, | |
3041 | // so both the subtraction and the unchecked StateID construction is | |
3042 | // OK. | |
3043 | self.to_index(StateID::new_unchecked(id.as_usize() - min)) | |
3044 | } | |
3045 | ||
3046 | /// Return the accelerators for this DFA. | |
3047 | fn accels(&self) -> Accels<&[u32]> { | |
3048 | self.accels.as_ref() | |
3049 | } | |
3050 | ||
3051 | /// Return this DFA's transition table as a slice. | |
3052 | fn trans(&self) -> &[StateID] { | |
3053 | self.tt.table() | |
3054 | } | |
3055 | } | |
3056 | ||
3057 | impl<T: AsRef<[u32]>> fmt::Debug for DFA<T> { | |
3058 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { | |
3059 | writeln!(f, "dense::DFA(")?; | |
3060 | for state in self.states() { | |
3061 | fmt_state_indicator(f, self, state.id())?; | |
3062 | let id = if f.alternate() { | |
3063 | state.id().as_usize() | |
3064 | } else { | |
3065 | self.to_index(state.id()) | |
3066 | }; | |
3067 | write!(f, "{:06?}: ", id)?; | |
3068 | state.fmt(f)?; | |
3069 | write!(f, "\n")?; | |
3070 | } | |
3071 | writeln!(f, "")?; | |
781aab86 | 3072 | for (i, (start_id, anchored, sty)) in self.starts().enumerate() { |
487cf647 FG |
3073 | let id = if f.alternate() { |
3074 | start_id.as_usize() | |
3075 | } else { | |
3076 | self.to_index(start_id) | |
3077 | }; | |
3078 | if i % self.st.stride == 0 { | |
781aab86 FG |
3079 | match anchored { |
3080 | Anchored::No => writeln!(f, "START-GROUP(unanchored)")?, | |
3081 | Anchored::Yes => writeln!(f, "START-GROUP(anchored)")?, | |
3082 | Anchored::Pattern(pid) => { | |
487cf647 FG |
3083 | writeln!(f, "START_GROUP(pattern: {:?})", pid)? |
3084 | } | |
3085 | } | |
3086 | } | |
3087 | writeln!(f, " {:?} => {:06?}", sty, id)?; | |
3088 | } | |
781aab86 | 3089 | if self.pattern_len() > 1 { |
487cf647 | 3090 | writeln!(f, "")?; |
781aab86 | 3091 | for i in 0..self.ms.len() { |
487cf647 FG |
3092 | let id = self.ms.match_state_id(self, i); |
3093 | let id = if f.alternate() { | |
3094 | id.as_usize() | |
3095 | } else { | |
3096 | self.to_index(id) | |
3097 | }; | |
3098 | write!(f, "MATCH({:06?}): ", id)?; | |
3099 | for (i, &pid) in self.ms.pattern_id_slice(i).iter().enumerate() | |
3100 | { | |
3101 | if i > 0 { | |
3102 | write!(f, ", ")?; | |
3103 | } | |
3104 | write!(f, "{:?}", pid)?; | |
3105 | } | |
3106 | writeln!(f, "")?; | |
3107 | } | |
3108 | } | |
781aab86 FG |
3109 | writeln!(f, "state length: {:?}", self.state_len())?; |
3110 | writeln!(f, "pattern length: {:?}", self.pattern_len())?; | |
3111 | writeln!(f, "flags: {:?}", self.flags)?; | |
487cf647 FG |
3112 | writeln!(f, ")")?; |
3113 | Ok(()) | |
3114 | } | |
3115 | } | |
3116 | ||
781aab86 | 3117 | // SAFETY: We assert that our implementation of each method is correct. |
487cf647 | 3118 | unsafe impl<T: AsRef<[u32]>> Automaton for DFA<T> { |
781aab86 | 3119 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3120 | fn is_special_state(&self, id: StateID) -> bool { |
3121 | self.special.is_special_state(id) | |
3122 | } | |
3123 | ||
781aab86 | 3124 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3125 | fn is_dead_state(&self, id: StateID) -> bool { |
3126 | self.special.is_dead_state(id) | |
3127 | } | |
3128 | ||
781aab86 | 3129 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3130 | fn is_quit_state(&self, id: StateID) -> bool { |
3131 | self.special.is_quit_state(id) | |
3132 | } | |
3133 | ||
781aab86 | 3134 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3135 | fn is_match_state(&self, id: StateID) -> bool { |
3136 | self.special.is_match_state(id) | |
3137 | } | |
3138 | ||
781aab86 | 3139 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3140 | fn is_start_state(&self, id: StateID) -> bool { |
3141 | self.special.is_start_state(id) | |
3142 | } | |
3143 | ||
781aab86 | 3144 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3145 | fn is_accel_state(&self, id: StateID) -> bool { |
3146 | self.special.is_accel_state(id) | |
3147 | } | |
3148 | ||
781aab86 | 3149 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3150 | fn next_state(&self, current: StateID, input: u8) -> StateID { |
3151 | let input = self.byte_classes().get(input); | |
3152 | let o = current.as_usize() + usize::from(input); | |
3153 | self.trans()[o] | |
3154 | } | |
3155 | ||
781aab86 | 3156 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3157 | unsafe fn next_state_unchecked( |
3158 | &self, | |
3159 | current: StateID, | |
781aab86 | 3160 | byte: u8, |
487cf647 | 3161 | ) -> StateID { |
781aab86 FG |
3162 | // We don't (or shouldn't) need an unchecked variant for the byte |
3163 | // class mapping, since bound checks should be omitted automatically | |
3164 | // by virtue of its representation. If this ends up not being true as | |
3165 | // confirmed by codegen, please file an issue. ---AG | |
3166 | let class = self.byte_classes().get(byte); | |
3167 | let o = current.as_usize() + usize::from(class); | |
3168 | let next = *self.trans().get_unchecked(o); | |
3169 | next | |
487cf647 FG |
3170 | } |
3171 | ||
781aab86 | 3172 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3173 | fn next_eoi_state(&self, current: StateID) -> StateID { |
3174 | let eoi = self.byte_classes().eoi().as_usize(); | |
3175 | let o = current.as_usize() + eoi; | |
3176 | self.trans()[o] | |
3177 | } | |
3178 | ||
781aab86 FG |
3179 | #[cfg_attr(feature = "perf-inline", inline(always))] |
3180 | fn pattern_len(&self) -> usize { | |
3181 | self.ms.pattern_len | |
487cf647 FG |
3182 | } |
3183 | ||
781aab86 FG |
3184 | #[cfg_attr(feature = "perf-inline", inline(always))] |
3185 | fn match_len(&self, id: StateID) -> usize { | |
487cf647 FG |
3186 | self.match_pattern_len(id) |
3187 | } | |
3188 | ||
781aab86 | 3189 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3190 | fn match_pattern(&self, id: StateID, match_index: usize) -> PatternID { |
3191 | // This is an optimization for the very common case of a DFA with a | |
3192 | // single pattern. This conditional avoids a somewhat more costly path | |
3193 | // that finds the pattern ID from the state machine, which requires | |
3194 | // a bit of slicing/pointer-chasing. This optimization tends to only | |
3195 | // matter when matches are frequent. | |
781aab86 | 3196 | if self.ms.pattern_len == 1 { |
487cf647 FG |
3197 | return PatternID::ZERO; |
3198 | } | |
3199 | let state_index = self.match_state_index(id); | |
3200 | self.ms.pattern_id(state_index, match_index) | |
3201 | } | |
3202 | ||
781aab86 FG |
3203 | #[cfg_attr(feature = "perf-inline", inline(always))] |
3204 | fn has_empty(&self) -> bool { | |
3205 | self.flags.has_empty | |
3206 | } | |
3207 | ||
3208 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
3209 | fn is_utf8(&self) -> bool { | |
3210 | self.flags.is_utf8 | |
3211 | } | |
3212 | ||
3213 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
3214 | fn is_always_start_anchored(&self) -> bool { | |
3215 | self.flags.is_always_start_anchored | |
3216 | } | |
3217 | ||
3218 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
487cf647 FG |
3219 | fn start_state_forward( |
3220 | &self, | |
781aab86 FG |
3221 | input: &Input<'_>, |
3222 | ) -> Result<StateID, MatchError> { | |
3223 | if !self.quitset.is_empty() && input.start() > 0 { | |
3224 | let offset = input.start() - 1; | |
3225 | let byte = input.haystack()[offset]; | |
3226 | if self.quitset.contains(byte) { | |
3227 | return Err(MatchError::quit(byte, offset)); | |
3228 | } | |
3229 | } | |
3230 | let start = self.st.start_map.fwd(&input); | |
3231 | self.st.start(input, start) | |
487cf647 FG |
3232 | } |
3233 | ||
781aab86 | 3234 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
3235 | fn start_state_reverse( |
3236 | &self, | |
781aab86 FG |
3237 | input: &Input<'_>, |
3238 | ) -> Result<StateID, MatchError> { | |
3239 | if !self.quitset.is_empty() && input.end() < input.haystack().len() { | |
3240 | let offset = input.end(); | |
3241 | let byte = input.haystack()[offset]; | |
3242 | if self.quitset.contains(byte) { | |
3243 | return Err(MatchError::quit(byte, offset)); | |
3244 | } | |
3245 | } | |
3246 | let start = self.st.start_map.rev(&input); | |
3247 | self.st.start(input, start) | |
487cf647 FG |
3248 | } |
3249 | ||
781aab86 FG |
3250 | #[cfg_attr(feature = "perf-inline", inline(always))] |
3251 | fn universal_start_state(&self, mode: Anchored) -> Option<StateID> { | |
3252 | match mode { | |
3253 | Anchored::No => self.st.universal_start_unanchored, | |
3254 | Anchored::Yes => self.st.universal_start_anchored, | |
3255 | Anchored::Pattern(_) => None, | |
3256 | } | |
3257 | } | |
3258 | ||
3259 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
487cf647 FG |
3260 | fn accelerator(&self, id: StateID) -> &[u8] { |
3261 | if !self.is_accel_state(id) { | |
3262 | return &[]; | |
3263 | } | |
3264 | self.accels.needles(self.accelerator_index(id)) | |
3265 | } | |
781aab86 FG |
3266 | |
3267 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
3268 | fn get_prefilter(&self) -> Option<&Prefilter> { | |
3269 | self.pre.as_ref() | |
3270 | } | |
487cf647 FG |
3271 | } |
3272 | ||
3273 | /// The transition table portion of a dense DFA. | |
3274 | /// | |
3275 | /// The transition table is the core part of the DFA in that it describes how | |
3276 | /// to move from one state to another based on the input sequence observed. | |
3277 | #[derive(Clone)] | |
3278 | pub(crate) struct TransitionTable<T> { | |
3279 | /// A contiguous region of memory representing the transition table in | |
3280 | /// row-major order. The representation is dense. That is, every state | |
3281 | /// has precisely the same number of transitions. The maximum number of | |
3282 | /// transitions per state is 257 (256 for each possible byte value, plus 1 | |
3283 | /// for the special EOI transition). If a DFA has been instructed to use | |
3284 | /// byte classes (the default), then the number of transitions is usually | |
3285 | /// substantially fewer. | |
3286 | /// | |
3287 | /// In practice, T is either `Vec<u32>` or `&[u32]`. | |
3288 | table: T, | |
3289 | /// A set of equivalence classes, where a single equivalence class | |
3290 | /// represents a set of bytes that never discriminate between a match | |
3291 | /// and a non-match in the DFA. Each equivalence class corresponds to a | |
3292 | /// single character in this DFA's alphabet, where the maximum number of | |
3293 | /// characters is 257 (each possible value of a byte plus the special | |
3294 | /// EOI transition). Consequently, the number of equivalence classes | |
3295 | /// corresponds to the number of transitions for each DFA state. Note | |
3296 | /// though that the *space* used by each DFA state in the transition table | |
3297 | /// may be larger. The total space used by each DFA state is known as the | |
3298 | /// stride. | |
3299 | /// | |
3300 | /// The only time the number of equivalence classes is fewer than 257 is if | |
3301 | /// the DFA's kind uses byte classes (which is the default). Equivalence | |
3302 | /// classes should generally only be disabled when debugging, so that | |
3303 | /// the transitions themselves aren't obscured. Disabling them has no | |
3304 | /// other benefit, since the equivalence class map is always used while | |
3305 | /// searching. In the vast majority of cases, the number of equivalence | |
3306 | /// classes is substantially smaller than 257, particularly when large | |
3307 | /// Unicode classes aren't used. | |
3308 | classes: ByteClasses, | |
3309 | /// The stride of each DFA state, expressed as a power-of-two exponent. | |
3310 | /// | |
3311 | /// The stride of a DFA corresponds to the total amount of space used by | |
3312 | /// each DFA state in the transition table. This may be bigger than the | |
3313 | /// size of a DFA's alphabet, since the stride is always the smallest | |
3314 | /// power of two greater than or equal to the alphabet size. | |
3315 | /// | |
3316 | /// While this wastes space, this avoids the need for integer division | |
3317 | /// to convert between premultiplied state IDs and their corresponding | |
3318 | /// indices. Instead, we can use simple bit-shifts. | |
3319 | /// | |
3320 | /// See the docs for the `stride2` method for more details. | |
3321 | /// | |
3322 | /// The minimum `stride2` value is `1` (corresponding to a stride of `2`) | |
3323 | /// while the maximum `stride2` value is `9` (corresponding to a stride of | |
3324 | /// `512`). The maximum is not `8` since the maximum alphabet size is `257` | |
3325 | /// when accounting for the special EOI transition. However, an alphabet | |
3326 | /// length of that size is exceptionally rare since the alphabet is shrunk | |
3327 | /// into equivalence classes. | |
3328 | stride2: usize, | |
3329 | } | |
3330 | ||
3331 | impl<'a> TransitionTable<&'a [u32]> { | |
3332 | /// Deserialize a transition table starting at the beginning of `slice`. | |
3333 | /// Upon success, return the total number of bytes read along with the | |
3334 | /// transition table. | |
3335 | /// | |
3336 | /// If there was a problem deserializing any part of the transition table, | |
3337 | /// then this returns an error. Notably, if the given slice does not have | |
3338 | /// the same alignment as `StateID`, then this will return an error (among | |
3339 | /// other possible errors). | |
3340 | /// | |
3341 | /// This is guaranteed to execute in constant time. | |
3342 | /// | |
3343 | /// # Safety | |
3344 | /// | |
781aab86 | 3345 | /// This routine is not safe because it does not check the validity of the |
487cf647 FG |
3346 | /// transition table itself. In particular, the transition table can be |
3347 | /// quite large, so checking its validity can be somewhat expensive. An | |
3348 | /// invalid transition table is not safe because other code may rely on the | |
3349 | /// transition table being correct (such as explicit bounds check elision). | |
3350 | /// Therefore, an invalid transition table can lead to undefined behavior. | |
3351 | /// | |
3352 | /// Callers that use this function must either pass on the safety invariant | |
3353 | /// or guarantee that the bytes given contain a valid transition table. | |
3354 | /// This guarantee is upheld by the bytes written by `write_to`. | |
3355 | unsafe fn from_bytes_unchecked( | |
3356 | mut slice: &'a [u8], | |
3357 | ) -> Result<(TransitionTable<&'a [u32]>, usize), DeserializeError> { | |
781aab86 | 3358 | let slice_start = slice.as_ptr().as_usize(); |
487cf647 | 3359 | |
781aab86 FG |
3360 | let (state_len, nr) = |
3361 | wire::try_read_u32_as_usize(slice, "state length")?; | |
487cf647 FG |
3362 | slice = &slice[nr..]; |
3363 | ||
781aab86 | 3364 | let (stride2, nr) = wire::try_read_u32_as_usize(slice, "stride2")?; |
487cf647 FG |
3365 | slice = &slice[nr..]; |
3366 | ||
3367 | let (classes, nr) = ByteClasses::from_bytes(slice)?; | |
3368 | slice = &slice[nr..]; | |
3369 | ||
3370 | // The alphabet length (determined by the byte class map) cannot be | |
3371 | // bigger than the stride (total space used by each DFA state). | |
3372 | if stride2 > 9 { | |
3373 | return Err(DeserializeError::generic( | |
3374 | "dense DFA has invalid stride2 (too big)", | |
3375 | )); | |
3376 | } | |
3377 | // It also cannot be zero, since even a DFA that never matches anything | |
3378 | // has a non-zero number of states with at least two equivalence | |
3379 | // classes: one for all 256 byte values and another for the EOI | |
3380 | // sentinel. | |
3381 | if stride2 < 1 { | |
3382 | return Err(DeserializeError::generic( | |
3383 | "dense DFA has invalid stride2 (too small)", | |
3384 | )); | |
3385 | } | |
3386 | // This is OK since 1 <= stride2 <= 9. | |
3387 | let stride = | |
3388 | 1usize.checked_shl(u32::try_from(stride2).unwrap()).unwrap(); | |
3389 | if classes.alphabet_len() > stride { | |
3390 | return Err(DeserializeError::generic( | |
3391 | "alphabet size cannot be bigger than transition table stride", | |
3392 | )); | |
3393 | } | |
3394 | ||
781aab86 FG |
3395 | let trans_len = |
3396 | wire::shl(state_len, stride2, "dense table transition length")?; | |
3397 | let table_bytes_len = wire::mul( | |
3398 | trans_len, | |
487cf647 | 3399 | StateID::SIZE, |
781aab86 | 3400 | "dense table state byte length", |
487cf647 | 3401 | )?; |
781aab86 FG |
3402 | wire::check_slice_len(slice, table_bytes_len, "transition table")?; |
3403 | wire::check_alignment::<StateID>(slice)?; | |
487cf647 FG |
3404 | let table_bytes = &slice[..table_bytes_len]; |
3405 | slice = &slice[table_bytes_len..]; | |
3406 | // SAFETY: Since StateID is always representable as a u32, all we need | |
3407 | // to do is ensure that we have the proper length and alignment. We've | |
3408 | // checked both above, so the cast below is safe. | |
3409 | // | |
781aab86 FG |
3410 | // N.B. This is the only not-safe code in this function. |
3411 | let table = core::slice::from_raw_parts( | |
3412 | table_bytes.as_ptr().cast::<u32>(), | |
3413 | trans_len, | |
3414 | ); | |
487cf647 | 3415 | let tt = TransitionTable { table, classes, stride2 }; |
781aab86 | 3416 | Ok((tt, slice.as_ptr().as_usize() - slice_start)) |
487cf647 FG |
3417 | } |
3418 | } | |
3419 | ||
781aab86 | 3420 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
3421 | impl TransitionTable<Vec<u32>> { |
3422 | /// Create a minimal transition table with just two states: a dead state | |
3423 | /// and a quit state. The alphabet length and stride of the transition | |
3424 | /// table is determined by the given set of equivalence classes. | |
3425 | fn minimal(classes: ByteClasses) -> TransitionTable<Vec<u32>> { | |
3426 | let mut tt = TransitionTable { | |
3427 | table: vec![], | |
3428 | classes, | |
3429 | stride2: classes.stride2(), | |
3430 | }; | |
3431 | // Two states, regardless of alphabet size, can always fit into u32. | |
3432 | tt.add_empty_state().unwrap(); // dead state | |
3433 | tt.add_empty_state().unwrap(); // quit state | |
3434 | tt | |
3435 | } | |
3436 | ||
3437 | /// Set a transition in this table. Both the `from` and `to` states must | |
3438 | /// already exist, otherwise this panics. `unit` should correspond to the | |
3439 | /// transition out of `from` to set to `to`. | |
3440 | fn set(&mut self, from: StateID, unit: alphabet::Unit, to: StateID) { | |
3441 | assert!(self.is_valid(from), "invalid 'from' state"); | |
3442 | assert!(self.is_valid(to), "invalid 'to' state"); | |
3443 | self.table[from.as_usize() + self.classes.get_by_unit(unit)] = | |
3444 | to.as_u32(); | |
3445 | } | |
3446 | ||
3447 | /// Add an empty state (a state where all transitions lead to a dead state) | |
3448 | /// and return its identifier. The identifier returned is guaranteed to | |
3449 | /// not point to any other existing state. | |
3450 | /// | |
3451 | /// If adding a state would exhaust the state identifier space, then this | |
3452 | /// returns an error. | |
781aab86 | 3453 | fn add_empty_state(&mut self) -> Result<StateID, BuildError> { |
487cf647 FG |
3454 | // Normally, to get a fresh state identifier, we would just |
3455 | // take the index of the next state added to the transition | |
3456 | // table. However, we actually perform an optimization here | |
3457 | // that premultiplies state IDs by the stride, such that they | |
3458 | // point immediately at the beginning of their transitions in | |
3459 | // the transition table. This avoids an extra multiplication | |
3460 | // instruction for state lookup at search time. | |
3461 | // | |
3462 | // Premultiplied identifiers means that instead of your matching | |
3463 | // loop looking something like this: | |
3464 | // | |
3465 | // state = dfa.start | |
3466 | // for byte in haystack: | |
3467 | // next = dfa.transitions[state * stride + byte] | |
3468 | // if dfa.is_match(next): | |
3469 | // return true | |
3470 | // return false | |
3471 | // | |
3472 | // it can instead look like this: | |
3473 | // | |
3474 | // state = dfa.start | |
3475 | // for byte in haystack: | |
3476 | // next = dfa.transitions[state + byte] | |
3477 | // if dfa.is_match(next): | |
3478 | // return true | |
3479 | // return false | |
3480 | // | |
3481 | // In other words, we save a multiplication instruction in the | |
3482 | // critical path. This turns out to be a decent performance win. | |
3483 | // The cost of using premultiplied state ids is that they can | |
3484 | // require a bigger state id representation. (And they also make | |
3485 | // the code a bit more complex, especially during minimization and | |
3486 | // when reshuffling states, as one needs to convert back and forth | |
3487 | // between state IDs and state indices.) | |
3488 | // | |
3489 | // To do this, we simply take the index of the state into the | |
3490 | // entire transition table, rather than the index of the state | |
3491 | // itself. e.g., If the stride is 64, then the ID of the 3rd state | |
3492 | // is 192, not 2. | |
3493 | let next = self.table.len(); | |
781aab86 FG |
3494 | let id = |
3495 | StateID::new(next).map_err(|_| BuildError::too_many_states())?; | |
487cf647 FG |
3496 | self.table.extend(iter::repeat(0).take(self.stride())); |
3497 | Ok(id) | |
3498 | } | |
3499 | ||
3500 | /// Swap the two states given in this transition table. | |
3501 | /// | |
3502 | /// This routine does not do anything to check the correctness of this | |
3503 | /// swap. Callers must ensure that other states pointing to id1 and id2 are | |
3504 | /// updated appropriately. | |
3505 | /// | |
3506 | /// Both id1 and id2 must point to valid states, otherwise this panics. | |
3507 | fn swap(&mut self, id1: StateID, id2: StateID) { | |
3508 | assert!(self.is_valid(id1), "invalid 'id1' state: {:?}", id1); | |
3509 | assert!(self.is_valid(id2), "invalid 'id2' state: {:?}", id2); | |
3510 | // We only need to swap the parts of the state that are used. So if the | |
3511 | // stride is 64, but the alphabet length is only 33, then we save a lot | |
3512 | // of work. | |
3513 | for b in 0..self.classes.alphabet_len() { | |
3514 | self.table.swap(id1.as_usize() + b, id2.as_usize() + b); | |
3515 | } | |
3516 | } | |
3517 | ||
781aab86 FG |
3518 | /// Remap the transitions for the state given according to the function |
3519 | /// given. This applies the given map function to every transition in the | |
3520 | /// given state and changes the transition in place to the result of the | |
3521 | /// map function for that transition. | |
3522 | fn remap(&mut self, id: StateID, map: impl Fn(StateID) -> StateID) { | |
3523 | for byte in 0..self.alphabet_len() { | |
3524 | let i = id.as_usize() + byte; | |
3525 | let next = self.table()[i]; | |
3526 | self.table_mut()[id.as_usize() + byte] = map(next); | |
3527 | } | |
3528 | } | |
3529 | ||
3530 | /// Truncate the states in this transition table to the given length. | |
487cf647 FG |
3531 | /// |
3532 | /// This routine does not do anything to check the correctness of this | |
3533 | /// truncation. Callers must ensure that other states pointing to truncated | |
3534 | /// states are updated appropriately. | |
781aab86 FG |
3535 | fn truncate(&mut self, len: usize) { |
3536 | self.table.truncate(len << self.stride2); | |
487cf647 FG |
3537 | } |
3538 | } | |
3539 | ||
3540 | impl<T: AsRef<[u32]>> TransitionTable<T> { | |
3541 | /// Writes a serialized form of this transition table to the buffer given. | |
3542 | /// If the buffer is too small, then an error is returned. To determine | |
3543 | /// how big the buffer must be, use `write_to_len`. | |
3544 | fn write_to<E: Endian>( | |
3545 | &self, | |
3546 | mut dst: &mut [u8], | |
3547 | ) -> Result<usize, SerializeError> { | |
3548 | let nwrite = self.write_to_len(); | |
3549 | if dst.len() < nwrite { | |
3550 | return Err(SerializeError::buffer_too_small("transition table")); | |
3551 | } | |
3552 | dst = &mut dst[..nwrite]; | |
3553 | ||
781aab86 | 3554 | // write state length |
487cf647 | 3555 | // Unwrap is OK since number of states is guaranteed to fit in a u32. |
781aab86 | 3556 | E::write_u32(u32::try_from(self.len()).unwrap(), dst); |
487cf647 FG |
3557 | dst = &mut dst[size_of::<u32>()..]; |
3558 | ||
3559 | // write state stride (as power of 2) | |
3560 | // Unwrap is OK since stride2 is guaranteed to be <= 9. | |
3561 | E::write_u32(u32::try_from(self.stride2).unwrap(), dst); | |
3562 | dst = &mut dst[size_of::<u32>()..]; | |
3563 | ||
3564 | // write byte class map | |
3565 | let n = self.classes.write_to(dst)?; | |
3566 | dst = &mut dst[n..]; | |
3567 | ||
3568 | // write actual transitions | |
3569 | for &sid in self.table() { | |
781aab86 | 3570 | let n = wire::write_state_id::<E>(sid, &mut dst); |
487cf647 FG |
3571 | dst = &mut dst[n..]; |
3572 | } | |
3573 | Ok(nwrite) | |
3574 | } | |
3575 | ||
3576 | /// Returns the number of bytes the serialized form of this transition | |
3577 | /// table will use. | |
3578 | fn write_to_len(&self) -> usize { | |
781aab86 | 3579 | size_of::<u32>() // state length |
487cf647 FG |
3580 | + size_of::<u32>() // stride2 |
3581 | + self.classes.write_to_len() | |
3582 | + (self.table().len() * StateID::SIZE) | |
3583 | } | |
3584 | ||
3585 | /// Validates that every state ID in this transition table is valid. | |
3586 | /// | |
3587 | /// That is, every state ID can be used to correctly index a state in this | |
3588 | /// table. | |
781aab86 | 3589 | fn validate(&self, sp: &Special) -> Result<(), DeserializeError> { |
487cf647 | 3590 | for state in self.states() { |
781aab86 FG |
3591 | // We check that the ID itself is well formed. That is, if it's |
3592 | // a special state then it must actually be a quit, dead, accel, | |
3593 | // match or start state. | |
3594 | if sp.is_special_state(state.id()) { | |
3595 | let is_actually_special = sp.is_dead_state(state.id()) | |
3596 | || sp.is_quit_state(state.id()) | |
3597 | || sp.is_match_state(state.id()) | |
3598 | || sp.is_start_state(state.id()) | |
3599 | || sp.is_accel_state(state.id()); | |
3600 | if !is_actually_special { | |
3601 | // This is kind of a cryptic error message... | |
3602 | return Err(DeserializeError::generic( | |
3603 | "found dense state tagged as special but \ | |
3604 | wasn't actually special", | |
3605 | )); | |
3606 | } | |
3607 | } | |
487cf647 FG |
3608 | for (_, to) in state.transitions() { |
3609 | if !self.is_valid(to) { | |
3610 | return Err(DeserializeError::generic( | |
3611 | "found invalid state ID in transition table", | |
3612 | )); | |
3613 | } | |
3614 | } | |
3615 | } | |
3616 | Ok(()) | |
3617 | } | |
3618 | ||
3619 | /// Converts this transition table to a borrowed value. | |
3620 | fn as_ref(&self) -> TransitionTable<&'_ [u32]> { | |
3621 | TransitionTable { | |
3622 | table: self.table.as_ref(), | |
3623 | classes: self.classes.clone(), | |
3624 | stride2: self.stride2, | |
3625 | } | |
3626 | } | |
3627 | ||
3628 | /// Converts this transition table to an owned value. | |
3629 | #[cfg(feature = "alloc")] | |
781aab86 | 3630 | fn to_owned(&self) -> TransitionTable<alloc::vec::Vec<u32>> { |
487cf647 FG |
3631 | TransitionTable { |
3632 | table: self.table.as_ref().to_vec(), | |
3633 | classes: self.classes.clone(), | |
3634 | stride2: self.stride2, | |
3635 | } | |
3636 | } | |
3637 | ||
3638 | /// Return the state for the given ID. If the given ID is not valid, then | |
3639 | /// this panics. | |
3640 | fn state(&self, id: StateID) -> State<'_> { | |
3641 | assert!(self.is_valid(id)); | |
3642 | ||
3643 | let i = id.as_usize(); | |
3644 | State { | |
3645 | id, | |
3646 | stride2: self.stride2, | |
3647 | transitions: &self.table()[i..i + self.alphabet_len()], | |
3648 | } | |
3649 | } | |
3650 | ||
3651 | /// Returns an iterator over all states in this transition table. | |
3652 | /// | |
3653 | /// This iterator yields a tuple for each state. The first element of the | |
3654 | /// tuple corresponds to a state's identifier, and the second element | |
3655 | /// corresponds to the state itself (comprised of its transitions). | |
3656 | fn states(&self) -> StateIter<'_, T> { | |
3657 | StateIter { | |
3658 | tt: self, | |
3659 | it: self.table().chunks(self.stride()).enumerate(), | |
3660 | } | |
3661 | } | |
3662 | ||
3663 | /// Convert a state identifier to an index to a state (in the range | |
781aab86 | 3664 | /// 0..self.len()). |
487cf647 FG |
3665 | /// |
3666 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state | |
3667 | /// to some other information (such as a remapped state ID). | |
3668 | /// | |
3669 | /// If the given ID is not valid, then this may panic or produce an | |
3670 | /// incorrect index. | |
3671 | fn to_index(&self, id: StateID) -> usize { | |
3672 | id.as_usize() >> self.stride2 | |
3673 | } | |
3674 | ||
781aab86 | 3675 | /// Convert an index to a state (in the range 0..self.len()) to an actual |
487cf647 FG |
3676 | /// state identifier. |
3677 | /// | |
3678 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state | |
3679 | /// to some other information (such as a remapped state ID). | |
3680 | /// | |
3681 | /// If the given index is not in the specified range, then this may panic | |
3682 | /// or produce an incorrect state ID. | |
781aab86 | 3683 | fn to_state_id(&self, index: usize) -> StateID { |
487cf647 FG |
3684 | // CORRECTNESS: If the given index is not valid, then it is not |
3685 | // required for this to panic or return a valid state ID. | |
3686 | StateID::new_unchecked(index << self.stride2) | |
3687 | } | |
3688 | ||
3689 | /// Returns the state ID for the state immediately following the one given. | |
3690 | /// | |
3691 | /// This does not check whether the state ID returned is invalid. In fact, | |
3692 | /// if the state ID given is the last state in this DFA, then the state ID | |
3693 | /// returned is guaranteed to be invalid. | |
781aab86 | 3694 | #[cfg(feature = "dfa-build")] |
487cf647 | 3695 | fn next_state_id(&self, id: StateID) -> StateID { |
781aab86 | 3696 | self.to_state_id(self.to_index(id).checked_add(1).unwrap()) |
487cf647 FG |
3697 | } |
3698 | ||
3699 | /// Returns the state ID for the state immediately preceding the one given. | |
3700 | /// | |
3701 | /// If the dead ID given (which is zero), then this panics. | |
781aab86 | 3702 | #[cfg(feature = "dfa-build")] |
487cf647 | 3703 | fn prev_state_id(&self, id: StateID) -> StateID { |
781aab86 | 3704 | self.to_state_id(self.to_index(id).checked_sub(1).unwrap()) |
487cf647 FG |
3705 | } |
3706 | ||
3707 | /// Returns the table as a slice of state IDs. | |
3708 | fn table(&self) -> &[StateID] { | |
781aab86 | 3709 | wire::u32s_to_state_ids(self.table.as_ref()) |
487cf647 FG |
3710 | } |
3711 | ||
3712 | /// Returns the total number of states in this transition table. | |
3713 | /// | |
3714 | /// Note that a DFA always has at least two states: the dead and quit | |
3715 | /// states. In particular, the dead state always has ID 0 and is | |
3716 | /// correspondingly always the first state. The dead state is never a match | |
3717 | /// state. | |
781aab86 | 3718 | fn len(&self) -> usize { |
487cf647 FG |
3719 | self.table().len() >> self.stride2 |
3720 | } | |
3721 | ||
3722 | /// Returns the total stride for every state in this DFA. This corresponds | |
3723 | /// to the total number of transitions used by each state in this DFA's | |
3724 | /// transition table. | |
3725 | fn stride(&self) -> usize { | |
3726 | 1 << self.stride2 | |
3727 | } | |
3728 | ||
3729 | /// Returns the total number of elements in the alphabet for this | |
3730 | /// transition table. This is always less than or equal to `self.stride()`. | |
3731 | /// It is only equal when the alphabet length is a power of 2. Otherwise, | |
3732 | /// it is always strictly less. | |
3733 | fn alphabet_len(&self) -> usize { | |
3734 | self.classes.alphabet_len() | |
3735 | } | |
3736 | ||
3737 | /// Returns true if and only if the given state ID is valid for this | |
3738 | /// transition table. Validity in this context means that the given ID can | |
3739 | /// be used as a valid offset with `self.stride()` to index this transition | |
3740 | /// table. | |
3741 | fn is_valid(&self, id: StateID) -> bool { | |
3742 | let id = id.as_usize(); | |
3743 | id < self.table().len() && id % self.stride() == 0 | |
3744 | } | |
3745 | ||
3746 | /// Return the memory usage, in bytes, of this transition table. | |
3747 | /// | |
3748 | /// This does not include the size of a `TransitionTable` value itself. | |
3749 | fn memory_usage(&self) -> usize { | |
3750 | self.table().len() * StateID::SIZE | |
3751 | } | |
3752 | } | |
3753 | ||
781aab86 | 3754 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
3755 | impl<T: AsMut<[u32]>> TransitionTable<T> { |
3756 | /// Returns the table as a slice of state IDs. | |
3757 | fn table_mut(&mut self) -> &mut [StateID] { | |
781aab86 | 3758 | wire::u32s_to_state_ids_mut(self.table.as_mut()) |
487cf647 FG |
3759 | } |
3760 | } | |
3761 | ||
3762 | /// The set of all possible starting states in a DFA. | |
3763 | /// | |
3764 | /// The set of starting states corresponds to the possible choices one can make | |
3765 | /// in terms of starting a DFA. That is, before following the first transition, | |
3766 | /// you first need to select the state that you start in. | |
3767 | /// | |
3768 | /// Normally, a DFA converted from an NFA that has a single starting state | |
3769 | /// would itself just have one starting state. However, our support for look | |
3770 | /// around generally requires more starting states. The correct starting state | |
3771 | /// is chosen based on certain properties of the position at which we begin | |
3772 | /// our search. | |
3773 | /// | |
3774 | /// Before listing those properties, we first must define two terms: | |
3775 | /// | |
3776 | /// * `haystack` - The bytes to execute the search. The search always starts | |
3777 | /// at the beginning of `haystack` and ends before or at the end of | |
3778 | /// `haystack`. | |
3779 | /// * `context` - The (possibly empty) bytes surrounding `haystack`. `haystack` | |
3780 | /// must be contained within `context` such that `context` is at least as big | |
3781 | /// as `haystack`. | |
3782 | /// | |
3783 | /// This split is crucial for dealing with look-around. For example, consider | |
3784 | /// the context `foobarbaz`, the haystack `bar` and the regex `^bar$`. This | |
3785 | /// regex should _not_ match the haystack since `bar` does not appear at the | |
3786 | /// beginning of the input. Similarly, the regex `\Bbar\B` should match the | |
3787 | /// haystack because `bar` is not surrounded by word boundaries. But a search | |
3788 | /// that does not take context into account would not permit `\B` to match | |
3789 | /// since the beginning of any string matches a word boundary. Similarly, a | |
3790 | /// search that does not take context into account when searching `^bar$` in | |
3791 | /// the haystack `bar` would produce a match when it shouldn't. | |
3792 | /// | |
3793 | /// Thus, it follows that the starting state is chosen based on the following | |
3794 | /// criteria, derived from the position at which the search starts in the | |
3795 | /// `context` (corresponding to the start of `haystack`): | |
3796 | /// | |
3797 | /// 1. If the search starts at the beginning of `context`, then the `Text` | |
3798 | /// start state is used. (Since `^` corresponds to | |
781aab86 | 3799 | /// `hir::Anchor::Start`.) |
487cf647 FG |
3800 | /// 2. If the search starts at a position immediately following a line |
3801 | /// terminator, then the `Line` start state is used. (Since `(?m:^)` | |
781aab86 | 3802 | /// corresponds to `hir::Anchor::StartLF`.) |
487cf647 FG |
3803 | /// 3. If the search starts at a position immediately following a byte |
3804 | /// classified as a "word" character (`[_0-9a-zA-Z]`), then the `WordByte` | |
3805 | /// start state is used. (Since `(?-u:\b)` corresponds to a word boundary.) | |
3806 | /// 4. Otherwise, if the search starts at a position immediately following | |
3807 | /// a byte that is not classified as a "word" character (`[^_0-9a-zA-Z]`), | |
3808 | /// then the `NonWordByte` start state is used. (Since `(?-u:\B)` | |
3809 | /// corresponds to a not-word-boundary.) | |
3810 | /// | |
3811 | /// (N.B. Unicode word boundaries are not supported by the DFA because they | |
3812 | /// require multi-byte look-around and this is difficult to support in a DFA.) | |
3813 | /// | |
3814 | /// To further complicate things, we also support constructing individual | |
3815 | /// anchored start states for each pattern in the DFA. (Which is required to | |
3816 | /// implement overlapping regexes correctly, but is also generally useful.) | |
3817 | /// Thus, when individual start states for each pattern are enabled, then the | |
3818 | /// total number of start states represented is `4 + (4 * #patterns)`, where | |
3819 | /// the 4 comes from each of the 4 possibilities above. The first 4 represents | |
3820 | /// the starting states for the entire DFA, which support searching for | |
3821 | /// multiple patterns simultaneously (possibly unanchored). | |
3822 | /// | |
3823 | /// If individual start states are disabled, then this will only store 4 | |
3824 | /// start states. Typically, individual start states are only enabled when | |
3825 | /// constructing the reverse DFA for regex matching. But they are also useful | |
3826 | /// for building DFAs that can search for a specific pattern or even to support | |
3827 | /// both anchored and unanchored searches with the same DFA. | |
3828 | /// | |
3829 | /// Note though that while the start table always has either `4` or | |
3830 | /// `4 + (4 * #patterns)` starting state *ids*, the total number of states | |
3831 | /// might be considerably smaller. That is, many of the IDs may be duplicative. | |
3832 | /// (For example, if a regex doesn't have a `\b` sub-pattern, then there's no | |
3833 | /// reason to generate a unique starting state for handling word boundaries. | |
3834 | /// Similarly for start/end anchors.) | |
3835 | #[derive(Clone)] | |
3836 | pub(crate) struct StartTable<T> { | |
3837 | /// The initial start state IDs. | |
3838 | /// | |
3839 | /// In practice, T is either `Vec<u32>` or `&[u32]`. | |
3840 | /// | |
781aab86 FG |
3841 | /// The first `2 * stride` (currently always 8) entries always correspond |
3842 | /// to the starts states for the entire DFA, with the first 4 entries being | |
3843 | /// for unanchored searches and the second 4 entries being for anchored | |
3844 | /// searches. To keep things simple, we always use 8 entries even if the | |
3845 | /// `StartKind` is not both. | |
3846 | /// | |
3847 | /// After that, there are `stride * patterns` state IDs, where `patterns` | |
3848 | /// may be zero in the case of a DFA with no patterns or in the case where | |
3849 | /// the DFA was built without enabling starting states for each pattern. | |
487cf647 | 3850 | table: T, |
781aab86 FG |
3851 | /// The starting state configuration supported. When 'both', both |
3852 | /// unanchored and anchored searches work. When 'unanchored', anchored | |
3853 | /// searches panic. When 'anchored', unanchored searches panic. | |
3854 | kind: StartKind, | |
3855 | /// The start state configuration for every possible byte. | |
3856 | start_map: StartByteMap, | |
487cf647 FG |
3857 | /// The number of starting state IDs per pattern. |
3858 | stride: usize, | |
3859 | /// The total number of patterns for which starting states are encoded. | |
781aab86 FG |
3860 | /// This is `None` for DFAs that were built without start states for each |
3861 | /// pattern. Thus, one cannot use this field to say how many patterns | |
3862 | /// are in the DFA in all cases. It is specific to how many patterns are | |
3863 | /// represented in this start table. | |
3864 | pattern_len: Option<usize>, | |
3865 | /// The universal starting state for unanchored searches. This is only | |
3866 | /// present when the DFA supports unanchored searches and when all starting | |
3867 | /// state IDs for an unanchored search are equivalent. | |
3868 | universal_start_unanchored: Option<StateID>, | |
3869 | /// The universal starting state for anchored searches. This is only | |
3870 | /// present when the DFA supports anchored searches and when all starting | |
3871 | /// state IDs for an anchored search are equivalent. | |
3872 | universal_start_anchored: Option<StateID>, | |
487cf647 FG |
3873 | } |
3874 | ||
781aab86 | 3875 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
3876 | impl StartTable<Vec<u32>> { |
3877 | /// Create a valid set of start states all pointing to the dead state. | |
3878 | /// | |
3879 | /// When the corresponding DFA is constructed with start states for each | |
3880 | /// pattern, then `patterns` should be the number of patterns. Otherwise, | |
3881 | /// it should be zero. | |
3882 | /// | |
3883 | /// If the total table size could exceed the allocatable limit, then this | |
3884 | /// returns an error. In practice, this is unlikely to be able to occur, | |
3885 | /// since it's likely that allocation would have failed long before it got | |
3886 | /// to this point. | |
781aab86 FG |
3887 | fn dead( |
3888 | kind: StartKind, | |
3889 | lookm: &LookMatcher, | |
3890 | pattern_len: Option<usize>, | |
3891 | ) -> Result<StartTable<Vec<u32>>, BuildError> { | |
3892 | if let Some(len) = pattern_len { | |
3893 | assert!(len <= PatternID::LIMIT); | |
3894 | } | |
3895 | let stride = Start::len(); | |
3896 | // OK because 2*4 is never going to overflow anything. | |
3897 | let starts_len = stride.checked_mul(2).unwrap(); | |
3898 | let pattern_starts_len = | |
3899 | match stride.checked_mul(pattern_len.unwrap_or(0)) { | |
3900 | Some(x) => x, | |
3901 | None => return Err(BuildError::too_many_start_states()), | |
3902 | }; | |
3903 | let table_len = match starts_len.checked_add(pattern_starts_len) { | |
487cf647 | 3904 | Some(x) => x, |
781aab86 | 3905 | None => return Err(BuildError::too_many_start_states()), |
487cf647 | 3906 | }; |
781aab86 FG |
3907 | if let Err(_) = isize::try_from(table_len) { |
3908 | return Err(BuildError::too_many_start_states()); | |
487cf647 FG |
3909 | } |
3910 | let table = vec![DEAD.as_u32(); table_len]; | |
781aab86 FG |
3911 | let start_map = StartByteMap::new(lookm); |
3912 | Ok(StartTable { | |
3913 | table, | |
3914 | kind, | |
3915 | start_map, | |
3916 | stride, | |
3917 | pattern_len, | |
3918 | universal_start_unanchored: None, | |
3919 | universal_start_anchored: None, | |
3920 | }) | |
487cf647 FG |
3921 | } |
3922 | } | |
3923 | ||
3924 | impl<'a> StartTable<&'a [u32]> { | |
3925 | /// Deserialize a table of start state IDs starting at the beginning of | |
3926 | /// `slice`. Upon success, return the total number of bytes read along with | |
3927 | /// the table of starting state IDs. | |
3928 | /// | |
3929 | /// If there was a problem deserializing any part of the starting IDs, | |
3930 | /// then this returns an error. Notably, if the given slice does not have | |
3931 | /// the same alignment as `StateID`, then this will return an error (among | |
3932 | /// other possible errors). | |
3933 | /// | |
3934 | /// This is guaranteed to execute in constant time. | |
3935 | /// | |
3936 | /// # Safety | |
3937 | /// | |
781aab86 | 3938 | /// This routine is not safe because it does not check the validity of the |
487cf647 FG |
3939 | /// starting state IDs themselves. In particular, the number of starting |
3940 | /// IDs can be of variable length, so it's possible that checking their | |
3941 | /// validity cannot be done in constant time. An invalid starting state | |
3942 | /// ID is not safe because other code may rely on the starting IDs being | |
3943 | /// correct (such as explicit bounds check elision). Therefore, an invalid | |
3944 | /// start ID can lead to undefined behavior. | |
3945 | /// | |
3946 | /// Callers that use this function must either pass on the safety invariant | |
3947 | /// or guarantee that the bytes given contain valid starting state IDs. | |
3948 | /// This guarantee is upheld by the bytes written by `write_to`. | |
3949 | unsafe fn from_bytes_unchecked( | |
3950 | mut slice: &'a [u8], | |
3951 | ) -> Result<(StartTable<&'a [u32]>, usize), DeserializeError> { | |
781aab86 | 3952 | let slice_start = slice.as_ptr().as_usize(); |
487cf647 | 3953 | |
781aab86 | 3954 | let (kind, nr) = StartKind::from_bytes(slice)?; |
487cf647 FG |
3955 | slice = &slice[nr..]; |
3956 | ||
781aab86 | 3957 | let (start_map, nr) = StartByteMap::from_bytes(slice)?; |
487cf647 FG |
3958 | slice = &slice[nr..]; |
3959 | ||
781aab86 FG |
3960 | let (stride, nr) = |
3961 | wire::try_read_u32_as_usize(slice, "start table stride")?; | |
3962 | slice = &slice[nr..]; | |
3963 | if stride != Start::len() { | |
487cf647 FG |
3964 | return Err(DeserializeError::generic( |
3965 | "invalid starting table stride", | |
3966 | )); | |
3967 | } | |
781aab86 FG |
3968 | |
3969 | let (maybe_pattern_len, nr) = | |
3970 | wire::try_read_u32_as_usize(slice, "start table patterns")?; | |
3971 | slice = &slice[nr..]; | |
3972 | let pattern_len = if maybe_pattern_len.as_u32() == u32::MAX { | |
3973 | None | |
3974 | } else { | |
3975 | Some(maybe_pattern_len) | |
3976 | }; | |
3977 | if pattern_len.map_or(false, |len| len > PatternID::LIMIT) { | |
487cf647 FG |
3978 | return Err(DeserializeError::generic( |
3979 | "invalid number of patterns", | |
3980 | )); | |
3981 | } | |
781aab86 FG |
3982 | |
3983 | let (universal_unanchored, nr) = | |
3984 | wire::try_read_u32(slice, "universal unanchored start")?; | |
3985 | slice = &slice[nr..]; | |
3986 | let universal_start_unanchored = if universal_unanchored == u32::MAX { | |
3987 | None | |
3988 | } else { | |
3989 | Some(StateID::try_from(universal_unanchored).map_err(|e| { | |
3990 | DeserializeError::state_id_error( | |
3991 | e, | |
3992 | "universal unanchored start", | |
3993 | ) | |
3994 | })?) | |
3995 | }; | |
3996 | ||
3997 | let (universal_anchored, nr) = | |
3998 | wire::try_read_u32(slice, "universal anchored start")?; | |
3999 | slice = &slice[nr..]; | |
4000 | let universal_start_anchored = if universal_anchored == u32::MAX { | |
4001 | None | |
4002 | } else { | |
4003 | Some(StateID::try_from(universal_anchored).map_err(|e| { | |
4004 | DeserializeError::state_id_error(e, "universal anchored start") | |
4005 | })?) | |
4006 | }; | |
4007 | ||
4008 | let pattern_table_size = wire::mul( | |
487cf647 | 4009 | stride, |
781aab86 FG |
4010 | pattern_len.unwrap_or(0), |
4011 | "invalid pattern length", | |
4012 | )?; | |
4013 | // Our start states always start with a two stride of start states for | |
4014 | // the entire automaton. The first stride is for unanchored starting | |
4015 | // states and the second stride is for anchored starting states. What | |
4016 | // follows it are an optional set of start states for each pattern. | |
4017 | let start_state_len = wire::add( | |
4018 | wire::mul(2, stride, "start state stride too big")?, | |
487cf647 FG |
4019 | pattern_table_size, |
4020 | "invalid 'any' pattern starts size", | |
4021 | )?; | |
781aab86 FG |
4022 | let table_bytes_len = wire::mul( |
4023 | start_state_len, | |
487cf647 FG |
4024 | StateID::SIZE, |
4025 | "pattern table bytes length", | |
4026 | )?; | |
781aab86 FG |
4027 | wire::check_slice_len(slice, table_bytes_len, "start ID table")?; |
4028 | wire::check_alignment::<StateID>(slice)?; | |
487cf647 FG |
4029 | let table_bytes = &slice[..table_bytes_len]; |
4030 | slice = &slice[table_bytes_len..]; | |
4031 | // SAFETY: Since StateID is always representable as a u32, all we need | |
4032 | // to do is ensure that we have the proper length and alignment. We've | |
4033 | // checked both above, so the cast below is safe. | |
4034 | // | |
781aab86 FG |
4035 | // N.B. This is the only not-safe code in this function. |
4036 | let table = core::slice::from_raw_parts( | |
4037 | table_bytes.as_ptr().cast::<u32>(), | |
4038 | start_state_len, | |
4039 | ); | |
4040 | let st = StartTable { | |
4041 | table, | |
4042 | kind, | |
4043 | start_map, | |
4044 | stride, | |
4045 | pattern_len, | |
4046 | universal_start_unanchored, | |
4047 | universal_start_anchored, | |
487cf647 | 4048 | }; |
781aab86 | 4049 | Ok((st, slice.as_ptr().as_usize() - slice_start)) |
487cf647 FG |
4050 | } |
4051 | } | |
4052 | ||
4053 | impl<T: AsRef<[u32]>> StartTable<T> { | |
4054 | /// Writes a serialized form of this start table to the buffer given. If | |
4055 | /// the buffer is too small, then an error is returned. To determine how | |
4056 | /// big the buffer must be, use `write_to_len`. | |
4057 | fn write_to<E: Endian>( | |
4058 | &self, | |
4059 | mut dst: &mut [u8], | |
4060 | ) -> Result<usize, SerializeError> { | |
4061 | let nwrite = self.write_to_len(); | |
4062 | if dst.len() < nwrite { | |
4063 | return Err(SerializeError::buffer_too_small( | |
4064 | "starting table ids", | |
4065 | )); | |
4066 | } | |
4067 | dst = &mut dst[..nwrite]; | |
4068 | ||
781aab86 FG |
4069 | // write start kind |
4070 | let nw = self.kind.write_to::<E>(dst)?; | |
4071 | dst = &mut dst[nw..]; | |
4072 | // write start byte map | |
4073 | let nw = self.start_map.write_to(dst)?; | |
4074 | dst = &mut dst[nw..]; | |
487cf647 FG |
4075 | // write stride |
4076 | // Unwrap is OK since the stride is always 4 (currently). | |
4077 | E::write_u32(u32::try_from(self.stride).unwrap(), dst); | |
4078 | dst = &mut dst[size_of::<u32>()..]; | |
781aab86 | 4079 | // write pattern length |
487cf647 | 4080 | // Unwrap is OK since number of patterns is guaranteed to fit in a u32. |
781aab86 FG |
4081 | E::write_u32( |
4082 | u32::try_from(self.pattern_len.unwrap_or(0xFFFF_FFFF)).unwrap(), | |
4083 | dst, | |
4084 | ); | |
4085 | dst = &mut dst[size_of::<u32>()..]; | |
4086 | // write universal start unanchored state id, u32::MAX if absent | |
4087 | E::write_u32( | |
4088 | self.universal_start_unanchored | |
4089 | .map_or(u32::MAX, |sid| sid.as_u32()), | |
4090 | dst, | |
4091 | ); | |
4092 | dst = &mut dst[size_of::<u32>()..]; | |
4093 | // write universal start anchored state id, u32::MAX if absent | |
4094 | E::write_u32( | |
4095 | self.universal_start_anchored.map_or(u32::MAX, |sid| sid.as_u32()), | |
4096 | dst, | |
4097 | ); | |
487cf647 FG |
4098 | dst = &mut dst[size_of::<u32>()..]; |
4099 | // write start IDs | |
4100 | for &sid in self.table() { | |
781aab86 | 4101 | let n = wire::write_state_id::<E>(sid, &mut dst); |
487cf647 FG |
4102 | dst = &mut dst[n..]; |
4103 | } | |
4104 | Ok(nwrite) | |
4105 | } | |
4106 | ||
4107 | /// Returns the number of bytes the serialized form of this start ID table | |
4108 | /// will use. | |
4109 | fn write_to_len(&self) -> usize { | |
781aab86 FG |
4110 | self.kind.write_to_len() |
4111 | + self.start_map.write_to_len() | |
4112 | + size_of::<u32>() // stride | |
487cf647 | 4113 | + size_of::<u32>() // # patterns |
781aab86 FG |
4114 | + size_of::<u32>() // universal unanchored start |
4115 | + size_of::<u32>() // universal anchored start | |
487cf647 FG |
4116 | + (self.table().len() * StateID::SIZE) |
4117 | } | |
4118 | ||
4119 | /// Validates that every state ID in this start table is valid by checking | |
4120 | /// it against the given transition table (which must be for the same DFA). | |
4121 | /// | |
4122 | /// That is, every state ID can be used to correctly index a state. | |
4123 | fn validate( | |
4124 | &self, | |
4125 | tt: &TransitionTable<T>, | |
4126 | ) -> Result<(), DeserializeError> { | |
781aab86 FG |
4127 | if !self.universal_start_unanchored.map_or(true, |s| tt.is_valid(s)) { |
4128 | return Err(DeserializeError::generic( | |
4129 | "found invalid universal unanchored starting state ID", | |
4130 | )); | |
4131 | } | |
4132 | if !self.universal_start_anchored.map_or(true, |s| tt.is_valid(s)) { | |
4133 | return Err(DeserializeError::generic( | |
4134 | "found invalid universal anchored starting state ID", | |
4135 | )); | |
4136 | } | |
487cf647 FG |
4137 | for &id in self.table() { |
4138 | if !tt.is_valid(id) { | |
4139 | return Err(DeserializeError::generic( | |
4140 | "found invalid starting state ID", | |
4141 | )); | |
4142 | } | |
4143 | } | |
4144 | Ok(()) | |
4145 | } | |
4146 | ||
4147 | /// Converts this start list to a borrowed value. | |
4148 | fn as_ref(&self) -> StartTable<&'_ [u32]> { | |
4149 | StartTable { | |
4150 | table: self.table.as_ref(), | |
781aab86 FG |
4151 | kind: self.kind, |
4152 | start_map: self.start_map.clone(), | |
487cf647 | 4153 | stride: self.stride, |
781aab86 FG |
4154 | pattern_len: self.pattern_len, |
4155 | universal_start_unanchored: self.universal_start_unanchored, | |
4156 | universal_start_anchored: self.universal_start_anchored, | |
487cf647 FG |
4157 | } |
4158 | } | |
4159 | ||
4160 | /// Converts this start list to an owned value. | |
4161 | #[cfg(feature = "alloc")] | |
781aab86 | 4162 | fn to_owned(&self) -> StartTable<alloc::vec::Vec<u32>> { |
487cf647 FG |
4163 | StartTable { |
4164 | table: self.table.as_ref().to_vec(), | |
781aab86 FG |
4165 | kind: self.kind, |
4166 | start_map: self.start_map.clone(), | |
487cf647 | 4167 | stride: self.stride, |
781aab86 FG |
4168 | pattern_len: self.pattern_len, |
4169 | universal_start_unanchored: self.universal_start_unanchored, | |
4170 | universal_start_anchored: self.universal_start_anchored, | |
487cf647 FG |
4171 | } |
4172 | } | |
4173 | ||
781aab86 FG |
4174 | /// Return the start state for the given input and starting configuration. |
4175 | /// This returns an error if the input configuration is not supported by | |
4176 | /// this DFA. For example, requesting an unanchored search when the DFA was | |
4177 | /// not built with unanchored starting states. Or asking for an anchored | |
4178 | /// pattern search with an invalid pattern ID or on a DFA that was not | |
4179 | /// built with start states for each pattern. | |
4180 | #[cfg_attr(feature = "perf-inline", inline(always))] | |
4181 | fn start( | |
4182 | &self, | |
4183 | input: &Input<'_>, | |
4184 | start: Start, | |
4185 | ) -> Result<StateID, MatchError> { | |
4186 | let start_index = start.as_usize(); | |
4187 | let mode = input.get_anchored(); | |
4188 | let index = match mode { | |
4189 | Anchored::No => { | |
4190 | if !self.kind.has_unanchored() { | |
4191 | return Err(MatchError::unsupported_anchored(mode)); | |
4192 | } | |
4193 | start_index | |
4194 | } | |
4195 | Anchored::Yes => { | |
4196 | if !self.kind.has_anchored() { | |
4197 | return Err(MatchError::unsupported_anchored(mode)); | |
4198 | } | |
4199 | self.stride + start_index | |
4200 | } | |
4201 | Anchored::Pattern(pid) => { | |
4202 | let len = match self.pattern_len { | |
4203 | None => { | |
4204 | return Err(MatchError::unsupported_anchored(mode)) | |
4205 | } | |
4206 | Some(len) => len, | |
4207 | }; | |
4208 | if pid.as_usize() >= len { | |
4209 | return Ok(DEAD); | |
4210 | } | |
4211 | (2 * self.stride) | |
4212 | + (self.stride * pid.as_usize()) | |
4213 | + start_index | |
487cf647 FG |
4214 | } |
4215 | }; | |
781aab86 | 4216 | Ok(self.table()[index]) |
487cf647 FG |
4217 | } |
4218 | ||
4219 | /// Returns an iterator over all start state IDs in this table. | |
4220 | /// | |
4221 | /// Each item is a triple of: start state ID, the start state type and the | |
4222 | /// pattern ID (if any). | |
4223 | fn iter(&self) -> StartStateIter<'_> { | |
4224 | StartStateIter { st: self.as_ref(), i: 0 } | |
4225 | } | |
4226 | ||
4227 | /// Returns the table as a slice of state IDs. | |
4228 | fn table(&self) -> &[StateID] { | |
781aab86 | 4229 | wire::u32s_to_state_ids(self.table.as_ref()) |
487cf647 FG |
4230 | } |
4231 | ||
4232 | /// Return the memory usage, in bytes, of this start list. | |
4233 | /// | |
4234 | /// This does not include the size of a `StartList` value itself. | |
4235 | fn memory_usage(&self) -> usize { | |
4236 | self.table().len() * StateID::SIZE | |
4237 | } | |
4238 | } | |
4239 | ||
781aab86 | 4240 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
4241 | impl<T: AsMut<[u32]>> StartTable<T> { |
4242 | /// Set the start state for the given index and pattern. | |
4243 | /// | |
4244 | /// If the pattern ID or state ID are not valid, then this will panic. | |
781aab86 FG |
4245 | fn set_start(&mut self, anchored: Anchored, start: Start, id: StateID) { |
4246 | let start_index = start.as_usize(); | |
4247 | let index = match anchored { | |
4248 | Anchored::No => start_index, | |
4249 | Anchored::Yes => self.stride + start_index, | |
4250 | Anchored::Pattern(pid) => { | |
4251 | let pid = pid.as_usize(); | |
4252 | let len = self | |
4253 | .pattern_len | |
4254 | .expect("start states for each pattern enabled"); | |
4255 | assert!(pid < len, "invalid pattern ID {:?}", pid); | |
4256 | self.stride | |
4257 | .checked_mul(pid) | |
4258 | .unwrap() | |
4259 | .checked_add(self.stride.checked_mul(2).unwrap()) | |
4260 | .unwrap() | |
4261 | .checked_add(start_index) | |
4262 | .unwrap() | |
4263 | } | |
487cf647 FG |
4264 | }; |
4265 | self.table_mut()[index] = id; | |
4266 | } | |
4267 | ||
4268 | /// Returns the table as a mutable slice of state IDs. | |
4269 | fn table_mut(&mut self) -> &mut [StateID] { | |
781aab86 | 4270 | wire::u32s_to_state_ids_mut(self.table.as_mut()) |
487cf647 FG |
4271 | } |
4272 | } | |
4273 | ||
4274 | /// An iterator over start state IDs. | |
4275 | /// | |
781aab86 FG |
4276 | /// This iterator yields a triple of start state ID, the anchored mode and the |
4277 | /// start state type. If a pattern ID is relevant, then the anchored mode will | |
4278 | /// contain it. Start states with an anchored mode containing a pattern ID will | |
4279 | /// only occur when the DFA was compiled with start states for each pattern | |
4280 | /// (which is disabled by default). | |
487cf647 FG |
4281 | pub(crate) struct StartStateIter<'a> { |
4282 | st: StartTable<&'a [u32]>, | |
4283 | i: usize, | |
4284 | } | |
4285 | ||
4286 | impl<'a> Iterator for StartStateIter<'a> { | |
781aab86 | 4287 | type Item = (StateID, Anchored, Start); |
487cf647 | 4288 | |
781aab86 | 4289 | fn next(&mut self) -> Option<(StateID, Anchored, Start)> { |
487cf647 FG |
4290 | let i = self.i; |
4291 | let table = self.st.table(); | |
4292 | if i >= table.len() { | |
4293 | return None; | |
4294 | } | |
4295 | self.i += 1; | |
4296 | ||
4297 | // This unwrap is okay since the stride of the starting state table | |
4298 | // must always match the number of start state types. | |
4299 | let start_type = Start::from_usize(i % self.st.stride).unwrap(); | |
781aab86 FG |
4300 | let anchored = if i < self.st.stride { |
4301 | Anchored::No | |
4302 | } else if i < (2 * self.st.stride) { | |
4303 | Anchored::Yes | |
487cf647 | 4304 | } else { |
781aab86 FG |
4305 | let pid = (i - (2 * self.st.stride)) / self.st.stride; |
4306 | Anchored::Pattern(PatternID::new(pid).unwrap()) | |
487cf647 | 4307 | }; |
781aab86 | 4308 | Some((table[i], anchored, start_type)) |
487cf647 FG |
4309 | } |
4310 | } | |
4311 | ||
4312 | /// This type represents that patterns that should be reported whenever a DFA | |
4313 | /// enters a match state. This structure exists to support DFAs that search for | |
4314 | /// matches for multiple regexes. | |
4315 | /// | |
4316 | /// This structure relies on the fact that all match states in a DFA occur | |
4317 | /// contiguously in the DFA's transition table. (See dfa/special.rs for a more | |
4318 | /// detailed breakdown of the representation.) Namely, when a match occurs, we | |
4319 | /// know its state ID. Since we know the start and end of the contiguous region | |
4320 | /// of match states, we can use that to compute the position at which the match | |
4321 | /// state occurs. That in turn is used as an offset into this structure. | |
4322 | #[derive(Clone, Debug)] | |
4323 | struct MatchStates<T> { | |
4324 | /// Slices is a flattened sequence of pairs, where each pair points to a | |
4325 | /// sub-slice of pattern_ids. The first element of the pair is an offset | |
4326 | /// into pattern_ids and the second element of the pair is the number | |
4327 | /// of 32-bit pattern IDs starting at that position. That is, each pair | |
4328 | /// corresponds to a single DFA match state and its corresponding match | |
4329 | /// IDs. The number of pairs always corresponds to the number of distinct | |
4330 | /// DFA match states. | |
4331 | /// | |
4332 | /// In practice, T is either Vec<u32> or &[u32]. | |
4333 | slices: T, | |
4334 | /// A flattened sequence of pattern IDs for each DFA match state. The only | |
4335 | /// way to correctly read this sequence is indirectly via `slices`. | |
4336 | /// | |
4337 | /// In practice, T is either Vec<u32> or &[u32]. | |
4338 | pattern_ids: T, | |
4339 | /// The total number of unique patterns represented by these match states. | |
781aab86 | 4340 | pattern_len: usize, |
487cf647 FG |
4341 | } |
4342 | ||
4343 | impl<'a> MatchStates<&'a [u32]> { | |
4344 | unsafe fn from_bytes_unchecked( | |
4345 | mut slice: &'a [u8], | |
4346 | ) -> Result<(MatchStates<&'a [u32]>, usize), DeserializeError> { | |
781aab86 | 4347 | let slice_start = slice.as_ptr().as_usize(); |
487cf647 FG |
4348 | |
4349 | // Read the total number of match states. | |
781aab86 FG |
4350 | let (state_len, nr) = |
4351 | wire::try_read_u32_as_usize(slice, "match state length")?; | |
487cf647 FG |
4352 | slice = &slice[nr..]; |
4353 | ||
4354 | // Read the slice start/length pairs. | |
781aab86 FG |
4355 | let pair_len = wire::mul(2, state_len, "match state offset pairs")?; |
4356 | let slices_bytes_len = wire::mul( | |
4357 | pair_len, | |
487cf647 FG |
4358 | PatternID::SIZE, |
4359 | "match state slice offset byte length", | |
4360 | )?; | |
781aab86 FG |
4361 | wire::check_slice_len(slice, slices_bytes_len, "match state slices")?; |
4362 | wire::check_alignment::<PatternID>(slice)?; | |
487cf647 FG |
4363 | let slices_bytes = &slice[..slices_bytes_len]; |
4364 | slice = &slice[slices_bytes_len..]; | |
4365 | // SAFETY: Since PatternID is always representable as a u32, all we | |
4366 | // need to do is ensure that we have the proper length and alignment. | |
4367 | // We've checked both above, so the cast below is safe. | |
4368 | // | |
781aab86 FG |
4369 | // N.B. This is one of the few not-safe snippets in this function, |
4370 | // so we mark it explicitly to call it out. | |
4371 | let slices = core::slice::from_raw_parts( | |
4372 | slices_bytes.as_ptr().cast::<u32>(), | |
4373 | pair_len, | |
4374 | ); | |
487cf647 FG |
4375 | |
4376 | // Read the total number of unique pattern IDs (which is always 1 more | |
4377 | // than the maximum pattern ID in this automaton, since pattern IDs are | |
4378 | // handed out contiguously starting at 0). | |
781aab86 FG |
4379 | let (pattern_len, nr) = |
4380 | wire::try_read_u32_as_usize(slice, "pattern length")?; | |
487cf647 FG |
4381 | slice = &slice[nr..]; |
4382 | ||
781aab86 | 4383 | // Now read the pattern ID length. We don't need to store this |
487cf647 | 4384 | // explicitly, but we need it to know how many pattern IDs to read. |
781aab86 FG |
4385 | let (idlen, nr) = |
4386 | wire::try_read_u32_as_usize(slice, "pattern ID length")?; | |
487cf647 FG |
4387 | slice = &slice[nr..]; |
4388 | ||
4389 | // Read the actual pattern IDs. | |
4390 | let pattern_ids_len = | |
781aab86 FG |
4391 | wire::mul(idlen, PatternID::SIZE, "pattern ID byte length")?; |
4392 | wire::check_slice_len(slice, pattern_ids_len, "match pattern IDs")?; | |
4393 | wire::check_alignment::<PatternID>(slice)?; | |
487cf647 FG |
4394 | let pattern_ids_bytes = &slice[..pattern_ids_len]; |
4395 | slice = &slice[pattern_ids_len..]; | |
4396 | // SAFETY: Since PatternID is always representable as a u32, all we | |
4397 | // need to do is ensure that we have the proper length and alignment. | |
4398 | // We've checked both above, so the cast below is safe. | |
4399 | // | |
781aab86 FG |
4400 | // N.B. This is one of the few not-safe snippets in this function, |
4401 | // so we mark it explicitly to call it out. | |
4402 | let pattern_ids = core::slice::from_raw_parts( | |
4403 | pattern_ids_bytes.as_ptr().cast::<u32>(), | |
4404 | idlen, | |
4405 | ); | |
487cf647 | 4406 | |
781aab86 FG |
4407 | let ms = MatchStates { slices, pattern_ids, pattern_len }; |
4408 | Ok((ms, slice.as_ptr().as_usize() - slice_start)) | |
487cf647 FG |
4409 | } |
4410 | } | |
4411 | ||
781aab86 | 4412 | #[cfg(feature = "dfa-build")] |
487cf647 | 4413 | impl MatchStates<Vec<u32>> { |
781aab86 FG |
4414 | fn empty(pattern_len: usize) -> MatchStates<Vec<u32>> { |
4415 | assert!(pattern_len <= PatternID::LIMIT); | |
4416 | MatchStates { slices: vec![], pattern_ids: vec![], pattern_len } | |
487cf647 FG |
4417 | } |
4418 | ||
4419 | fn new( | |
4420 | matches: &BTreeMap<StateID, Vec<PatternID>>, | |
781aab86 FG |
4421 | pattern_len: usize, |
4422 | ) -> Result<MatchStates<Vec<u32>>, BuildError> { | |
4423 | let mut m = MatchStates::empty(pattern_len); | |
487cf647 FG |
4424 | for (_, pids) in matches.iter() { |
4425 | let start = PatternID::new(m.pattern_ids.len()) | |
781aab86 | 4426 | .map_err(|_| BuildError::too_many_match_pattern_ids())?; |
487cf647 FG |
4427 | m.slices.push(start.as_u32()); |
4428 | // This is always correct since the number of patterns in a single | |
4429 | // match state can never exceed maximum number of allowable | |
4430 | // patterns. Why? Because a pattern can only appear once in a | |
4431 | // particular match state, by construction. (And since our pattern | |
4432 | // ID limit is one less than u32::MAX, we're guaranteed that the | |
4433 | // length fits in a u32.) | |
4434 | m.slices.push(u32::try_from(pids.len()).unwrap()); | |
4435 | for &pid in pids { | |
4436 | m.pattern_ids.push(pid.as_u32()); | |
4437 | } | |
4438 | } | |
781aab86 | 4439 | m.pattern_len = pattern_len; |
487cf647 FG |
4440 | Ok(m) |
4441 | } | |
4442 | ||
4443 | fn new_with_map( | |
4444 | &self, | |
4445 | matches: &BTreeMap<StateID, Vec<PatternID>>, | |
781aab86 FG |
4446 | ) -> Result<MatchStates<Vec<u32>>, BuildError> { |
4447 | MatchStates::new(matches, self.pattern_len) | |
487cf647 FG |
4448 | } |
4449 | } | |
4450 | ||
4451 | impl<T: AsRef<[u32]>> MatchStates<T> { | |
4452 | /// Writes a serialized form of these match states to the buffer given. If | |
4453 | /// the buffer is too small, then an error is returned. To determine how | |
4454 | /// big the buffer must be, use `write_to_len`. | |
4455 | fn write_to<E: Endian>( | |
4456 | &self, | |
4457 | mut dst: &mut [u8], | |
4458 | ) -> Result<usize, SerializeError> { | |
4459 | let nwrite = self.write_to_len(); | |
4460 | if dst.len() < nwrite { | |
4461 | return Err(SerializeError::buffer_too_small("match states")); | |
4462 | } | |
4463 | dst = &mut dst[..nwrite]; | |
4464 | ||
781aab86 | 4465 | // write state ID length |
487cf647 | 4466 | // Unwrap is OK since number of states is guaranteed to fit in a u32. |
781aab86 | 4467 | E::write_u32(u32::try_from(self.len()).unwrap(), dst); |
487cf647 FG |
4468 | dst = &mut dst[size_of::<u32>()..]; |
4469 | ||
4470 | // write slice offset pairs | |
4471 | for &pid in self.slices() { | |
781aab86 | 4472 | let n = wire::write_pattern_id::<E>(pid, &mut dst); |
487cf647 FG |
4473 | dst = &mut dst[n..]; |
4474 | } | |
4475 | ||
781aab86 | 4476 | // write unique pattern ID length |
487cf647 | 4477 | // Unwrap is OK since number of patterns is guaranteed to fit in a u32. |
781aab86 | 4478 | E::write_u32(u32::try_from(self.pattern_len).unwrap(), dst); |
487cf647 FG |
4479 | dst = &mut dst[size_of::<u32>()..]; |
4480 | ||
781aab86 | 4481 | // write pattern ID length |
487cf647 FG |
4482 | // Unwrap is OK since we check at construction (and deserialization) |
4483 | // that the number of patterns is representable as a u32. | |
4484 | E::write_u32(u32::try_from(self.pattern_ids().len()).unwrap(), dst); | |
4485 | dst = &mut dst[size_of::<u32>()..]; | |
4486 | ||
4487 | // write pattern IDs | |
4488 | for &pid in self.pattern_ids() { | |
781aab86 | 4489 | let n = wire::write_pattern_id::<E>(pid, &mut dst); |
487cf647 FG |
4490 | dst = &mut dst[n..]; |
4491 | } | |
4492 | ||
4493 | Ok(nwrite) | |
4494 | } | |
4495 | ||
781aab86 FG |
4496 | /// Returns the number of bytes the serialized form of these match states |
4497 | /// will use. | |
487cf647 | 4498 | fn write_to_len(&self) -> usize { |
781aab86 | 4499 | size_of::<u32>() // match state length |
487cf647 | 4500 | + (self.slices().len() * PatternID::SIZE) |
781aab86 FG |
4501 | + size_of::<u32>() // unique pattern ID length |
4502 | + size_of::<u32>() // pattern ID length | |
487cf647 FG |
4503 | + (self.pattern_ids().len() * PatternID::SIZE) |
4504 | } | |
4505 | ||
4506 | /// Valides that the match state info is itself internally consistent and | |
4507 | /// consistent with the recorded match state region in the given DFA. | |
4508 | fn validate(&self, dfa: &DFA<T>) -> Result<(), DeserializeError> { | |
781aab86 | 4509 | if self.len() != dfa.special.match_len(dfa.stride()) { |
487cf647 | 4510 | return Err(DeserializeError::generic( |
781aab86 | 4511 | "match state length mismatch", |
487cf647 FG |
4512 | )); |
4513 | } | |
781aab86 | 4514 | for si in 0..self.len() { |
487cf647 FG |
4515 | let start = self.slices()[si * 2].as_usize(); |
4516 | let len = self.slices()[si * 2 + 1].as_usize(); | |
4517 | if start >= self.pattern_ids().len() { | |
4518 | return Err(DeserializeError::generic( | |
4519 | "invalid pattern ID start offset", | |
4520 | )); | |
4521 | } | |
4522 | if start + len > self.pattern_ids().len() { | |
4523 | return Err(DeserializeError::generic( | |
4524 | "invalid pattern ID length", | |
4525 | )); | |
4526 | } | |
4527 | for mi in 0..len { | |
4528 | let pid = self.pattern_id(si, mi); | |
781aab86 | 4529 | if pid.as_usize() >= self.pattern_len { |
487cf647 FG |
4530 | return Err(DeserializeError::generic( |
4531 | "invalid pattern ID", | |
4532 | )); | |
4533 | } | |
4534 | } | |
4535 | } | |
4536 | Ok(()) | |
4537 | } | |
4538 | ||
4539 | /// Converts these match states back into their map form. This is useful | |
4540 | /// when shuffling states, as the normal MatchStates representation is not | |
4541 | /// amenable to easy state swapping. But with this map, to swap id1 and | |
4542 | /// id2, all you need to do is: | |
4543 | /// | |
4544 | /// if let Some(pids) = map.remove(&id1) { | |
4545 | /// map.insert(id2, pids); | |
4546 | /// } | |
4547 | /// | |
4548 | /// Once shuffling is done, use MatchStates::new to convert back. | |
781aab86 | 4549 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
4550 | fn to_map(&self, dfa: &DFA<T>) -> BTreeMap<StateID, Vec<PatternID>> { |
4551 | let mut map = BTreeMap::new(); | |
781aab86 | 4552 | for i in 0..self.len() { |
487cf647 FG |
4553 | let mut pids = vec![]; |
4554 | for j in 0..self.pattern_len(i) { | |
4555 | pids.push(self.pattern_id(i, j)); | |
4556 | } | |
4557 | map.insert(self.match_state_id(dfa, i), pids); | |
4558 | } | |
4559 | map | |
4560 | } | |
4561 | ||
4562 | /// Converts these match states to a borrowed value. | |
4563 | fn as_ref(&self) -> MatchStates<&'_ [u32]> { | |
4564 | MatchStates { | |
4565 | slices: self.slices.as_ref(), | |
4566 | pattern_ids: self.pattern_ids.as_ref(), | |
781aab86 | 4567 | pattern_len: self.pattern_len, |
487cf647 FG |
4568 | } |
4569 | } | |
4570 | ||
4571 | /// Converts these match states to an owned value. | |
4572 | #[cfg(feature = "alloc")] | |
781aab86 | 4573 | fn to_owned(&self) -> MatchStates<alloc::vec::Vec<u32>> { |
487cf647 FG |
4574 | MatchStates { |
4575 | slices: self.slices.as_ref().to_vec(), | |
4576 | pattern_ids: self.pattern_ids.as_ref().to_vec(), | |
781aab86 | 4577 | pattern_len: self.pattern_len, |
487cf647 FG |
4578 | } |
4579 | } | |
4580 | ||
4581 | /// Returns the match state ID given the match state index. (Where the | |
4582 | /// first match state corresponds to index 0.) | |
4583 | /// | |
4584 | /// This panics if there is no match state at the given index. | |
4585 | fn match_state_id(&self, dfa: &DFA<T>, index: usize) -> StateID { | |
4586 | assert!(dfa.special.matches(), "no match states to index"); | |
4587 | // This is one of the places where we rely on the fact that match | |
4588 | // states are contiguous in the transition table. Namely, that the | |
4589 | // first match state ID always corresponds to dfa.special.min_start. | |
4590 | // From there, since we know the stride, we can compute the ID of any | |
4591 | // match state given its index. | |
4592 | let stride2 = u32::try_from(dfa.stride2()).unwrap(); | |
4593 | let offset = index.checked_shl(stride2).unwrap(); | |
4594 | let id = dfa.special.min_match.as_usize().checked_add(offset).unwrap(); | |
4595 | let sid = StateID::new(id).unwrap(); | |
4596 | assert!(dfa.is_match_state(sid)); | |
4597 | sid | |
4598 | } | |
4599 | ||
4600 | /// Returns the pattern ID at the given match index for the given match | |
4601 | /// state. | |
4602 | /// | |
4603 | /// The match state index is the state index minus the state index of the | |
4604 | /// first match state in the DFA. | |
4605 | /// | |
4606 | /// The match index is the index of the pattern ID for the given state. | |
4607 | /// The index must be less than `self.pattern_len(state_index)`. | |
781aab86 | 4608 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
4609 | fn pattern_id(&self, state_index: usize, match_index: usize) -> PatternID { |
4610 | self.pattern_id_slice(state_index)[match_index] | |
4611 | } | |
4612 | ||
4613 | /// Returns the number of patterns in the given match state. | |
4614 | /// | |
4615 | /// The match state index is the state index minus the state index of the | |
4616 | /// first match state in the DFA. | |
781aab86 | 4617 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
4618 | fn pattern_len(&self, state_index: usize) -> usize { |
4619 | self.slices()[state_index * 2 + 1].as_usize() | |
4620 | } | |
4621 | ||
4622 | /// Returns all of the pattern IDs for the given match state index. | |
4623 | /// | |
4624 | /// The match state index is the state index minus the state index of the | |
4625 | /// first match state in the DFA. | |
781aab86 | 4626 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 FG |
4627 | fn pattern_id_slice(&self, state_index: usize) -> &[PatternID] { |
4628 | let start = self.slices()[state_index * 2].as_usize(); | |
4629 | let len = self.pattern_len(state_index); | |
4630 | &self.pattern_ids()[start..start + len] | |
4631 | } | |
4632 | ||
4633 | /// Returns the pattern ID offset slice of u32 as a slice of PatternID. | |
781aab86 | 4634 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 | 4635 | fn slices(&self) -> &[PatternID] { |
781aab86 | 4636 | wire::u32s_to_pattern_ids(self.slices.as_ref()) |
487cf647 FG |
4637 | } |
4638 | ||
4639 | /// Returns the total number of match states. | |
781aab86 FG |
4640 | #[cfg_attr(feature = "perf-inline", inline(always))] |
4641 | fn len(&self) -> usize { | |
487cf647 FG |
4642 | assert_eq!(0, self.slices().len() % 2); |
4643 | self.slices().len() / 2 | |
4644 | } | |
4645 | ||
4646 | /// Returns the pattern ID slice of u32 as a slice of PatternID. | |
781aab86 | 4647 | #[cfg_attr(feature = "perf-inline", inline(always))] |
487cf647 | 4648 | fn pattern_ids(&self) -> &[PatternID] { |
781aab86 | 4649 | wire::u32s_to_pattern_ids(self.pattern_ids.as_ref()) |
487cf647 FG |
4650 | } |
4651 | ||
4652 | /// Return the memory usage, in bytes, of these match pairs. | |
4653 | fn memory_usage(&self) -> usize { | |
4654 | (self.slices().len() + self.pattern_ids().len()) * PatternID::SIZE | |
4655 | } | |
4656 | } | |
4657 | ||
781aab86 FG |
4658 | /// A common set of flags for both dense and sparse DFAs. This primarily |
4659 | /// centralizes the serialization format of these flags at a bitset. | |
4660 | #[derive(Clone, Copy, Debug)] | |
4661 | pub(crate) struct Flags { | |
4662 | /// Whether the DFA can match the empty string. When this is false, all | |
4663 | /// matches returned by this DFA are guaranteed to have non-zero length. | |
4664 | pub(crate) has_empty: bool, | |
4665 | /// Whether the DFA should only produce matches with spans that correspond | |
4666 | /// to valid UTF-8. This also includes omitting any zero-width matches that | |
4667 | /// split the UTF-8 encoding of a codepoint. | |
4668 | pub(crate) is_utf8: bool, | |
4669 | /// Whether the DFA is always anchored or not, regardless of `Input` | |
4670 | /// configuration. This is useful for avoiding a reverse scan even when | |
4671 | /// executing unanchored searches. | |
4672 | pub(crate) is_always_start_anchored: bool, | |
4673 | } | |
4674 | ||
4675 | impl Flags { | |
4676 | /// Creates a set of flags for a DFA from an NFA. | |
4677 | /// | |
4678 | /// N.B. This constructor was defined at the time of writing because all | |
4679 | /// of the flags are derived directly from the NFA. If this changes in the | |
4680 | /// future, we might be more thoughtful about how the `Flags` value is | |
4681 | /// itself built. | |
4682 | #[cfg(feature = "dfa-build")] | |
4683 | fn from_nfa(nfa: &thompson::NFA) -> Flags { | |
4684 | Flags { | |
4685 | has_empty: nfa.has_empty(), | |
4686 | is_utf8: nfa.is_utf8(), | |
4687 | is_always_start_anchored: nfa.is_always_start_anchored(), | |
4688 | } | |
4689 | } | |
4690 | ||
4691 | /// Deserializes the flags from the given slice. On success, this also | |
4692 | /// returns the number of bytes read from the slice. | |
4693 | pub(crate) fn from_bytes( | |
4694 | slice: &[u8], | |
4695 | ) -> Result<(Flags, usize), DeserializeError> { | |
4696 | let (bits, nread) = wire::try_read_u32(slice, "flag bitset")?; | |
4697 | let flags = Flags { | |
4698 | has_empty: bits & (1 << 0) != 0, | |
4699 | is_utf8: bits & (1 << 1) != 0, | |
4700 | is_always_start_anchored: bits & (1 << 2) != 0, | |
4701 | }; | |
4702 | Ok((flags, nread)) | |
4703 | } | |
4704 | ||
4705 | /// Writes these flags to the given byte slice. If the buffer is too small, | |
4706 | /// then an error is returned. To determine how big the buffer must be, | |
4707 | /// use `write_to_len`. | |
4708 | pub(crate) fn write_to<E: Endian>( | |
4709 | &self, | |
4710 | dst: &mut [u8], | |
4711 | ) -> Result<usize, SerializeError> { | |
4712 | fn bool_to_int(b: bool) -> u32 { | |
4713 | if b { | |
4714 | 1 | |
4715 | } else { | |
4716 | 0 | |
4717 | } | |
4718 | } | |
4719 | ||
4720 | let nwrite = self.write_to_len(); | |
4721 | if dst.len() < nwrite { | |
4722 | return Err(SerializeError::buffer_too_small("flag bitset")); | |
4723 | } | |
4724 | let bits = (bool_to_int(self.has_empty) << 0) | |
4725 | | (bool_to_int(self.is_utf8) << 1) | |
4726 | | (bool_to_int(self.is_always_start_anchored) << 2); | |
4727 | E::write_u32(bits, dst); | |
4728 | Ok(nwrite) | |
4729 | } | |
4730 | ||
4731 | /// Returns the number of bytes the serialized form of these flags | |
4732 | /// will use. | |
4733 | pub(crate) fn write_to_len(&self) -> usize { | |
4734 | size_of::<u32>() | |
4735 | } | |
4736 | } | |
4737 | ||
487cf647 FG |
4738 | /// An iterator over all states in a DFA. |
4739 | /// | |
4740 | /// This iterator yields a tuple for each state. The first element of the | |
4741 | /// tuple corresponds to a state's identifier, and the second element | |
4742 | /// corresponds to the state itself (comprised of its transitions). | |
4743 | /// | |
4744 | /// `'a` corresponding to the lifetime of original DFA, `T` corresponds to | |
4745 | /// the type of the transition table itself. | |
4746 | pub(crate) struct StateIter<'a, T> { | |
4747 | tt: &'a TransitionTable<T>, | |
4748 | it: iter::Enumerate<slice::Chunks<'a, StateID>>, | |
4749 | } | |
4750 | ||
4751 | impl<'a, T: AsRef<[u32]>> Iterator for StateIter<'a, T> { | |
4752 | type Item = State<'a>; | |
4753 | ||
4754 | fn next(&mut self) -> Option<State<'a>> { | |
4755 | self.it.next().map(|(index, _)| { | |
781aab86 | 4756 | let id = self.tt.to_state_id(index); |
487cf647 FG |
4757 | self.tt.state(id) |
4758 | }) | |
4759 | } | |
4760 | } | |
4761 | ||
4762 | /// An immutable representation of a single DFA state. | |
4763 | /// | |
4764 | /// `'a` correspondings to the lifetime of a DFA's transition table. | |
4765 | pub(crate) struct State<'a> { | |
4766 | id: StateID, | |
4767 | stride2: usize, | |
4768 | transitions: &'a [StateID], | |
4769 | } | |
4770 | ||
4771 | impl<'a> State<'a> { | |
4772 | /// Return an iterator over all transitions in this state. This yields | |
4773 | /// a number of transitions equivalent to the alphabet length of the | |
4774 | /// corresponding DFA. | |
4775 | /// | |
4776 | /// Each transition is represented by a tuple. The first element is | |
4777 | /// the input byte for that transition and the second element is the | |
4778 | /// transitions itself. | |
4779 | pub(crate) fn transitions(&self) -> StateTransitionIter<'_> { | |
4780 | StateTransitionIter { | |
4781 | len: self.transitions.len(), | |
4782 | it: self.transitions.iter().enumerate(), | |
4783 | } | |
4784 | } | |
4785 | ||
4786 | /// Return an iterator over a sparse representation of the transitions in | |
4787 | /// this state. Only non-dead transitions are returned. | |
4788 | /// | |
4789 | /// The "sparse" representation in this case corresponds to a sequence of | |
4790 | /// triples. The first two elements of the triple comprise an inclusive | |
4791 | /// byte range while the last element corresponds to the transition taken | |
4792 | /// for all bytes in the range. | |
4793 | /// | |
4794 | /// This is somewhat more condensed than the classical sparse | |
4795 | /// representation (where you have an element for every non-dead | |
4796 | /// transition), but in practice, checking if a byte is in a range is very | |
4797 | /// cheap and using ranges tends to conserve quite a bit more space. | |
4798 | pub(crate) fn sparse_transitions(&self) -> StateSparseTransitionIter<'_> { | |
4799 | StateSparseTransitionIter { dense: self.transitions(), cur: None } | |
4800 | } | |
4801 | ||
4802 | /// Returns the identifier for this state. | |
4803 | pub(crate) fn id(&self) -> StateID { | |
4804 | self.id | |
4805 | } | |
4806 | ||
4807 | /// Analyzes this state to determine whether it can be accelerated. If so, | |
4808 | /// it returns an accelerator that contains at least one byte. | |
781aab86 | 4809 | #[cfg(feature = "dfa-build")] |
487cf647 FG |
4810 | fn accelerate(&self, classes: &ByteClasses) -> Option<Accel> { |
4811 | // We just try to add bytes to our accelerator. Once adding fails | |
4812 | // (because we've added too many bytes), then give up. | |
4813 | let mut accel = Accel::new(); | |
4814 | for (class, id) in self.transitions() { | |
4815 | if id == self.id() { | |
4816 | continue; | |
4817 | } | |
4818 | for unit in classes.elements(class) { | |
4819 | if let Some(byte) = unit.as_u8() { | |
4820 | if !accel.add(byte) { | |
4821 | return None; | |
4822 | } | |
4823 | } | |
4824 | } | |
4825 | } | |
4826 | if accel.is_empty() { | |
4827 | None | |
4828 | } else { | |
4829 | Some(accel) | |
4830 | } | |
4831 | } | |
4832 | } | |
4833 | ||
4834 | impl<'a> fmt::Debug for State<'a> { | |
4835 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { | |
781aab86 FG |
4836 | for (i, (start, end, sid)) in self.sparse_transitions().enumerate() { |
4837 | let id = if f.alternate() { | |
4838 | sid.as_usize() | |
487cf647 | 4839 | } else { |
781aab86 | 4840 | sid.as_usize() >> self.stride2 |
487cf647 FG |
4841 | }; |
4842 | if i > 0 { | |
4843 | write!(f, ", ")?; | |
4844 | } | |
4845 | if start == end { | |
781aab86 | 4846 | write!(f, "{:?} => {:?}", start, id)?; |
487cf647 | 4847 | } else { |
781aab86 | 4848 | write!(f, "{:?}-{:?} => {:?}", start, end, id)?; |
487cf647 FG |
4849 | } |
4850 | } | |
4851 | Ok(()) | |
4852 | } | |
4853 | } | |
4854 | ||
487cf647 FG |
4855 | /// An iterator over all transitions in a single DFA state. This yields |
4856 | /// a number of transitions equivalent to the alphabet length of the | |
4857 | /// corresponding DFA. | |
4858 | /// | |
4859 | /// Each transition is represented by a tuple. The first element is the input | |
4860 | /// byte for that transition and the second element is the transition itself. | |
4861 | #[derive(Debug)] | |
4862 | pub(crate) struct StateTransitionIter<'a> { | |
4863 | len: usize, | |
4864 | it: iter::Enumerate<slice::Iter<'a, StateID>>, | |
4865 | } | |
4866 | ||
4867 | impl<'a> Iterator for StateTransitionIter<'a> { | |
4868 | type Item = (alphabet::Unit, StateID); | |
4869 | ||
4870 | fn next(&mut self) -> Option<(alphabet::Unit, StateID)> { | |
4871 | self.it.next().map(|(i, &id)| { | |
4872 | let unit = if i + 1 == self.len { | |
4873 | alphabet::Unit::eoi(i) | |
4874 | } else { | |
4875 | let b = u8::try_from(i) | |
4876 | .expect("raw byte alphabet is never exceeded"); | |
4877 | alphabet::Unit::u8(b) | |
4878 | }; | |
4879 | (unit, id) | |
4880 | }) | |
4881 | } | |
4882 | } | |
4883 | ||
487cf647 FG |
4884 | /// An iterator over all non-DEAD transitions in a single DFA state using a |
4885 | /// sparse representation. | |
4886 | /// | |
4887 | /// Each transition is represented by a triple. The first two elements of the | |
4888 | /// triple comprise an inclusive byte range while the last element corresponds | |
4889 | /// to the transition taken for all bytes in the range. | |
4890 | /// | |
4891 | /// As a convenience, this always returns `alphabet::Unit` values of the same | |
4892 | /// type. That is, you'll never get a (byte, EOI) or a (EOI, byte). Only (byte, | |
4893 | /// byte) and (EOI, EOI) values are yielded. | |
4894 | #[derive(Debug)] | |
4895 | pub(crate) struct StateSparseTransitionIter<'a> { | |
4896 | dense: StateTransitionIter<'a>, | |
4897 | cur: Option<(alphabet::Unit, alphabet::Unit, StateID)>, | |
4898 | } | |
4899 | ||
4900 | impl<'a> Iterator for StateSparseTransitionIter<'a> { | |
4901 | type Item = (alphabet::Unit, alphabet::Unit, StateID); | |
4902 | ||
4903 | fn next(&mut self) -> Option<(alphabet::Unit, alphabet::Unit, StateID)> { | |
4904 | while let Some((unit, next)) = self.dense.next() { | |
4905 | let (prev_start, prev_end, prev_next) = match self.cur { | |
4906 | Some(t) => t, | |
4907 | None => { | |
4908 | self.cur = Some((unit, unit, next)); | |
4909 | continue; | |
4910 | } | |
4911 | }; | |
4912 | if prev_next == next && !unit.is_eoi() { | |
4913 | self.cur = Some((prev_start, unit, prev_next)); | |
4914 | } else { | |
4915 | self.cur = Some((unit, unit, next)); | |
4916 | if prev_next != DEAD { | |
4917 | return Some((prev_start, prev_end, prev_next)); | |
4918 | } | |
4919 | } | |
4920 | } | |
4921 | if let Some((start, end, next)) = self.cur.take() { | |
4922 | if next != DEAD { | |
4923 | return Some((start, end, next)); | |
4924 | } | |
4925 | } | |
4926 | None | |
4927 | } | |
4928 | } | |
4929 | ||
781aab86 FG |
4930 | /// An error that occurred during the construction of a DFA. |
4931 | /// | |
4932 | /// This error does not provide many introspection capabilities. There are | |
4933 | /// generally only two things you can do with it: | |
4934 | /// | |
4935 | /// * Obtain a human readable message via its `std::fmt::Display` impl. | |
4936 | /// * Access an underlying [`nfa::thompson::BuildError`](thompson::BuildError) | |
4937 | /// type from its `source` method via the `std::error::Error` trait. This error | |
4938 | /// only occurs when using convenience routines for building a DFA directly | |
4939 | /// from a pattern string. | |
4940 | /// | |
4941 | /// When the `std` feature is enabled, this implements the `std::error::Error` | |
4942 | /// trait. | |
4943 | #[cfg(feature = "dfa-build")] | |
4944 | #[derive(Clone, Debug)] | |
4945 | pub struct BuildError { | |
4946 | kind: BuildErrorKind, | |
487cf647 FG |
4947 | } |
4948 | ||
781aab86 | 4949 | /// The kind of error that occurred during the construction of a DFA. |
487cf647 | 4950 | /// |
781aab86 FG |
4951 | /// Note that this error is non-exhaustive. Adding new variants is not |
4952 | /// considered a breaking change. | |
4953 | #[cfg(feature = "dfa-build")] | |
4954 | #[derive(Clone, Debug)] | |
4955 | enum BuildErrorKind { | |
4956 | /// An error that occurred while constructing an NFA as a precursor step | |
4957 | /// before a DFA is compiled. | |
4958 | NFA(thompson::BuildError), | |
4959 | /// An error that occurred because an unsupported regex feature was used. | |
4960 | /// The message string describes which unsupported feature was used. | |
4961 | /// | |
4962 | /// The primary regex feature that is unsupported by DFAs is the Unicode | |
4963 | /// word boundary look-around assertion (`\b`). This can be worked around | |
4964 | /// by either using an ASCII word boundary (`(?-u:\b)`) or by enabling | |
4965 | /// Unicode word boundaries when building a DFA. | |
4966 | Unsupported(&'static str), | |
4967 | /// An error that occurs if too many states are produced while building a | |
4968 | /// DFA. | |
4969 | TooManyStates, | |
4970 | /// An error that occurs if too many start states are needed while building | |
4971 | /// a DFA. | |
4972 | /// | |
4973 | /// This is a kind of oddball error that occurs when building a DFA with | |
4974 | /// start states enabled for each pattern and enough patterns to cause | |
4975 | /// the table of start states to overflow `usize`. | |
4976 | TooManyStartStates, | |
4977 | /// This is another oddball error that can occur if there are too many | |
4978 | /// patterns spread out across too many match states. | |
4979 | TooManyMatchPatternIDs, | |
4980 | /// An error that occurs if the DFA got too big during determinization. | |
4981 | DFAExceededSizeLimit { limit: usize }, | |
4982 | /// An error that occurs if auxiliary storage (not the DFA) used during | |
4983 | /// determinization got too big. | |
4984 | DeterminizeExceededSizeLimit { limit: usize }, | |
487cf647 FG |
4985 | } |
4986 | ||
781aab86 FG |
4987 | #[cfg(feature = "dfa-build")] |
4988 | impl BuildError { | |
4989 | /// Return the kind of this error. | |
4990 | fn kind(&self) -> &BuildErrorKind { | |
4991 | &self.kind | |
4992 | } | |
4993 | ||
4994 | pub(crate) fn nfa(err: thompson::BuildError) -> BuildError { | |
4995 | BuildError { kind: BuildErrorKind::NFA(err) } | |
4996 | } | |
4997 | ||
4998 | pub(crate) fn unsupported_dfa_word_boundary_unicode() -> BuildError { | |
4999 | let msg = "cannot build DFAs for regexes with Unicode word \ | |
5000 | boundaries; switch to ASCII word boundaries, or \ | |
5001 | heuristically enable Unicode word boundaries or use a \ | |
5002 | different regex engine"; | |
5003 | BuildError { kind: BuildErrorKind::Unsupported(msg) } | |
5004 | } | |
5005 | ||
5006 | pub(crate) fn too_many_states() -> BuildError { | |
5007 | BuildError { kind: BuildErrorKind::TooManyStates } | |
5008 | } | |
5009 | ||
5010 | pub(crate) fn too_many_start_states() -> BuildError { | |
5011 | BuildError { kind: BuildErrorKind::TooManyStartStates } | |
5012 | } | |
5013 | ||
5014 | pub(crate) fn too_many_match_pattern_ids() -> BuildError { | |
5015 | BuildError { kind: BuildErrorKind::TooManyMatchPatternIDs } | |
5016 | } | |
5017 | ||
5018 | pub(crate) fn dfa_exceeded_size_limit(limit: usize) -> BuildError { | |
5019 | BuildError { kind: BuildErrorKind::DFAExceededSizeLimit { limit } } | |
5020 | } | |
5021 | ||
5022 | pub(crate) fn determinize_exceeded_size_limit(limit: usize) -> BuildError { | |
5023 | BuildError { | |
5024 | kind: BuildErrorKind::DeterminizeExceededSizeLimit { limit }, | |
487cf647 FG |
5025 | } |
5026 | } | |
781aab86 | 5027 | } |
487cf647 | 5028 | |
781aab86 FG |
5029 | #[cfg(all(feature = "std", feature = "dfa-build"))] |
5030 | impl std::error::Error for BuildError { | |
5031 | fn source(&self) -> Option<&(dyn std::error::Error + 'static)> { | |
5032 | match self.kind() { | |
5033 | BuildErrorKind::NFA(ref err) => Some(err), | |
5034 | _ => None, | |
5035 | } | |
487cf647 | 5036 | } |
781aab86 | 5037 | } |
487cf647 | 5038 | |
781aab86 FG |
5039 | #[cfg(feature = "dfa-build")] |
5040 | impl core::fmt::Display for BuildError { | |
5041 | fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { | |
5042 | match self.kind() { | |
5043 | BuildErrorKind::NFA(_) => write!(f, "error building NFA"), | |
5044 | BuildErrorKind::Unsupported(ref msg) => { | |
5045 | write!(f, "unsupported regex feature for DFAs: {}", msg) | |
487cf647 | 5046 | } |
781aab86 FG |
5047 | BuildErrorKind::TooManyStates => write!( |
5048 | f, | |
5049 | "number of DFA states exceeds limit of {}", | |
5050 | StateID::LIMIT, | |
5051 | ), | |
5052 | BuildErrorKind::TooManyStartStates => { | |
5053 | let stride = Start::len(); | |
5054 | // The start table has `stride` entries for starting states for | |
5055 | // the entire DFA, and then `stride` entries for each pattern | |
5056 | // if start states for each pattern are enabled (which is the | |
5057 | // only way this error can occur). Thus, the total number of | |
5058 | // patterns that can fit in the table is `stride` less than | |
5059 | // what we can allocate. | |
5060 | let max = usize::try_from(core::isize::MAX).unwrap(); | |
5061 | let limit = (max - stride) / stride; | |
5062 | write!( | |
5063 | f, | |
5064 | "compiling DFA with start states exceeds pattern \ | |
5065 | pattern limit of {}", | |
5066 | limit, | |
5067 | ) | |
487cf647 | 5068 | } |
781aab86 FG |
5069 | BuildErrorKind::TooManyMatchPatternIDs => write!( |
5070 | f, | |
5071 | "compiling DFA with total patterns in all match states \ | |
5072 | exceeds limit of {}", | |
5073 | PatternID::LIMIT, | |
5074 | ), | |
5075 | BuildErrorKind::DFAExceededSizeLimit { limit } => write!( | |
5076 | f, | |
5077 | "DFA exceeded size limit of {:?} during determinization", | |
5078 | limit, | |
5079 | ), | |
5080 | BuildErrorKind::DeterminizeExceededSizeLimit { limit } => { | |
5081 | write!(f, "determinization exceeded size limit of {:?}", limit) | |
487cf647 FG |
5082 | } |
5083 | } | |
487cf647 FG |
5084 | } |
5085 | } | |
5086 | ||
781aab86 | 5087 | #[cfg(all(test, feature = "syntax", feature = "dfa-build"))] |
487cf647 FG |
5088 | mod tests { |
5089 | use super::*; | |
5090 | ||
5091 | #[test] | |
5092 | fn errors_with_unicode_word_boundary() { | |
5093 | let pattern = r"\b"; | |
5094 | assert!(Builder::new().build(pattern).is_err()); | |
5095 | } | |
5096 | ||
5097 | #[test] | |
5098 | fn roundtrip_never_match() { | |
5099 | let dfa = DFA::never_match().unwrap(); | |
5100 | let (buf, _) = dfa.to_bytes_native_endian(); | |
5101 | let dfa: DFA<&[u32]> = DFA::from_bytes(&buf).unwrap().0; | |
5102 | ||
781aab86 | 5103 | assert_eq!(None, dfa.try_search_fwd(&Input::new("foo12345")).unwrap()); |
487cf647 FG |
5104 | } |
5105 | ||
5106 | #[test] | |
5107 | fn roundtrip_always_match() { | |
5108 | use crate::HalfMatch; | |
5109 | ||
5110 | let dfa = DFA::always_match().unwrap(); | |
5111 | let (buf, _) = dfa.to_bytes_native_endian(); | |
5112 | let dfa: DFA<&[u32]> = DFA::from_bytes(&buf).unwrap().0; | |
5113 | ||
5114 | assert_eq!( | |
5115 | Some(HalfMatch::must(0, 0)), | |
781aab86 | 5116 | dfa.try_search_fwd(&Input::new("foo12345")).unwrap() |
487cf647 FG |
5117 | ); |
5118 | } | |
781aab86 FG |
5119 | |
5120 | // See the analogous test in src/hybrid/dfa.rs. | |
5121 | #[test] | |
5122 | fn heuristic_unicode_reverse() { | |
5123 | let dfa = DFA::builder() | |
5124 | .configure(DFA::config().unicode_word_boundary(true)) | |
5125 | .thompson(thompson::Config::new().reverse(true)) | |
5126 | .build(r"\b[0-9]+\b") | |
5127 | .unwrap(); | |
5128 | ||
5129 | let input = Input::new("β123").range(2..); | |
5130 | let expected = MatchError::quit(0xB2, 1); | |
5131 | let got = dfa.try_search_rev(&input); | |
5132 | assert_eq!(Err(expected), got); | |
5133 | ||
5134 | let input = Input::new("123β").range(..3); | |
5135 | let expected = MatchError::quit(0xCE, 3); | |
5136 | let got = dfa.try_search_rev(&input); | |
5137 | assert_eq!(Err(expected), got); | |
5138 | } | |
487cf647 | 5139 | } |