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 |
7 | [`dense::Config`](Config) for building and configuring a dense DFA. |
8 | */ |
9 | |
10 | #[cfg (feature = "dfa-build" )] |
11 | use core::cmp; |
12 | use core::{convert::TryFrom, fmt, iter, mem::size_of, slice}; |
13 | |
14 | #[cfg (feature = "dfa-build" )] |
15 | use alloc::{ |
16 | collections::{BTreeMap, BTreeSet}, |
17 | vec, |
18 | vec::Vec, |
19 | }; |
20 | |
21 | #[cfg (feature = "dfa-build" )] |
22 | use crate::{ |
23 | dfa::{ |
24 | accel::Accel, determinize, minimize::Minimizer, remapper::Remapper, |
25 | sparse, |
26 | }, |
27 | nfa::thompson, |
28 | util::{look::LookMatcher, search::MatchKind}, |
29 | }; |
30 | use crate::{ |
31 | dfa::{ |
32 | accel::Accels, |
33 | automaton::{fmt_state_indicator, Automaton, StartError}, |
34 | special::Special, |
35 | start::StartKind, |
36 | DEAD, |
37 | }, |
38 | util::{ |
39 | alphabet::{self, ByteClasses, ByteSet}, |
40 | int::{Pointer, Usize}, |
41 | prefilter::Prefilter, |
42 | primitives::{PatternID, StateID}, |
43 | search::Anchored, |
44 | start::{self, Start, StartByteMap}, |
45 | wire::{self, DeserializeError, Endian, SerializeError}, |
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 | /// |
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 | /// |
64 | /// A dense DFA configuration is a simple data object that is typically used |
65 | /// with [`dense::Builder::configure`](self::Builder::configure). |
66 | /// |
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 |
70 | /// not by default) and an [`Anchored::Pattern`] mode is requested via |
71 | /// [`Input`](crate::Input). |
72 | #[cfg (feature = "dfa-build" )] |
73 | #[derive(Clone, Debug, Default)] |
74 | pub struct Config { |
75 | // As with other configuration types in this crate, we put all our knobs |
76 | // in options so that we can distinguish between "default" and "not set." |
77 | // This makes it possible to easily combine multiple configurations |
78 | // without default values overwriting explicitly specified values. See the |
79 | // 'overwrite' method. |
80 | // |
81 | // For docs on the fields below, see the corresponding method setters. |
82 | accelerate: Option<bool>, |
83 | pre: Option<Option<Prefilter>>, |
84 | minimize: Option<bool>, |
85 | match_kind: Option<MatchKind>, |
86 | start_kind: Option<StartKind>, |
87 | starts_for_each_pattern: Option<bool>, |
88 | byte_classes: Option<bool>, |
89 | unicode_word_boundary: Option<bool>, |
90 | quitset: Option<ByteSet>, |
91 | specialize_start_states: Option<bool>, |
92 | dfa_size_limit: Option<Option<usize>>, |
93 | determinize_size_limit: Option<Option<usize>>, |
94 | } |
95 | |
96 | #[cfg (feature = "dfa-build" )] |
97 | impl Config { |
98 | /// Return a new default dense DFA compiler configuration. |
99 | pub fn new() -> Config { |
100 | Config::default() |
101 | } |
102 | |
103 | /// Enable state acceleration. |
104 | /// |
105 | /// When enabled, DFA construction will analyze each state to determine |
106 | /// whether it is eligible for simple acceleration. Acceleration typically |
107 | /// occurs when most of a state's transitions loop back to itself, leaving |
108 | /// only a select few bytes that will exit the state. When this occurs, |
109 | /// other routines like `memchr` can be used to look for those bytes which |
110 | /// may be much faster than traversing the DFA. |
111 | /// |
112 | /// Callers may elect to disable this if consistent performance is more |
113 | /// desirable than variable performance. Namely, acceleration can sometimes |
114 | /// make searching slower than it otherwise would be if the transitions |
115 | /// that leave accelerated states are traversed frequently. |
116 | /// |
117 | /// See [`Automaton::accelerator`] for 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 | |
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 | |
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 | /// ``` |
277 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
278 | /// use regex_automata::{ |
279 | /// dfa::{Automaton, OverlappingState, dense}, |
280 | /// HalfMatch, Input, MatchKind, |
281 | /// }; |
282 | /// |
283 | /// let dfa = dense::Builder::new() |
284 | /// .configure(dense::Config::new().match_kind(MatchKind::All)) |
285 | /// .build_many(&[r"\w+$" , r"\S+$" ])?; |
286 | /// let input = Input::new("@foo" ); |
287 | /// let mut state = OverlappingState::start(); |
288 | /// |
289 | /// let expected = Some(HalfMatch::must(1, 4)); |
290 | /// dfa.try_search_overlapping_fwd(&input, &mut state)?; |
291 | /// assert_eq!(expected, state.get_match()); |
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)); |
299 | /// dfa.try_search_overlapping_fwd(&input, &mut state)?; |
300 | /// assert_eq!(expected, state.get_match()); |
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 | /// ``` |
317 | /// use regex_automata::{ |
318 | /// dfa::{dense, Automaton, StartKind}, |
319 | /// nfa::thompson::NFA, |
320 | /// Anchored, HalfMatch, Input, MatchKind, |
321 | /// }; |
322 | /// |
323 | /// let haystack = "123foobar456" .as_bytes(); |
324 | /// let pattern = r"[a-z]+r" ; |
325 | /// |
326 | /// let dfa_fwd = dense::DFA::new(pattern)?; |
327 | /// let dfa_rev = dense::Builder::new() |
328 | /// .thompson(NFA::config().reverse(true)) |
329 | /// .configure(dense::Config::new() |
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) |
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); |
341 | /// let got_fwd = dfa_fwd.try_search_fwd(&Input::new(haystack))?.unwrap(); |
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. |
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(); |
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 | |
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 | |
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::{ |
442 | /// dfa::{dense, Automaton}, |
443 | /// Anchored, HalfMatch, PatternID, Input, |
444 | /// }; |
445 | /// |
446 | /// let dfa = dense::Builder::new() |
447 | /// .configure(dense::Config::new().starts_for_each_pattern(true)) |
448 | /// .build(r"foo[0-9]+" )?; |
449 | /// let haystack = "quux foo123" ; |
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); |
455 | /// let input = Input::new(haystack); |
456 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&input)?); |
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. |
462 | /// let input = Input::new(haystack) |
463 | /// .anchored(Anchored::Pattern(PatternID::must(0))); |
464 | /// assert_eq!(None, dfa.try_search_fwd(&input)?); |
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: |
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)?); |
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 |
491 | /// equivalence class because they never discriminate between a match and a |
492 | /// non-match. |
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 |
518 | /// [`MatchError::quit`](crate::MatchError::quit) error is returned. |
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 |
542 | /// [`MatchError::quit`](crate::MatchError::quit) error will never be |
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}, |
556 | /// HalfMatch, Input, MatchError, |
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. |
565 | /// let haystack = "foo 123 ☃" .as_bytes(); |
566 | /// let expected = Some(HalfMatch::must(0, 7)); |
567 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
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. |
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); |
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 |
608 | /// a [`MatchError::quit`](crate::MatchError::quit) error indicating the |
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 | /// ``` |
650 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
651 | /// use regex_automata::{dfa::{Automaton, dense}, Input, MatchError}; |
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. |
661 | /// let expected = MatchError::quit(b' \n' , 3); |
662 | /// let got = dfa.try_search_fwd(&Input::new(haystack)).unwrap_err(); |
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 | } |
674 | if self.quitset.is_none() { |
675 | self.quitset = Some(ByteSet::empty()); |
676 | } |
677 | if yes { |
678 | self.quitset.as_mut().unwrap().add(byte); |
679 | } else { |
680 | self.quitset.as_mut().unwrap().remove(byte); |
681 | } |
682 | self |
683 | } |
684 | |
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 | |
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 | /// ``` |
793 | /// # if cfg!(miri) { return Ok(()); } // miri takes too long |
794 | /// use regex_automata::{dfa::{dense, Automaton}, Input}; |
795 | /// |
796 | /// // 6MB isn't enough! |
797 | /// dense::Builder::new() |
798 | /// .configure(dense::Config::new().dfa_size_limit(Some(6_000_000))) |
799 | /// .build(r"\w{20}" ) |
800 | /// .unwrap_err(); |
801 | /// |
802 | /// // ... but 7MB probably is! |
803 | /// // (Note that DFA sizes aren't necessarily stable between releases.) |
804 | /// let dfa = dense::Builder::new() |
805 | /// .configure(dense::Config::new().dfa_size_limit(Some(7_000_000))) |
806 | /// .build(r"\w{20}" )?; |
807 | /// let haystack = "A" .repeat(20).into_bytes(); |
808 | /// assert!(dfa.try_search_fwd(&Input::new(&haystack))?.is_some()); |
809 | /// |
810 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
811 | /// ``` |
812 | /// |
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 |
815 | /// `(?-u:\w{20})`. So only use Unicode if you need it! |
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 | /// ``` |
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 | /// ``` |
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}; |
884 | /// |
885 | /// // 700KB isn't enough! |
886 | /// dense::Builder::new() |
887 | /// .configure(dense::Config::new() |
888 | /// .determinize_size_limit(Some(700_000)) |
889 | /// ) |
890 | /// .build(r"\w{20}" ) |
891 | /// .unwrap_err(); |
892 | /// |
893 | /// // ... but 800KB 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(800_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) |
927 | /// ) |
928 | /// .build(r"\w{20}" ) |
929 | /// .unwrap_err(); |
930 | /// |
931 | /// // ... but 300KB probably is! |
932 | /// // (Note that auxiliary storage sizes aren't necessarily stable between |
933 | /// // releases.) |
934 | /// let dfa = dense::Builder::new() |
935 | /// .configure(dense::Config::new() |
936 | /// .determinize_size_limit(Some(300_000)) |
937 | /// .start_kind(StartKind::Anchored) |
938 | /// ) |
939 | /// .build(r"\w{20}" )?; |
940 | /// let haystack = "A" .repeat(20).into_bytes(); |
941 | /// let input = Input::new(&haystack).anchored(Anchored::Yes); |
942 | /// assert!(dfa.try_search_fwd(&input)?.is_some()); |
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 | |
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 | |
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 | |
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 | |
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 | |
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 { |
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) |
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. |
1043 | pub(crate) fn overwrite(&self, o: Config) -> Config { |
1044 | Config { |
1045 | accelerate: o.accelerate.or(self.accelerate), |
1046 | pre: o.pre.or_else(|| self.pre.clone()), |
1047 | minimize: o.minimize.or(self.minimize), |
1048 | match_kind: o.match_kind.or(self.match_kind), |
1049 | start_kind: o.start_kind.or(self.start_kind), |
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), |
1057 | quitset: o.quitset.or(self.quitset), |
1058 | specialize_start_states: o |
1059 | .specialize_start_states |
1060 | .or(self.specialize_start_states), |
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. |
1111 | /// |
1112 | /// ``` |
1113 | /// use regex_automata::{ |
1114 | /// dfa::{Automaton, dense}, |
1115 | /// util::syntax, |
1116 | /// HalfMatch, Input, |
1117 | /// }; |
1118 | /// |
1119 | /// let dfa = dense::Builder::new() |
1120 | /// .configure(dense::Config::new().minimize(false)) |
1121 | /// .syntax(syntax::Config::new().unicode(false).utf8(false)) |
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)); |
1126 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
1127 | /// assert_eq!(expected, got); |
1128 | /// |
1129 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1130 | /// ``` |
1131 | #[cfg (feature = "dfa-build" )] |
1132 | #[derive(Clone, Debug)] |
1133 | pub struct Builder { |
1134 | config: Config, |
1135 | #[cfg (feature = "syntax" )] |
1136 | thompson: thompson::Compiler, |
1137 | } |
1138 | |
1139 | #[cfg (feature = "dfa-build" )] |
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(), |
1145 | #[cfg (feature = "syntax" )] |
1146 | thompson: thompson::Compiler::new(), |
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. |
1154 | #[cfg (feature = "syntax" )] |
1155 | pub fn build(&self, pattern: &str) -> Result<OwnedDFA, BuildError> { |
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. |
1163 | #[cfg (feature = "syntax" )] |
1164 | pub fn build_many<P: AsRef<str>>( |
1165 | &self, |
1166 | patterns: &[P], |
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)?; |
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}, |
1192 | /// nfa::thompson::NFA, |
1193 | /// HalfMatch, Input, |
1194 | /// }; |
1195 | /// |
1196 | /// let haystack = "foo123bar" .as_bytes(); |
1197 | /// |
1198 | /// // This shows how to set non-default options for building an NFA. |
1199 | /// let nfa = NFA::compiler() |
1200 | /// .configure(NFA::config().shrink(true)) |
1201 | /// .build(r"[0-9]+" )?; |
1202 | /// let dfa = dense::Builder::new().build_from_nfa(&nfa)?; |
1203 | /// let expected = Some(HalfMatch::must(0, 6)); |
1204 | /// let got = dfa.try_search_fwd(&Input::new(haystack))?; |
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, |
1212 | ) -> Result<OwnedDFA, BuildError> { |
1213 | let mut quitset = self.config.quitset.unwrap_or(ByteSet::empty()); |
1214 | if self.config.get_unicode_word_boundary() |
1215 | && nfa.look_set_any().contains_word_unicode() |
1216 | { |
1217 | for b in 0x80..=0xFF { |
1218 | quitset.add(b); |
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 |
1232 | // class as a quit byte, and thus cause the DFA to stop when it |
1233 | // shouldn't. |
1234 | // |
1235 | // Test case: |
1236 | // |
1237 | // regex-cli find match dense --unicode-word-boundary \ |
1238 | // -p '^#' -p '\b10\.55\.182\.100\b' -y @conn.json.1000x.log |
1239 | if !quitset.is_empty() { |
1240 | set.add_set(&quitset); |
1241 | } |
1242 | set.byte_classes() |
1243 | }; |
1244 | |
1245 | let mut dfa = DFA::initial( |
1246 | classes, |
1247 | nfa.pattern_len(), |
1248 | self.config.get_starts(), |
1249 | nfa.look_matcher(), |
1250 | self.config.get_starts_for_each_pattern(), |
1251 | self.config.get_prefilter().map(|p| p.clone()), |
1252 | quitset, |
1253 | Flags::from_nfa(&nfa), |
1254 | )?; |
1255 | determinize::Config::new() |
1256 | .match_kind(self.config.get_match_kind()) |
1257 | .quit(quitset) |
1258 | .dfa_size_limit(self.config.get_dfa_size_limit()) |
1259 | .determinize_size_limit(self.config.get_determinize_size_limit()) |
1260 | .run(nfa, &mut dfa)?; |
1261 | if self.config.get_minimize() { |
1262 | dfa.minimize(); |
1263 | } |
1264 | if self.config.get_accelerate() { |
1265 | dfa.accelerate(); |
1266 | } |
1267 | // The state shuffling done before this point always assumes that start |
1268 | // states should be marked as "special," even though it isn't the |
1269 | // default configuration. State shuffling is complex enough as it is, |
1270 | // so it's simpler to just "fix" our special state ID ranges to not |
1271 | // include starting states after-the-fact. |
1272 | if !self.config.get_specialize_start_states() { |
1273 | dfa.special.set_no_special_start_states(); |
1274 | } |
1275 | // Look for and set the universal starting states. |
1276 | dfa.set_universal_starts(); |
1277 | Ok(dfa) |
1278 | } |
1279 | |
1280 | /// Apply the given dense DFA configuration options to this builder. |
1281 | pub fn configure(&mut self, config: Config) -> &mut Builder { |
1282 | self.config = self.config.overwrite(config); |
1283 | self |
1284 | } |
1285 | |
1286 | /// Set the syntax configuration for this builder using |
1287 | /// [`syntax::Config`](crate::util::syntax::Config). |
1288 | /// |
1289 | /// This permits setting things like case insensitivity, Unicode and multi |
1290 | /// line mode. |
1291 | /// |
1292 | /// These settings only apply when constructing a DFA directly from a |
1293 | /// pattern. |
1294 | #[cfg (feature = "syntax" )] |
1295 | pub fn syntax( |
1296 | &mut self, |
1297 | config: crate::util::syntax::Config, |
1298 | ) -> &mut Builder { |
1299 | self.thompson.syntax(config); |
1300 | self |
1301 | } |
1302 | |
1303 | /// Set the Thompson NFA configuration for this builder using |
1304 | /// [`nfa::thompson::Config`](crate::nfa::thompson::Config). |
1305 | /// |
1306 | /// This permits setting things like whether the DFA should match the regex |
1307 | /// in reverse or if additional time should be spent shrinking the size of |
1308 | /// the NFA. |
1309 | /// |
1310 | /// These settings only apply when constructing a DFA directly from a |
1311 | /// pattern. |
1312 | #[cfg (feature = "syntax" )] |
1313 | pub fn thompson(&mut self, config: thompson::Config) -> &mut Builder { |
1314 | self.thompson.configure(config); |
1315 | self |
1316 | } |
1317 | } |
1318 | |
1319 | #[cfg (feature = "dfa-build" )] |
1320 | impl Default for Builder { |
1321 | fn default() -> Builder { |
1322 | Builder::new() |
1323 | } |
1324 | } |
1325 | |
1326 | /// A convenience alias for an owned DFA. We use this particular instantiation |
1327 | /// a lot in this crate, so it's worth giving it a name. This instantiation |
1328 | /// is commonly used for mutable APIs on the DFA while building it. The main |
1329 | /// reason for making DFAs generic is no_std support, and more generally, |
1330 | /// making it possible to load a DFA from an arbitrary slice of bytes. |
1331 | #[cfg (feature = "alloc" )] |
1332 | pub(crate) type OwnedDFA = DFA<alloc::vec::Vec<u32>>; |
1333 | |
1334 | /// A dense table-based deterministic finite automaton (DFA). |
1335 | /// |
1336 | /// All dense DFAs have one or more start states, zero or more match states |
1337 | /// and a transition table that maps the current state and the current byte |
1338 | /// of input to the next state. A DFA can use this information to implement |
1339 | /// fast searching. In particular, the use of a dense DFA generally makes the |
1340 | /// trade off that match speed is the most valuable characteristic, even if |
1341 | /// building the DFA may take significant time *and* space. (More concretely, |
1342 | /// building a DFA takes time and space that is exponential in the size of the |
1343 | /// pattern in the worst case.) As such, the processing of every byte of input |
1344 | /// is done with a small constant number of operations that does not vary with |
1345 | /// the pattern, its size or the size of the alphabet. If your needs don't line |
1346 | /// up with this trade off, then a dense DFA may not be an adequate solution to |
1347 | /// your problem. |
1348 | /// |
1349 | /// In contrast, a [`sparse::DFA`] makes the opposite |
1350 | /// trade off: it uses less space but will execute a variable number of |
1351 | /// instructions per byte at match time, which makes it slower for matching. |
1352 | /// (Note that space usage is still exponential in the size of the pattern in |
1353 | /// the worst case.) |
1354 | /// |
1355 | /// A DFA can be built using the default configuration via the |
1356 | /// [`DFA::new`] constructor. Otherwise, one can |
1357 | /// configure various aspects via [`dense::Builder`](Builder). |
1358 | /// |
1359 | /// A single DFA fundamentally supports the following operations: |
1360 | /// |
1361 | /// 1. Detection of a match. |
1362 | /// 2. Location of the end of a match. |
1363 | /// 3. In the case of a DFA with multiple patterns, which pattern matched is |
1364 | /// reported as well. |
1365 | /// |
1366 | /// A notable absence from the above list of capabilities is the location of |
1367 | /// the *start* of a match. In order to provide both the start and end of |
1368 | /// a match, *two* DFAs are required. This functionality is provided by a |
1369 | /// [`Regex`](crate::dfa::regex::Regex). |
1370 | /// |
1371 | /// # Type parameters |
1372 | /// |
1373 | /// A `DFA` has one type parameter, `T`, which is used to represent state IDs, |
1374 | /// pattern IDs and accelerators. `T` is typically a `Vec<u32>` or a `&[u32]`. |
1375 | /// |
1376 | /// # The `Automaton` trait |
1377 | /// |
1378 | /// This type implements the [`Automaton`] trait, which means it can be used |
1379 | /// for searching. For example: |
1380 | /// |
1381 | /// ``` |
1382 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1383 | /// |
1384 | /// let dfa = DFA::new("foo[0-9]+" )?; |
1385 | /// let expected = HalfMatch::must(0, 8); |
1386 | /// assert_eq!(Some(expected), dfa.try_search_fwd(&Input::new("foo12345" ))?); |
1387 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1388 | /// ``` |
1389 | #[derive(Clone)] |
1390 | pub struct DFA<T> { |
1391 | /// The transition table for this DFA. This includes the transitions |
1392 | /// themselves, along with the stride, number of states and the equivalence |
1393 | /// class mapping. |
1394 | tt: TransitionTable<T>, |
1395 | /// The set of starting state identifiers for this DFA. The starting state |
1396 | /// IDs act as pointers into the transition table. The specific starting |
1397 | /// state chosen for each search is dependent on the context at which the |
1398 | /// search begins. |
1399 | st: StartTable<T>, |
1400 | /// The set of match states and the patterns that match for each |
1401 | /// corresponding match state. |
1402 | /// |
1403 | /// This structure is technically only needed because of support for |
1404 | /// multi-regexes. Namely, multi-regexes require answering not just whether |
1405 | /// a match exists, but _which_ patterns match. So we need to store the |
1406 | /// matching pattern IDs for each match state. We do this even when there |
1407 | /// is only one pattern for the sake of simplicity. In practice, this uses |
1408 | /// up very little space for the case of one pattern. |
1409 | ms: MatchStates<T>, |
1410 | /// Information about which states are "special." Special states are states |
1411 | /// that are dead, quit, matching, starting or accelerated. For more info, |
1412 | /// see the docs for `Special`. |
1413 | special: Special, |
1414 | /// The accelerators for this DFA. |
1415 | /// |
1416 | /// If a state is accelerated, then there exist only a small number of |
1417 | /// bytes that can cause the DFA to leave the state. This permits searching |
1418 | /// to use optimized routines to find those specific bytes instead of using |
1419 | /// the transition table. |
1420 | /// |
1421 | /// All accelerated states exist in a contiguous range in the DFA's |
1422 | /// transition table. See dfa/special.rs for more details on how states are |
1423 | /// arranged. |
1424 | accels: Accels<T>, |
1425 | /// Any prefilter attached to this DFA. |
1426 | /// |
1427 | /// Note that currently prefilters are not serialized. When deserializing |
1428 | /// a DFA from bytes, this is always set to `None`. |
1429 | pre: Option<Prefilter>, |
1430 | /// The set of "quit" bytes for this DFA. |
1431 | /// |
1432 | /// This is only used when computing the start state for a particular |
1433 | /// position in a haystack. Namely, in the case where there is a quit |
1434 | /// byte immediately before the start of the search, this set needs to be |
1435 | /// explicitly consulted. In all other cases, quit bytes are detected by |
1436 | /// the DFA itself, by transitioning all quit bytes to a special "quit |
1437 | /// state." |
1438 | quitset: ByteSet, |
1439 | /// Various flags describing the behavior of this DFA. |
1440 | flags: Flags, |
1441 | } |
1442 | |
1443 | #[cfg (feature = "dfa-build" )] |
1444 | impl OwnedDFA { |
1445 | /// Parse the given regular expression using a default configuration and |
1446 | /// return the corresponding DFA. |
1447 | /// |
1448 | /// If you want a non-default configuration, then use the |
1449 | /// [`dense::Builder`](Builder) to set your own configuration. |
1450 | /// |
1451 | /// # Example |
1452 | /// |
1453 | /// ``` |
1454 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
1455 | /// |
1456 | /// let dfa = dense::DFA::new("foo[0-9]+bar" )?; |
1457 | /// let expected = Some(HalfMatch::must(0, 11)); |
1458 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345bar" ))?); |
1459 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1460 | /// ``` |
1461 | #[cfg (feature = "syntax" )] |
1462 | pub fn new(pattern: &str) -> Result<OwnedDFA, BuildError> { |
1463 | Builder::new().build(pattern) |
1464 | } |
1465 | |
1466 | /// Parse the given regular expressions using a default configuration and |
1467 | /// return the corresponding multi-DFA. |
1468 | /// |
1469 | /// If you want a non-default configuration, then use the |
1470 | /// [`dense::Builder`](Builder) to set your own configuration. |
1471 | /// |
1472 | /// # Example |
1473 | /// |
1474 | /// ``` |
1475 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
1476 | /// |
1477 | /// let dfa = dense::DFA::new_many(&["[0-9]+" , "[a-z]+" ])?; |
1478 | /// let expected = Some(HalfMatch::must(1, 3)); |
1479 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345bar" ))?); |
1480 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1481 | /// ``` |
1482 | #[cfg (feature = "syntax" )] |
1483 | pub fn new_many<P: AsRef<str>>( |
1484 | patterns: &[P], |
1485 | ) -> Result<OwnedDFA, BuildError> { |
1486 | Builder::new().build_many(patterns) |
1487 | } |
1488 | } |
1489 | |
1490 | #[cfg (feature = "dfa-build" )] |
1491 | impl OwnedDFA { |
1492 | /// Create a new DFA that matches every input. |
1493 | /// |
1494 | /// # Example |
1495 | /// |
1496 | /// ``` |
1497 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
1498 | /// |
1499 | /// let dfa = dense::DFA::always_match()?; |
1500 | /// |
1501 | /// let expected = Some(HalfMatch::must(0, 0)); |
1502 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("" ))?); |
1503 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo" ))?); |
1504 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1505 | /// ``` |
1506 | pub fn always_match() -> Result<OwnedDFA, BuildError> { |
1507 | let nfa = thompson::NFA::always_match(); |
1508 | Builder::new().build_from_nfa(&nfa) |
1509 | } |
1510 | |
1511 | /// Create a new DFA that never matches any input. |
1512 | /// |
1513 | /// # Example |
1514 | /// |
1515 | /// ``` |
1516 | /// use regex_automata::{dfa::{Automaton, dense}, Input}; |
1517 | /// |
1518 | /// let dfa = dense::DFA::never_match()?; |
1519 | /// assert_eq!(None, dfa.try_search_fwd(&Input::new("" ))?); |
1520 | /// assert_eq!(None, dfa.try_search_fwd(&Input::new("foo" ))?); |
1521 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1522 | /// ``` |
1523 | pub fn never_match() -> Result<OwnedDFA, BuildError> { |
1524 | let nfa = thompson::NFA::never_match(); |
1525 | Builder::new().build_from_nfa(&nfa) |
1526 | } |
1527 | |
1528 | /// Create an initial DFA with the given equivalence classes, pattern |
1529 | /// length and whether anchored starting states are enabled for each |
1530 | /// pattern. An initial DFA can be further mutated via determinization. |
1531 | fn initial( |
1532 | classes: ByteClasses, |
1533 | pattern_len: usize, |
1534 | starts: StartKind, |
1535 | lookm: &LookMatcher, |
1536 | starts_for_each_pattern: bool, |
1537 | pre: Option<Prefilter>, |
1538 | quitset: ByteSet, |
1539 | flags: Flags, |
1540 | ) -> Result<OwnedDFA, BuildError> { |
1541 | let start_pattern_len = |
1542 | if starts_for_each_pattern { Some(pattern_len) } else { None }; |
1543 | Ok(DFA { |
1544 | tt: TransitionTable::minimal(classes), |
1545 | st: StartTable::dead(starts, lookm, start_pattern_len)?, |
1546 | ms: MatchStates::empty(pattern_len), |
1547 | special: Special::new(), |
1548 | accels: Accels::empty(), |
1549 | pre, |
1550 | quitset, |
1551 | flags, |
1552 | }) |
1553 | } |
1554 | } |
1555 | |
1556 | #[cfg (feature = "dfa-build" )] |
1557 | impl DFA<&[u32]> { |
1558 | /// Return a new default dense DFA compiler configuration. |
1559 | /// |
1560 | /// This is a convenience routine to avoid needing to import the [`Config`] |
1561 | /// type when customizing the construction of a dense DFA. |
1562 | pub fn config() -> Config { |
1563 | Config::new() |
1564 | } |
1565 | |
1566 | /// Create a new dense DFA builder with the default configuration. |
1567 | /// |
1568 | /// This is a convenience routine to avoid needing to import the |
1569 | /// [`Builder`] type in common cases. |
1570 | pub fn builder() -> Builder { |
1571 | Builder::new() |
1572 | } |
1573 | } |
1574 | |
1575 | impl<T: AsRef<[u32]>> DFA<T> { |
1576 | /// Cheaply return a borrowed version of this dense DFA. Specifically, |
1577 | /// the DFA returned always uses `&[u32]` for its transition table. |
1578 | pub fn as_ref(&self) -> DFA<&'_ [u32]> { |
1579 | DFA { |
1580 | tt: self.tt.as_ref(), |
1581 | st: self.st.as_ref(), |
1582 | ms: self.ms.as_ref(), |
1583 | special: self.special, |
1584 | accels: self.accels(), |
1585 | pre: self.pre.clone(), |
1586 | quitset: self.quitset, |
1587 | flags: self.flags, |
1588 | } |
1589 | } |
1590 | |
1591 | /// Return an owned version of this sparse DFA. Specifically, the DFA |
1592 | /// returned always uses `Vec<u32>` for its transition table. |
1593 | /// |
1594 | /// Effectively, this returns a dense DFA whose transition table lives on |
1595 | /// the heap. |
1596 | #[cfg (feature = "alloc" )] |
1597 | pub fn to_owned(&self) -> OwnedDFA { |
1598 | DFA { |
1599 | tt: self.tt.to_owned(), |
1600 | st: self.st.to_owned(), |
1601 | ms: self.ms.to_owned(), |
1602 | special: self.special, |
1603 | accels: self.accels().to_owned(), |
1604 | pre: self.pre.clone(), |
1605 | quitset: self.quitset, |
1606 | flags: self.flags, |
1607 | } |
1608 | } |
1609 | |
1610 | /// Returns the starting state configuration for this DFA. |
1611 | /// |
1612 | /// The default is [`StartKind::Both`], which means the DFA supports both |
1613 | /// unanchored and anchored searches. However, this can generally lead to |
1614 | /// bigger DFAs. Therefore, a DFA might be compiled with support for just |
1615 | /// unanchored or anchored searches. In that case, running a search with |
1616 | /// an unsupported configuration will panic. |
1617 | pub fn start_kind(&self) -> StartKind { |
1618 | self.st.kind |
1619 | } |
1620 | |
1621 | /// Returns the start byte map used for computing the `Start` configuration |
1622 | /// at the beginning of a search. |
1623 | pub(crate) fn start_map(&self) -> &StartByteMap { |
1624 | &self.st.start_map |
1625 | } |
1626 | |
1627 | /// Returns true only if this DFA has starting states for each pattern. |
1628 | /// |
1629 | /// When a DFA has starting states for each pattern, then a search with the |
1630 | /// DFA can be configured to only look for anchored matches of a specific |
1631 | /// pattern. Specifically, APIs like [`Automaton::try_search_fwd`] can |
1632 | /// accept a non-None `pattern_id` if and only if this method returns true. |
1633 | /// Otherwise, calling `try_search_fwd` will panic. |
1634 | /// |
1635 | /// Note that if the DFA has no patterns, this always returns false. |
1636 | pub fn starts_for_each_pattern(&self) -> bool { |
1637 | self.st.pattern_len.is_some() |
1638 | } |
1639 | |
1640 | /// Returns the equivalence classes that make up the alphabet for this DFA. |
1641 | /// |
1642 | /// Unless [`Config::byte_classes`] was disabled, it is possible that |
1643 | /// multiple distinct bytes are grouped into the same equivalence class |
1644 | /// if it is impossible for them to discriminate between a match and a |
1645 | /// non-match. This has the effect of reducing the overall alphabet size |
1646 | /// and in turn potentially substantially reducing the size of the DFA's |
1647 | /// transition table. |
1648 | /// |
1649 | /// The downside of using equivalence classes like this is that every state |
1650 | /// transition will automatically use this map to convert an arbitrary |
1651 | /// byte to its corresponding equivalence class. In practice this has a |
1652 | /// negligible impact on performance. |
1653 | pub fn byte_classes(&self) -> &ByteClasses { |
1654 | &self.tt.classes |
1655 | } |
1656 | |
1657 | /// Returns the total number of elements in the alphabet for this DFA. |
1658 | /// |
1659 | /// That is, this returns the total number of transitions that each state |
1660 | /// in this DFA must have. Typically, a normal byte oriented DFA would |
1661 | /// always have an alphabet size of 256, corresponding to the number of |
1662 | /// unique values in a single byte. However, this implementation has two |
1663 | /// peculiarities that impact the alphabet length: |
1664 | /// |
1665 | /// * Every state has a special "EOI" transition that is only followed |
1666 | /// after the end of some haystack is reached. This EOI transition is |
1667 | /// necessary to account for one byte of look-ahead when implementing |
1668 | /// things like `\b` and `$`. |
1669 | /// * Bytes are grouped into equivalence classes such that no two bytes in |
1670 | /// the same class can distinguish a match from a non-match. For example, |
1671 | /// in the regex `^[a-z]+$`, the ASCII bytes `a-z` could all be in the |
1672 | /// same equivalence class. This leads to a massive space savings. |
1673 | /// |
1674 | /// Note though that the alphabet length does _not_ necessarily equal the |
1675 | /// total stride space taken up by a single DFA state in the transition |
1676 | /// table. Namely, for performance reasons, the stride is always the |
1677 | /// smallest power of two that is greater than or equal to the alphabet |
1678 | /// length. For this reason, [`DFA::stride`] or [`DFA::stride2`] are |
1679 | /// often more useful. The alphabet length is typically useful only for |
1680 | /// informational purposes. |
1681 | pub fn alphabet_len(&self) -> usize { |
1682 | self.tt.alphabet_len() |
1683 | } |
1684 | |
1685 | /// Returns the total stride for every state in this DFA, expressed as the |
1686 | /// exponent of a power of 2. The stride is the amount of space each state |
1687 | /// takes up in the transition table, expressed as a number of transitions. |
1688 | /// (Unused transitions map to dead states.) |
1689 | /// |
1690 | /// The stride of a DFA is always equivalent to the smallest power of 2 |
1691 | /// that is greater than or equal to the DFA's alphabet length. This |
1692 | /// definition uses extra space, but permits faster translation between |
1693 | /// premultiplied state identifiers and contiguous indices (by using shifts |
1694 | /// instead of relying on integer division). |
1695 | /// |
1696 | /// For example, if the DFA's stride is 16 transitions, then its `stride2` |
1697 | /// is `4` since `2^4 = 16`. |
1698 | /// |
1699 | /// The minimum `stride2` value is `1` (corresponding to a stride of `2`) |
1700 | /// while the maximum `stride2` value is `9` (corresponding to a stride of |
1701 | /// `512`). The maximum is not `8` since the maximum alphabet size is `257` |
1702 | /// when accounting for the special EOI transition. However, an alphabet |
1703 | /// length of that size is exceptionally rare since the alphabet is shrunk |
1704 | /// into equivalence classes. |
1705 | pub fn stride2(&self) -> usize { |
1706 | self.tt.stride2 |
1707 | } |
1708 | |
1709 | /// Returns the total stride for every state in this DFA. This corresponds |
1710 | /// to the total number of transitions used by each state in this DFA's |
1711 | /// transition table. |
1712 | /// |
1713 | /// Please see [`DFA::stride2`] for more information. In particular, this |
1714 | /// returns the stride as the number of transitions, where as `stride2` |
1715 | /// returns it as the exponent of a power of 2. |
1716 | pub fn stride(&self) -> usize { |
1717 | self.tt.stride() |
1718 | } |
1719 | |
1720 | /// Returns the memory usage, in bytes, of this DFA. |
1721 | /// |
1722 | /// The memory usage is computed based on the number of bytes used to |
1723 | /// represent this DFA. |
1724 | /// |
1725 | /// This does **not** include the stack size used up by this DFA. To |
1726 | /// compute that, use `std::mem::size_of::<dense::DFA>()`. |
1727 | pub fn memory_usage(&self) -> usize { |
1728 | self.tt.memory_usage() |
1729 | + self.st.memory_usage() |
1730 | + self.ms.memory_usage() |
1731 | + self.accels.memory_usage() |
1732 | } |
1733 | } |
1734 | |
1735 | /// Routines for converting a dense DFA to other representations, such as |
1736 | /// sparse DFAs or raw bytes suitable for persistent storage. |
1737 | impl<T: AsRef<[u32]>> DFA<T> { |
1738 | /// Convert this dense DFA to a sparse DFA. |
1739 | /// |
1740 | /// If a `StateID` is too small to represent all states in the sparse |
1741 | /// DFA, then this returns an error. In most cases, if a dense DFA is |
1742 | /// constructable with `StateID` then a sparse DFA will be as well. |
1743 | /// However, it is not guaranteed. |
1744 | /// |
1745 | /// # Example |
1746 | /// |
1747 | /// ``` |
1748 | /// use regex_automata::{dfa::{Automaton, dense}, HalfMatch, Input}; |
1749 | /// |
1750 | /// let dense = dense::DFA::new("foo[0-9]+" )?; |
1751 | /// let sparse = dense.to_sparse()?; |
1752 | /// |
1753 | /// let expected = Some(HalfMatch::must(0, 8)); |
1754 | /// assert_eq!(expected, sparse.try_search_fwd(&Input::new("foo12345" ))?); |
1755 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1756 | /// ``` |
1757 | #[cfg (feature = "dfa-build" )] |
1758 | pub fn to_sparse(&self) -> Result<sparse::DFA<Vec<u8>>, BuildError> { |
1759 | sparse::DFA::from_dense(self) |
1760 | } |
1761 | |
1762 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in little endian |
1763 | /// format. Upon success, the `Vec<u8>` and the initial padding length are |
1764 | /// returned. |
1765 | /// |
1766 | /// The written bytes are guaranteed to be deserialized correctly and |
1767 | /// without errors in a semver compatible release of this crate by a |
1768 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
1769 | /// deserialization APIs has been satisfied): |
1770 | /// |
1771 | /// * [`DFA::from_bytes`] |
1772 | /// * [`DFA::from_bytes_unchecked`] |
1773 | /// |
1774 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at |
1775 | /// an address that does not have the same alignment as `u32`. The padding |
1776 | /// corresponds to the number of leading bytes written to the returned |
1777 | /// `Vec<u8>`. |
1778 | /// |
1779 | /// # Example |
1780 | /// |
1781 | /// This example shows how to serialize and deserialize a DFA: |
1782 | /// |
1783 | /// ``` |
1784 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1785 | /// |
1786 | /// // Compile our original DFA. |
1787 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
1788 | /// |
1789 | /// // N.B. We use native endianness here to make the example work, but |
1790 | /// // using to_bytes_little_endian would work on a little endian target. |
1791 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); |
1792 | /// // Even if buf has initial padding, DFA::from_bytes will automatically |
1793 | /// // ignore it. |
1794 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; |
1795 | /// |
1796 | /// let expected = Some(HalfMatch::must(0, 8)); |
1797 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
1798 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1799 | /// ``` |
1800 | #[cfg (feature = "dfa-build" )] |
1801 | pub fn to_bytes_little_endian(&self) -> (Vec<u8>, usize) { |
1802 | self.to_bytes::<wire::LE>() |
1803 | } |
1804 | |
1805 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in big endian |
1806 | /// format. Upon success, the `Vec<u8>` and the initial padding length are |
1807 | /// returned. |
1808 | /// |
1809 | /// The written bytes are guaranteed to be deserialized correctly and |
1810 | /// without errors in a semver compatible release of this crate by a |
1811 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
1812 | /// deserialization APIs has been satisfied): |
1813 | /// |
1814 | /// * [`DFA::from_bytes`] |
1815 | /// * [`DFA::from_bytes_unchecked`] |
1816 | /// |
1817 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at |
1818 | /// an address that does not have the same alignment as `u32`. The padding |
1819 | /// corresponds to the number of leading bytes written to the returned |
1820 | /// `Vec<u8>`. |
1821 | /// |
1822 | /// # Example |
1823 | /// |
1824 | /// This example shows how to serialize and deserialize a DFA: |
1825 | /// |
1826 | /// ``` |
1827 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1828 | /// |
1829 | /// // Compile our original DFA. |
1830 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
1831 | /// |
1832 | /// // N.B. We use native endianness here to make the example work, but |
1833 | /// // using to_bytes_big_endian would work on a big endian target. |
1834 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); |
1835 | /// // Even if buf has initial padding, DFA::from_bytes will automatically |
1836 | /// // ignore it. |
1837 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; |
1838 | /// |
1839 | /// let expected = Some(HalfMatch::must(0, 8)); |
1840 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
1841 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1842 | /// ``` |
1843 | #[cfg (feature = "dfa-build" )] |
1844 | pub fn to_bytes_big_endian(&self) -> (Vec<u8>, usize) { |
1845 | self.to_bytes::<wire::BE>() |
1846 | } |
1847 | |
1848 | /// Serialize this DFA as raw bytes to a `Vec<u8>` in native endian |
1849 | /// format. Upon success, the `Vec<u8>` and the initial padding length are |
1850 | /// returned. |
1851 | /// |
1852 | /// The written bytes are guaranteed to be deserialized correctly and |
1853 | /// without errors in a semver compatible release of this crate by a |
1854 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
1855 | /// deserialization APIs has been satisfied): |
1856 | /// |
1857 | /// * [`DFA::from_bytes`] |
1858 | /// * [`DFA::from_bytes_unchecked`] |
1859 | /// |
1860 | /// The padding returned is non-zero if the returned `Vec<u8>` starts at |
1861 | /// an address that does not have the same alignment as `u32`. The padding |
1862 | /// corresponds to the number of leading bytes written to the returned |
1863 | /// `Vec<u8>`. |
1864 | /// |
1865 | /// Generally speaking, native endian format should only be used when |
1866 | /// you know that the target you're compiling the DFA for matches the |
1867 | /// endianness of the target on which you're compiling DFA. For example, |
1868 | /// if serialization and deserialization happen in the same process or on |
1869 | /// the same machine. Otherwise, when serializing a DFA for use in a |
1870 | /// portable environment, you'll almost certainly want to serialize _both_ |
1871 | /// a little endian and a big endian version and then load the correct one |
1872 | /// based on the target's configuration. |
1873 | /// |
1874 | /// # Example |
1875 | /// |
1876 | /// This example shows how to serialize and deserialize a DFA: |
1877 | /// |
1878 | /// ``` |
1879 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1880 | /// |
1881 | /// // Compile our original DFA. |
1882 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
1883 | /// |
1884 | /// let (buf, _) = original_dfa.to_bytes_native_endian(); |
1885 | /// // Even if buf has initial padding, DFA::from_bytes will automatically |
1886 | /// // ignore it. |
1887 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf)?.0; |
1888 | /// |
1889 | /// let expected = Some(HalfMatch::must(0, 8)); |
1890 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
1891 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1892 | /// ``` |
1893 | #[cfg (feature = "dfa-build" )] |
1894 | pub fn to_bytes_native_endian(&self) -> (Vec<u8>, usize) { |
1895 | self.to_bytes::<wire::NE>() |
1896 | } |
1897 | |
1898 | /// The implementation of the public `to_bytes` serialization methods, |
1899 | /// which is generic over endianness. |
1900 | #[cfg (feature = "dfa-build" )] |
1901 | fn to_bytes<E: Endian>(&self) -> (Vec<u8>, usize) { |
1902 | let len = self.write_to_len(); |
1903 | let (mut buf, padding) = wire::alloc_aligned_buffer::<u32>(len); |
1904 | // This should always succeed since the only possible serialization |
1905 | // error is providing a buffer that's too small, but we've ensured that |
1906 | // `buf` is big enough here. |
1907 | self.as_ref().write_to::<E>(&mut buf[padding..]).unwrap(); |
1908 | (buf, padding) |
1909 | } |
1910 | |
1911 | /// Serialize this DFA as raw bytes to the given slice, in little endian |
1912 | /// format. Upon success, the total number of bytes written to `dst` is |
1913 | /// returned. |
1914 | /// |
1915 | /// The written bytes are guaranteed to be deserialized correctly and |
1916 | /// without errors in a semver compatible release of this crate by a |
1917 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
1918 | /// deserialization APIs has been satisfied): |
1919 | /// |
1920 | /// * [`DFA::from_bytes`] |
1921 | /// * [`DFA::from_bytes_unchecked`] |
1922 | /// |
1923 | /// Note that unlike the various `to_byte_*` routines, this does not write |
1924 | /// any padding. Callers are responsible for handling alignment correctly. |
1925 | /// |
1926 | /// # Errors |
1927 | /// |
1928 | /// This returns an error if the given destination slice is not big enough |
1929 | /// to contain the full serialized DFA. If an error occurs, then nothing |
1930 | /// is written to `dst`. |
1931 | /// |
1932 | /// # Example |
1933 | /// |
1934 | /// This example shows how to serialize and deserialize a DFA without |
1935 | /// dynamic memory allocation. |
1936 | /// |
1937 | /// ``` |
1938 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1939 | /// |
1940 | /// // Compile our original DFA. |
1941 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
1942 | /// |
1943 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
1944 | /// // need to use a special type to force the alignment of our [u8; N] |
1945 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing |
1946 | /// // the DFA may fail because of an alignment mismatch. |
1947 | /// #[repr(C)] |
1948 | /// struct Aligned<B: ?Sized> { |
1949 | /// _align: [u32; 0], |
1950 | /// bytes: B, |
1951 | /// } |
1952 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; |
1953 | /// // N.B. We use native endianness here to make the example work, but |
1954 | /// // using write_to_little_endian would work on a little endian target. |
1955 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; |
1956 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; |
1957 | /// |
1958 | /// let expected = Some(HalfMatch::must(0, 8)); |
1959 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
1960 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
1961 | /// ``` |
1962 | pub fn write_to_little_endian( |
1963 | &self, |
1964 | dst: &mut [u8], |
1965 | ) -> Result<usize, SerializeError> { |
1966 | self.as_ref().write_to::<wire::LE>(dst) |
1967 | } |
1968 | |
1969 | /// Serialize this DFA as raw bytes to the given slice, in big endian |
1970 | /// format. Upon success, the total number of bytes written to `dst` is |
1971 | /// returned. |
1972 | /// |
1973 | /// The written bytes are guaranteed to be deserialized correctly and |
1974 | /// without errors in a semver compatible release of this crate by a |
1975 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
1976 | /// deserialization APIs has been satisfied): |
1977 | /// |
1978 | /// * [`DFA::from_bytes`] |
1979 | /// * [`DFA::from_bytes_unchecked`] |
1980 | /// |
1981 | /// Note that unlike the various `to_byte_*` routines, this does not write |
1982 | /// any padding. Callers are responsible for handling alignment correctly. |
1983 | /// |
1984 | /// # Errors |
1985 | /// |
1986 | /// This returns an error if the given destination slice is not big enough |
1987 | /// to contain the full serialized DFA. If an error occurs, then nothing |
1988 | /// is written to `dst`. |
1989 | /// |
1990 | /// # Example |
1991 | /// |
1992 | /// This example shows how to serialize and deserialize a DFA without |
1993 | /// dynamic memory allocation. |
1994 | /// |
1995 | /// ``` |
1996 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
1997 | /// |
1998 | /// // Compile our original DFA. |
1999 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
2000 | /// |
2001 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
2002 | /// // need to use a special type to force the alignment of our [u8; N] |
2003 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing |
2004 | /// // the DFA may fail because of an alignment mismatch. |
2005 | /// #[repr(C)] |
2006 | /// struct Aligned<B: ?Sized> { |
2007 | /// _align: [u32; 0], |
2008 | /// bytes: B, |
2009 | /// } |
2010 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; |
2011 | /// // N.B. We use native endianness here to make the example work, but |
2012 | /// // using write_to_big_endian would work on a big endian target. |
2013 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; |
2014 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; |
2015 | /// |
2016 | /// let expected = Some(HalfMatch::must(0, 8)); |
2017 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2018 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2019 | /// ``` |
2020 | pub fn write_to_big_endian( |
2021 | &self, |
2022 | dst: &mut [u8], |
2023 | ) -> Result<usize, SerializeError> { |
2024 | self.as_ref().write_to::<wire::BE>(dst) |
2025 | } |
2026 | |
2027 | /// Serialize this DFA as raw bytes to the given slice, in native endian |
2028 | /// format. Upon success, the total number of bytes written to `dst` is |
2029 | /// returned. |
2030 | /// |
2031 | /// The written bytes are guaranteed to be deserialized correctly and |
2032 | /// without errors in a semver compatible release of this crate by a |
2033 | /// `DFA`'s deserialization APIs (assuming all other criteria for the |
2034 | /// deserialization APIs has been satisfied): |
2035 | /// |
2036 | /// * [`DFA::from_bytes`] |
2037 | /// * [`DFA::from_bytes_unchecked`] |
2038 | /// |
2039 | /// Generally speaking, native endian format should only be used when |
2040 | /// you know that the target you're compiling the DFA for matches the |
2041 | /// endianness of the target on which you're compiling DFA. For example, |
2042 | /// if serialization and deserialization happen in the same process or on |
2043 | /// the same machine. Otherwise, when serializing a DFA for use in a |
2044 | /// portable environment, you'll almost certainly want to serialize _both_ |
2045 | /// a little endian and a big endian version and then load the correct one |
2046 | /// based on the target's configuration. |
2047 | /// |
2048 | /// Note that unlike the various `to_byte_*` routines, this does not write |
2049 | /// any padding. Callers are responsible for handling alignment correctly. |
2050 | /// |
2051 | /// # Errors |
2052 | /// |
2053 | /// This returns an error if the given destination slice is not big enough |
2054 | /// to contain the full serialized DFA. If an error occurs, then nothing |
2055 | /// is written to `dst`. |
2056 | /// |
2057 | /// # Example |
2058 | /// |
2059 | /// This example shows how to serialize and deserialize a DFA without |
2060 | /// dynamic memory allocation. |
2061 | /// |
2062 | /// ``` |
2063 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
2064 | /// |
2065 | /// // Compile our original DFA. |
2066 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
2067 | /// |
2068 | /// // Create a 4KB buffer on the stack to store our serialized DFA. We |
2069 | /// // need to use a special type to force the alignment of our [u8; N] |
2070 | /// // array to be aligned to a 4 byte boundary. Otherwise, deserializing |
2071 | /// // the DFA may fail because of an alignment mismatch. |
2072 | /// #[repr(C)] |
2073 | /// struct Aligned<B: ?Sized> { |
2074 | /// _align: [u32; 0], |
2075 | /// bytes: B, |
2076 | /// } |
2077 | /// let mut buf = Aligned { _align: [], bytes: [0u8; 4 * (1<<10)] }; |
2078 | /// let written = original_dfa.write_to_native_endian(&mut buf.bytes)?; |
2079 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&buf.bytes[..written])?.0; |
2080 | /// |
2081 | /// let expected = Some(HalfMatch::must(0, 8)); |
2082 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2083 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2084 | /// ``` |
2085 | pub fn write_to_native_endian( |
2086 | &self, |
2087 | dst: &mut [u8], |
2088 | ) -> Result<usize, SerializeError> { |
2089 | self.as_ref().write_to::<wire::NE>(dst) |
2090 | } |
2091 | |
2092 | /// Return the total number of bytes required to serialize this DFA. |
2093 | /// |
2094 | /// This is useful for determining the size of the buffer required to pass |
2095 | /// to one of the serialization routines: |
2096 | /// |
2097 | /// * [`DFA::write_to_little_endian`] |
2098 | /// * [`DFA::write_to_big_endian`] |
2099 | /// * [`DFA::write_to_native_endian`] |
2100 | /// |
2101 | /// Passing a buffer smaller than the size returned by this method will |
2102 | /// result in a serialization error. Serialization routines are guaranteed |
2103 | /// to succeed when the buffer is big enough. |
2104 | /// |
2105 | /// # Example |
2106 | /// |
2107 | /// This example shows how to dynamically allocate enough room to serialize |
2108 | /// a DFA. |
2109 | /// |
2110 | /// ``` |
2111 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
2112 | /// |
2113 | /// let original_dfa = DFA::new("foo[0-9]+" )?; |
2114 | /// |
2115 | /// let mut buf = vec![0; original_dfa.write_to_len()]; |
2116 | /// // This is guaranteed to succeed, because the only serialization error |
2117 | /// // that can occur is when the provided buffer is too small. But |
2118 | /// // write_to_len guarantees a correct size. |
2119 | /// let written = original_dfa.write_to_native_endian(&mut buf).unwrap(); |
2120 | /// // But this is not guaranteed to succeed! In particular, |
2121 | /// // deserialization requires proper alignment for &[u32], but our buffer |
2122 | /// // was allocated as a &[u8] whose required alignment is smaller than |
2123 | /// // &[u32]. However, it's likely to work in practice because of how most |
2124 | /// // allocators work. So if you write code like this, make sure to either |
2125 | /// // handle the error correctly and/or run it under Miri since Miri will |
2126 | /// // likely provoke the error by returning Vec<u8> buffers with alignment |
2127 | /// // less than &[u32]. |
2128 | /// let dfa: DFA<&[u32]> = match DFA::from_bytes(&buf[..written]) { |
2129 | /// // As mentioned above, it is legal for an error to be returned |
2130 | /// // here. It is quite difficult to get a Vec<u8> with a guaranteed |
2131 | /// // alignment equivalent to Vec<u32>. |
2132 | /// Err(_) => return Ok(()), |
2133 | /// Ok((dfa, _)) => dfa, |
2134 | /// }; |
2135 | /// |
2136 | /// let expected = Some(HalfMatch::must(0, 8)); |
2137 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2138 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2139 | /// ``` |
2140 | /// |
2141 | /// Note that this example isn't actually guaranteed to work! In |
2142 | /// particular, if `buf` is not aligned to a 4-byte boundary, then the |
2143 | /// `DFA::from_bytes` call will fail. If you need this to work, then you |
2144 | /// either need to deal with adding some initial padding yourself, or use |
2145 | /// one of the `to_bytes` methods, which will do it for you. |
2146 | pub fn write_to_len(&self) -> usize { |
2147 | wire::write_label_len(LABEL) |
2148 | + wire::write_endianness_check_len() |
2149 | + wire::write_version_len() |
2150 | + size_of::<u32>() // unused, intended for future flexibility |
2151 | + self.flags.write_to_len() |
2152 | + self.tt.write_to_len() |
2153 | + self.st.write_to_len() |
2154 | + self.ms.write_to_len() |
2155 | + self.special.write_to_len() |
2156 | + self.accels.write_to_len() |
2157 | + self.quitset.write_to_len() |
2158 | } |
2159 | } |
2160 | |
2161 | impl<'a> DFA<&'a [u32]> { |
2162 | /// Safely deserialize a DFA with a specific state identifier |
2163 | /// representation. Upon success, this returns both the deserialized DFA |
2164 | /// and the number of bytes read from the given slice. Namely, the contents |
2165 | /// of the slice beyond the DFA are not read. |
2166 | /// |
2167 | /// Deserializing a DFA using this routine will never allocate heap memory. |
2168 | /// For safety purposes, the DFA's transition table will be verified such |
2169 | /// that every transition points to a valid state. If this verification is |
2170 | /// too costly, then a [`DFA::from_bytes_unchecked`] API is provided, which |
2171 | /// will always execute in constant time. |
2172 | /// |
2173 | /// The bytes given must be generated by one of the serialization APIs |
2174 | /// of a `DFA` using a semver compatible release of this crate. Those |
2175 | /// include: |
2176 | /// |
2177 | /// * [`DFA::to_bytes_little_endian`] |
2178 | /// * [`DFA::to_bytes_big_endian`] |
2179 | /// * [`DFA::to_bytes_native_endian`] |
2180 | /// * [`DFA::write_to_little_endian`] |
2181 | /// * [`DFA::write_to_big_endian`] |
2182 | /// * [`DFA::write_to_native_endian`] |
2183 | /// |
2184 | /// The `to_bytes` methods allocate and return a `Vec<u8>` for you, along |
2185 | /// with handling alignment correctly. The `write_to` methods do not |
2186 | /// allocate and write to an existing slice (which may be on the stack). |
2187 | /// Since deserialization always uses the native endianness of the target |
2188 | /// platform, the serialization API you use should match the endianness of |
2189 | /// the target platform. (It's often a good idea to generate serialized |
2190 | /// DFAs for both forms of endianness and then load the correct one based |
2191 | /// on endianness.) |
2192 | /// |
2193 | /// # Errors |
2194 | /// |
2195 | /// Generally speaking, it's easier to state the conditions in which an |
2196 | /// error is _not_ returned. All of the following must be true: |
2197 | /// |
2198 | /// * The bytes given must be produced by one of the serialization APIs |
2199 | /// on this DFA, as mentioned above. |
2200 | /// * The endianness of the target platform matches the endianness used to |
2201 | /// serialized the provided DFA. |
2202 | /// * The slice given must have the same alignment as `u32`. |
2203 | /// |
2204 | /// If any of the above are not true, then an error will be returned. |
2205 | /// |
2206 | /// # Panics |
2207 | /// |
2208 | /// This routine will never panic for any input. |
2209 | /// |
2210 | /// # Example |
2211 | /// |
2212 | /// This example shows how to serialize a DFA to raw bytes, deserialize it |
2213 | /// and then use it for searching. |
2214 | /// |
2215 | /// ``` |
2216 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
2217 | /// |
2218 | /// let initial = DFA::new("foo[0-9]+" )?; |
2219 | /// let (bytes, _) = initial.to_bytes_native_endian(); |
2220 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&bytes)?.0; |
2221 | /// |
2222 | /// let expected = Some(HalfMatch::must(0, 8)); |
2223 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2224 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2225 | /// ``` |
2226 | /// |
2227 | /// # Example: dealing with alignment and padding |
2228 | /// |
2229 | /// In the above example, we used the `to_bytes_native_endian` method to |
2230 | /// serialize a DFA, but we ignored part of its return value corresponding |
2231 | /// to padding added to the beginning of the serialized DFA. This is OK |
2232 | /// because deserialization will skip this initial padding. What matters |
2233 | /// is that the address immediately following the padding has an alignment |
2234 | /// that matches `u32`. That is, the following is an equivalent but |
2235 | /// alternative way to write the above example: |
2236 | /// |
2237 | /// ``` |
2238 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
2239 | /// |
2240 | /// let initial = DFA::new("foo[0-9]+" )?; |
2241 | /// // Serialization returns the number of leading padding bytes added to |
2242 | /// // the returned Vec<u8>. |
2243 | /// let (bytes, pad) = initial.to_bytes_native_endian(); |
2244 | /// let dfa: DFA<&[u32]> = DFA::from_bytes(&bytes[pad..])?.0; |
2245 | /// |
2246 | /// let expected = Some(HalfMatch::must(0, 8)); |
2247 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2248 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2249 | /// ``` |
2250 | /// |
2251 | /// This padding is necessary because Rust's standard library does |
2252 | /// not expose any safe and robust way of creating a `Vec<u8>` with a |
2253 | /// guaranteed alignment other than 1. Now, in practice, the underlying |
2254 | /// allocator is likely to provide a `Vec<u8>` that meets our alignment |
2255 | /// requirements, which means `pad` is zero in practice most of the time. |
2256 | /// |
2257 | /// The purpose of exposing the padding like this is flexibility for the |
2258 | /// caller. For example, if one wants to embed a serialized DFA into a |
2259 | /// compiled program, then it's important to guarantee that it starts at a |
2260 | /// `u32`-aligned address. The simplest way to do this is to discard the |
2261 | /// padding bytes and set it up so that the serialized DFA itself begins at |
2262 | /// a properly aligned address. We can show this in two parts. The first |
2263 | /// part is serializing the DFA to a file: |
2264 | /// |
2265 | /// ```no_run |
2266 | /// use regex_automata::dfa::dense::DFA; |
2267 | /// |
2268 | /// let dfa = DFA::new("foo[0-9]+" )?; |
2269 | /// |
2270 | /// let (bytes, pad) = dfa.to_bytes_big_endian(); |
2271 | /// // Write the contents of the DFA *without* the initial padding. |
2272 | /// std::fs::write("foo.bigendian.dfa" , &bytes[pad..])?; |
2273 | /// |
2274 | /// // Do it again, but this time for little endian. |
2275 | /// let (bytes, pad) = dfa.to_bytes_little_endian(); |
2276 | /// std::fs::write("foo.littleendian.dfa" , &bytes[pad..])?; |
2277 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2278 | /// ``` |
2279 | /// |
2280 | /// And now the second part is embedding the DFA into the compiled program |
2281 | /// and deserializing it at runtime on first use. We use conditional |
2282 | /// compilation to choose the correct endianness. |
2283 | /// |
2284 | /// ```no_run |
2285 | /// use regex_automata::{ |
2286 | /// dfa::{Automaton, dense::DFA}, |
2287 | /// util::{lazy::Lazy, wire::AlignAs}, |
2288 | /// HalfMatch, Input, |
2289 | /// }; |
2290 | /// |
2291 | /// // This crate provides its own "lazy" type, kind of like |
2292 | /// // lazy_static! or once_cell::sync::Lazy. But it works in no-alloc |
2293 | /// // no-std environments and let's us write this using completely |
2294 | /// // safe code. |
2295 | /// static RE: Lazy<DFA<&'static [u32]>> = Lazy::new(|| { |
2296 | /// # const _: &str = stringify! { |
2297 | /// // This assignment is made possible (implicitly) via the |
2298 | /// // CoerceUnsized trait. This is what guarantees that our |
2299 | /// // bytes are stored in memory on a 4 byte boundary. You |
2300 | /// // *must* do this or something equivalent for correct |
2301 | /// // deserialization. |
2302 | /// static ALIGNED: &AlignAs<[u8], u32> = &AlignAs { |
2303 | /// _align: [], |
2304 | /// #[cfg(target_endian = "big" )] |
2305 | /// bytes: *include_bytes!("foo.bigendian.dfa" ), |
2306 | /// #[cfg(target_endian = "little" )] |
2307 | /// bytes: *include_bytes!("foo.littleendian.dfa" ), |
2308 | /// }; |
2309 | /// # }; |
2310 | /// # static ALIGNED: &AlignAs<[u8], u32> = &AlignAs { |
2311 | /// # _align: [], |
2312 | /// # bytes: [], |
2313 | /// # }; |
2314 | /// |
2315 | /// let (dfa, _) = DFA::from_bytes(&ALIGNED.bytes) |
2316 | /// .expect("serialized DFA should be valid" ); |
2317 | /// dfa |
2318 | /// }); |
2319 | /// |
2320 | /// let expected = Ok(Some(HalfMatch::must(0, 8))); |
2321 | /// assert_eq!(expected, RE.try_search_fwd(&Input::new("foo12345" ))); |
2322 | /// ``` |
2323 | /// |
2324 | /// An alternative to [`util::lazy::Lazy`](crate::util::lazy::Lazy) |
2325 | /// is [`lazy_static`](https://crates.io/crates/lazy_static) or |
2326 | /// [`once_cell`](https://crates.io/crates/once_cell), which provide |
2327 | /// stronger guarantees (like the initialization function only being |
2328 | /// executed once). And `once_cell` in particular provides a more |
2329 | /// expressive API. But a `Lazy` value from this crate is likely just fine |
2330 | /// in most circumstances. |
2331 | /// |
2332 | /// Note that regardless of which initialization method you use, you |
2333 | /// will still need to use the [`AlignAs`](crate::util::wire::AlignAs) |
2334 | /// trick above to force correct alignment, but this is safe to do and |
2335 | /// `from_bytes` will return an error if you get it wrong. |
2336 | pub fn from_bytes( |
2337 | slice: &'a [u8], |
2338 | ) -> Result<(DFA<&'a [u32]>, usize), DeserializeError> { |
2339 | // SAFETY: This is safe because we validate the transition table, start |
2340 | // table, match states and accelerators below. If any validation fails, |
2341 | // then we return an error. |
2342 | let (dfa, nread) = unsafe { DFA::from_bytes_unchecked(slice)? }; |
2343 | dfa.tt.validate(&dfa.special)?; |
2344 | dfa.st.validate(&dfa.tt)?; |
2345 | dfa.ms.validate(&dfa)?; |
2346 | dfa.accels.validate()?; |
2347 | // N.B. dfa.special doesn't have a way to do unchecked deserialization, |
2348 | // so it has already been validated. |
2349 | for state in dfa.states() { |
2350 | // If the state is an accel state, then it must have a non-empty |
2351 | // accelerator. |
2352 | if dfa.is_accel_state(state.id()) { |
2353 | let index = dfa.accelerator_index(state.id()); |
2354 | if index >= dfa.accels.len() { |
2355 | return Err(DeserializeError::generic( |
2356 | "found DFA state with invalid accelerator index" , |
2357 | )); |
2358 | } |
2359 | let needles = dfa.accels.needles(index); |
2360 | if !(1 <= needles.len() && needles.len() <= 3) { |
2361 | return Err(DeserializeError::generic( |
2362 | "accelerator needles has invalid length" , |
2363 | )); |
2364 | } |
2365 | } |
2366 | } |
2367 | Ok((dfa, nread)) |
2368 | } |
2369 | |
2370 | /// Deserialize a DFA with a specific state identifier representation in |
2371 | /// constant time by omitting the verification of the validity of the |
2372 | /// transition table and other data inside the DFA. |
2373 | /// |
2374 | /// This is just like [`DFA::from_bytes`], except it can potentially return |
2375 | /// a DFA that exhibits undefined behavior if its transition table contains |
2376 | /// invalid state identifiers. |
2377 | /// |
2378 | /// This routine is useful if you need to deserialize a DFA cheaply |
2379 | /// and cannot afford the transition table validation performed by |
2380 | /// `from_bytes`. |
2381 | /// |
2382 | /// # Example |
2383 | /// |
2384 | /// ``` |
2385 | /// use regex_automata::{dfa::{Automaton, dense::DFA}, HalfMatch, Input}; |
2386 | /// |
2387 | /// let initial = DFA::new("foo[0-9]+" )?; |
2388 | /// let (bytes, _) = initial.to_bytes_native_endian(); |
2389 | /// // SAFETY: This is guaranteed to be safe since the bytes given come |
2390 | /// // directly from a compatible serialization routine. |
2391 | /// let dfa: DFA<&[u32]> = unsafe { DFA::from_bytes_unchecked(&bytes)?.0 }; |
2392 | /// |
2393 | /// let expected = Some(HalfMatch::must(0, 8)); |
2394 | /// assert_eq!(expected, dfa.try_search_fwd(&Input::new("foo12345" ))?); |
2395 | /// # Ok::<(), Box<dyn std::error::Error>>(()) |
2396 | /// ``` |
2397 | pub unsafe fn from_bytes_unchecked( |
2398 | slice: &'a [u8], |
2399 | ) -> Result<(DFA<&'a [u32]>, usize), DeserializeError> { |
2400 | let mut nr = 0; |
2401 | |
2402 | nr += wire::skip_initial_padding(slice); |
2403 | wire::check_alignment::<StateID>(&slice[nr..])?; |
2404 | nr += wire::read_label(&slice[nr..], LABEL)?; |
2405 | nr += wire::read_endianness_check(&slice[nr..])?; |
2406 | nr += wire::read_version(&slice[nr..], VERSION)?; |
2407 | |
2408 | let _unused = wire::try_read_u32(&slice[nr..], "unused space" )?; |
2409 | nr += size_of::<u32>(); |
2410 | |
2411 | let (flags, nread) = Flags::from_bytes(&slice[nr..])?; |
2412 | nr += nread; |
2413 | |
2414 | let (tt, nread) = TransitionTable::from_bytes_unchecked(&slice[nr..])?; |
2415 | nr += nread; |
2416 | |
2417 | let (st, nread) = StartTable::from_bytes_unchecked(&slice[nr..])?; |
2418 | nr += nread; |
2419 | |
2420 | let (ms, nread) = MatchStates::from_bytes_unchecked(&slice[nr..])?; |
2421 | nr += nread; |
2422 | |
2423 | let (special, nread) = Special::from_bytes(&slice[nr..])?; |
2424 | nr += nread; |
2425 | special.validate_state_len(tt.len(), tt.stride2)?; |
2426 | |
2427 | let (accels, nread) = Accels::from_bytes_unchecked(&slice[nr..])?; |
2428 | nr += nread; |
2429 | |
2430 | let (quitset, nread) = ByteSet::from_bytes(&slice[nr..])?; |
2431 | nr += nread; |
2432 | |
2433 | // Prefilters don't support serialization, so they're always absent. |
2434 | let pre = None; |
2435 | Ok((DFA { tt, st, ms, special, accels, pre, quitset, flags }, nr)) |
2436 | } |
2437 | |
2438 | /// The implementation of the public `write_to` serialization methods, |
2439 | /// which is generic over endianness. |
2440 | /// |
2441 | /// This is defined only for &[u32] to reduce binary size/compilation time. |
2442 | fn write_to<E: Endian>( |
2443 | &self, |
2444 | mut dst: &mut [u8], |
2445 | ) -> Result<usize, SerializeError> { |
2446 | let nwrite = self.write_to_len(); |
2447 | if dst.len() < nwrite { |
2448 | return Err(SerializeError::buffer_too_small("dense DFA" )); |
2449 | } |
2450 | dst = &mut dst[..nwrite]; |
2451 | |
2452 | let mut nw = 0; |
2453 | nw += wire::write_label(LABEL, &mut dst[nw..])?; |
2454 | nw += wire::write_endianness_check::<E>(&mut dst[nw..])?; |
2455 | nw += wire::write_version::<E>(VERSION, &mut dst[nw..])?; |
2456 | nw += { |
2457 | // Currently unused, intended for future flexibility |
2458 | E::write_u32(0, &mut dst[nw..]); |
2459 | size_of::<u32>() |
2460 | }; |
2461 | nw += self.flags.write_to::<E>(&mut dst[nw..])?; |
2462 | nw += self.tt.write_to::<E>(&mut dst[nw..])?; |
2463 | nw += self.st.write_to::<E>(&mut dst[nw..])?; |
2464 | nw += self.ms.write_to::<E>(&mut dst[nw..])?; |
2465 | nw += self.special.write_to::<E>(&mut dst[nw..])?; |
2466 | nw += self.accels.write_to::<E>(&mut dst[nw..])?; |
2467 | nw += self.quitset.write_to::<E>(&mut dst[nw..])?; |
2468 | Ok(nw) |
2469 | } |
2470 | } |
2471 | |
2472 | // The following methods implement mutable routines on the internal |
2473 | // representation of a DFA. As such, we must fix the first type parameter to a |
2474 | // `Vec<u32>` since a generic `T: AsRef<[u32]>` does not permit mutation. We |
2475 | // can get away with this because these methods are internal to the crate and |
2476 | // are exclusively used during construction of the DFA. |
2477 | #[cfg (feature = "dfa-build" )] |
2478 | impl OwnedDFA { |
2479 | /// Add a start state of this DFA. |
2480 | pub(crate) fn set_start_state( |
2481 | &mut self, |
2482 | anchored: Anchored, |
2483 | start: Start, |
2484 | id: StateID, |
2485 | ) { |
2486 | assert!(self.tt.is_valid(id), "invalid start state" ); |
2487 | self.st.set_start(anchored, start, id); |
2488 | } |
2489 | |
2490 | /// Set the given transition to this DFA. Both the `from` and `to` states |
2491 | /// must already exist. |
2492 | pub(crate) fn set_transition( |
2493 | &mut self, |
2494 | from: StateID, |
2495 | byte: alphabet::Unit, |
2496 | to: StateID, |
2497 | ) { |
2498 | self.tt.set(from, byte, to); |
2499 | } |
2500 | |
2501 | /// An an empty state (a state where all transitions lead to a dead state) |
2502 | /// and return its identifier. The identifier returned is guaranteed to |
2503 | /// not point to any other existing state. |
2504 | /// |
2505 | /// If adding a state would exceed `StateID::LIMIT`, then this returns an |
2506 | /// error. |
2507 | pub(crate) fn add_empty_state(&mut self) -> Result<StateID, BuildError> { |
2508 | self.tt.add_empty_state() |
2509 | } |
2510 | |
2511 | /// Swap the two states given in the transition table. |
2512 | /// |
2513 | /// This routine does not do anything to check the correctness of this |
2514 | /// swap. Callers must ensure that other states pointing to id1 and id2 are |
2515 | /// updated appropriately. |
2516 | pub(crate) fn swap_states(&mut self, id1: StateID, id2: StateID) { |
2517 | self.tt.swap(id1, id2); |
2518 | } |
2519 | |
2520 | /// Remap all of the state identifiers in this DFA according to the map |
2521 | /// function given. This includes all transitions and all starting state |
2522 | /// identifiers. |
2523 | pub(crate) fn remap(&mut self, map: impl Fn(StateID) -> StateID) { |
2524 | // We could loop over each state ID and call 'remap_state' here, but |
2525 | // this is more direct: just map every transition directly. This |
2526 | // technically might do a little extra work since the alphabet length |
2527 | // is likely less than the stride, but if that is indeed an issue we |
2528 | // should benchmark it and fix it. |
2529 | for sid in self.tt.table_mut().iter_mut() { |
2530 | *sid = map(*sid); |
2531 | } |
2532 | for sid in self.st.table_mut().iter_mut() { |
2533 | *sid = map(*sid); |
2534 | } |
2535 | } |
2536 | |
2537 | /// Remap the transitions for the state given according to the function |
2538 | /// given. This applies the given map function to every transition in the |
2539 | /// given state and changes the transition in place to the result of the |
2540 | /// map function for that transition. |
2541 | pub(crate) fn remap_state( |
2542 | &mut self, |
2543 | id: StateID, |
2544 | map: impl Fn(StateID) -> StateID, |
2545 | ) { |
2546 | self.tt.remap(id, map); |
2547 | } |
2548 | |
2549 | /// Truncate the states in this DFA to the given length. |
2550 | /// |
2551 | /// This routine does not do anything to check the correctness of this |
2552 | /// truncation. Callers must ensure that other states pointing to truncated |
2553 | /// states are updated appropriately. |
2554 | pub(crate) fn truncate_states(&mut self, len: usize) { |
2555 | self.tt.truncate(len); |
2556 | } |
2557 | |
2558 | /// Minimize this DFA in place using Hopcroft's algorithm. |
2559 | pub(crate) fn minimize(&mut self) { |
2560 | Minimizer::new(self).run(); |
2561 | } |
2562 | |
2563 | /// Updates the match state pattern ID map to use the one provided. |
2564 | /// |
2565 | /// This is useful when it's convenient to manipulate matching states |
2566 | /// (and their corresponding pattern IDs) as a map. In particular, the |
2567 | /// representation used by a DFA for this map is not amenable to mutation, |
2568 | /// so if things need to be changed (like when shuffling states), it's |
2569 | /// often easier to work with the map form. |
2570 | pub(crate) fn set_pattern_map( |
2571 | &mut self, |
2572 | map: &BTreeMap<StateID, Vec<PatternID>>, |
2573 | ) -> Result<(), BuildError> { |
2574 | self.ms = self.ms.new_with_map(map)?; |
2575 | Ok(()) |
2576 | } |
2577 | |
2578 | /// Find states that have a small number of non-loop transitions and mark |
2579 | /// them as candidates for acceleration during search. |
2580 | pub(crate) fn accelerate(&mut self) { |
2581 | // dead and quit states can never be accelerated. |
2582 | if self.state_len() <= 2 { |
2583 | return; |
2584 | } |
2585 | |
2586 | // Go through every state and record their accelerator, if possible. |
2587 | let mut accels = BTreeMap::new(); |
2588 | // Count the number of accelerated match, start and non-match/start |
2589 | // states. |
2590 | let (mut cmatch, mut cstart, mut cnormal) = (0, 0, 0); |
2591 | for state in self.states() { |
2592 | if let Some(accel) = state.accelerate(self.byte_classes()) { |
2593 | debug!( |
2594 | "accelerating full DFA state {}: {:?}" , |
2595 | state.id().as_usize(), |
2596 | accel, |
2597 | ); |
2598 | accels.insert(state.id(), accel); |
2599 | if self.is_match_state(state.id()) { |
2600 | cmatch += 1; |
2601 | } else if self.is_start_state(state.id()) { |
2602 | cstart += 1; |
2603 | } else { |
2604 | assert!(!self.is_dead_state(state.id())); |
2605 | assert!(!self.is_quit_state(state.id())); |
2606 | cnormal += 1; |
2607 | } |
2608 | } |
2609 | } |
2610 | // If no states were able to be accelerated, then we're done. |
2611 | if accels.is_empty() { |
2612 | return; |
2613 | } |
2614 | let original_accels_len = accels.len(); |
2615 | |
2616 | // A remapper keeps track of state ID changes. Once we're done |
2617 | // shuffling, the remapper is used to rewrite all transitions in the |
2618 | // DFA based on the new positions of states. |
2619 | let mut remapper = Remapper::new(self); |
2620 | |
2621 | // As we swap states, if they are match states, we need to swap their |
2622 | // pattern ID lists too (for multi-regexes). We do this by converting |
2623 | // the lists to an easily swappable map, and then convert back to |
2624 | // MatchStates once we're done. |
2625 | let mut new_matches = self.ms.to_map(self); |
2626 | |
2627 | // There is at least one state that gets accelerated, so these are |
2628 | // guaranteed to get set to sensible values below. |
2629 | self.special.min_accel = StateID::MAX; |
2630 | self.special.max_accel = StateID::ZERO; |
2631 | let update_special_accel = |
2632 | |special: &mut Special, accel_id: StateID| { |
2633 | special.min_accel = cmp::min(special.min_accel, accel_id); |
2634 | special.max_accel = cmp::max(special.max_accel, accel_id); |
2635 | }; |
2636 | |
2637 | // Start by shuffling match states. Any match states that are |
2638 | // accelerated get moved to the end of the match state range. |
2639 | if cmatch > 0 && self.special.matches() { |
2640 | // N.B. special.{min,max}_match do not need updating, since the |
2641 | // range/number of match states does not change. Only the ordering |
2642 | // of match states may change. |
2643 | let mut next_id = self.special.max_match; |
2644 | let mut cur_id = next_id; |
2645 | while cur_id >= self.special.min_match { |
2646 | if let Some(accel) = accels.remove(&cur_id) { |
2647 | accels.insert(next_id, accel); |
2648 | update_special_accel(&mut self.special, next_id); |
2649 | |
2650 | // No need to do any actual swapping for equivalent IDs. |
2651 | if cur_id != next_id { |
2652 | remapper.swap(self, cur_id, next_id); |
2653 | |
2654 | // Swap pattern IDs for match states. |
2655 | let cur_pids = new_matches.remove(&cur_id).unwrap(); |
2656 | let next_pids = new_matches.remove(&next_id).unwrap(); |
2657 | new_matches.insert(cur_id, next_pids); |
2658 | new_matches.insert(next_id, cur_pids); |
2659 | } |
2660 | next_id = self.tt.prev_state_id(next_id); |
2661 | } |
2662 | cur_id = self.tt.prev_state_id(cur_id); |
2663 | } |
2664 | } |
2665 | |
2666 | // This is where it gets tricky. Without acceleration, start states |
2667 | // normally come right after match states. But we want accelerated |
2668 | // states to be a single contiguous range (to make it very fast |
2669 | // to determine whether a state *is* accelerated), while also keeping |
2670 | // match and starting states as contiguous ranges for the same reason. |
2671 | // So what we do here is shuffle states such that it looks like this: |
2672 | // |
2673 | // DQMMMMAAAAASSSSSSNNNNNNN |
2674 | // | | |
2675 | // |---------| |
2676 | // accelerated states |
2677 | // |
2678 | // Where: |
2679 | // D - dead state |
2680 | // Q - quit state |
2681 | // M - match state (may be accelerated) |
2682 | // A - normal state that is accelerated |
2683 | // S - start state (may be accelerated) |
2684 | // N - normal state that is NOT accelerated |
2685 | // |
2686 | // We implement this by shuffling states, which is done by a sequence |
2687 | // of pairwise swaps. We start by looking at all normal states to be |
2688 | // accelerated. When we find one, we swap it with the earliest starting |
2689 | // state, and then swap that with the earliest normal state. This |
2690 | // preserves the contiguous property. |
2691 | // |
2692 | // Once we're done looking for accelerated normal states, now we look |
2693 | // for accelerated starting states by moving them to the beginning |
2694 | // of the starting state range (just like we moved accelerated match |
2695 | // states to the end of the matching state range). |
2696 | // |
2697 | // For a more detailed/different perspective on this, see the docs |
2698 | // in dfa/special.rs. |
2699 | if cnormal > 0 { |
2700 | // our next available starting and normal states for swapping. |
2701 | let mut next_start_id = self.special.min_start; |
2702 | let mut cur_id = self.to_state_id(self.state_len() - 1); |
2703 | // This is guaranteed to exist since cnormal > 0. |
2704 | let mut next_norm_id = |
2705 | self.tt.next_state_id(self.special.max_start); |
2706 | while cur_id >= next_norm_id { |
2707 | if let Some(accel) = accels.remove(&cur_id) { |
2708 | remapper.swap(self, next_start_id, cur_id); |
2709 | remapper.swap(self, next_norm_id, cur_id); |
2710 | // Keep our accelerator map updated with new IDs if the |
2711 | // states we swapped were also accelerated. |
2712 | if let Some(accel2) = accels.remove(&next_norm_id) { |
2713 | accels.insert(cur_id, accel2); |
2714 | } |
2715 | if let Some(accel2) = accels.remove(&next_start_id) { |
2716 | accels.insert(next_norm_id, accel2); |
2717 | } |
2718 | accels.insert(next_start_id, accel); |
2719 | update_special_accel(&mut self.special, next_start_id); |
2720 | // Our start range shifts one to the right now. |
2721 | self.special.min_start = |
2722 | self.tt.next_state_id(self.special.min_start); |
2723 | self.special.max_start = |
2724 | self.tt.next_state_id(self.special.max_start); |
2725 | next_start_id = self.tt.next_state_id(next_start_id); |
2726 | next_norm_id = self.tt.next_state_id(next_norm_id); |
2727 | } |
2728 | // This is pretty tricky, but if our 'next_norm_id' state also |
2729 | // happened to be accelerated, then the result is that it is |
2730 | // now in the position of cur_id, so we need to consider it |
2731 | // again. This loop is still guaranteed to terminate though, |
2732 | // because when accels contains cur_id, we're guaranteed to |
2733 | // increment next_norm_id even if cur_id remains unchanged. |
2734 | if !accels.contains_key(&cur_id) { |
2735 | cur_id = self.tt.prev_state_id(cur_id); |
2736 | } |
2737 | } |
2738 | } |
2739 | // Just like we did for match states, but we want to move accelerated |
2740 | // start states to the beginning of the range instead of the end. |
2741 | if cstart > 0 { |
2742 | // N.B. special.{min,max}_start do not need updating, since the |
2743 | // range/number of start states does not change at this point. Only |
2744 | // the ordering of start states may change. |
2745 | let mut next_id = self.special.min_start; |
2746 | let mut cur_id = next_id; |
2747 | while cur_id <= self.special.max_start { |
2748 | if let Some(accel) = accels.remove(&cur_id) { |
2749 | remapper.swap(self, cur_id, next_id); |
2750 | accels.insert(next_id, accel); |
2751 | update_special_accel(&mut self.special, next_id); |
2752 | next_id = self.tt.next_state_id(next_id); |
2753 | } |
2754 | cur_id = self.tt.next_state_id(cur_id); |
2755 | } |
2756 | } |
2757 | |
2758 | // Remap all transitions in our DFA and assert some things. |
2759 | remapper.remap(self); |
2760 | // This unwrap is OK because acceleration never changes the number of |
2761 | // match states or patterns in those match states. Since acceleration |
2762 | // runs after the pattern map has been set at least once, we know that |
2763 | // our match states cannot error. |
2764 | self.set_pattern_map(&new_matches).unwrap(); |
2765 | self.special.set_max(); |
2766 | self.special.validate().expect("special state ranges should validate" ); |
2767 | self.special |
2768 | .validate_state_len(self.state_len(), self.stride2()) |
2769 | .expect( |
2770 | "special state ranges should be consistent with state length" , |
2771 | ); |
2772 | assert_eq!( |
2773 | self.special.accel_len(self.stride()), |
2774 | // We record the number of accelerated states initially detected |
2775 | // since the accels map is itself mutated in the process above. |
2776 | // If mutated incorrectly, its size may change, and thus can't be |
2777 | // trusted as a source of truth of how many accelerated states we |
2778 | // expected there to be. |
2779 | original_accels_len, |
2780 | "mismatch with expected number of accelerated states" , |
2781 | ); |
2782 | |
2783 | // And finally record our accelerators. We kept our accels map updated |
2784 | // as we shuffled states above, so the accelerators should now |
2785 | // correspond to a contiguous range in the state ID space. (Which we |
2786 | // assert.) |
2787 | let mut prev: Option<StateID> = None; |
2788 | for (id, accel) in accels { |
2789 | assert!(prev.map_or(true, |p| self.tt.next_state_id(p) == id)); |
2790 | prev = Some(id); |
2791 | self.accels.add(accel); |
2792 | } |
2793 | } |
2794 | |
2795 | /// Shuffle the states in this DFA so that starting states, match |
2796 | /// states and accelerated states are all contiguous. |
2797 | /// |
2798 | /// See dfa/special.rs for more details. |
2799 | pub(crate) fn shuffle( |
2800 | &mut self, |
2801 | mut matches: BTreeMap<StateID, Vec<PatternID>>, |
2802 | ) -> Result<(), BuildError> { |
2803 | // The determinizer always adds a quit state and it is always second. |
2804 | self.special.quit_id = self.to_state_id(1); |
2805 | // If all we have are the dead and quit states, then we're done and |
2806 | // the DFA will never produce a match. |
2807 | if self.state_len() <= 2 { |
2808 | self.special.set_max(); |
2809 | return Ok(()); |
2810 | } |
2811 | |
2812 | // Collect all our non-DEAD start states into a convenient set and |
2813 | // confirm there is no overlap with match states. In the classicl DFA |
2814 | // construction, start states can be match states. But because of |
2815 | // look-around, we delay all matches by a byte, which prevents start |
2816 | // states from being match states. |
2817 | let mut is_start: BTreeSet<StateID> = BTreeSet::new(); |
2818 | for (start_id, _, _) in self.starts() { |
2819 | // If a starting configuration points to a DEAD state, then we |
2820 | // don't want to shuffle it. The DEAD state is always the first |
2821 | // state with ID=0. So we can just leave it be. |
2822 | if start_id == DEAD { |
2823 | continue; |
2824 | } |
2825 | assert!( |
2826 | !matches.contains_key(&start_id), |
2827 | "{:?} is both a start and a match state, which is not allowed" , |
2828 | start_id, |
2829 | ); |
2830 | is_start.insert(start_id); |
2831 | } |
2832 | |
2833 | // We implement shuffling by a sequence of pairwise swaps of states. |
2834 | // Since we have a number of things referencing states via their |
2835 | // IDs and swapping them changes their IDs, we need to record every |
2836 | // swap we make so that we can remap IDs. The remapper handles this |
2837 | // book-keeping for us. |
2838 | let mut remapper = Remapper::new(self); |
2839 | |
2840 | // Shuffle matching states. |
2841 | if matches.is_empty() { |
2842 | self.special.min_match = DEAD; |
2843 | self.special.max_match = DEAD; |
2844 | } else { |
2845 | // The determinizer guarantees that the first two states are the |
2846 | // dead and quit states, respectively. We want our match states to |
2847 | // come right after quit. |
2848 | let mut next_id = self.to_state_id(2); |
2849 | let mut new_matches = BTreeMap::new(); |
2850 | self.special.min_match = next_id; |
2851 | for (id, pids) in matches { |
2852 | remapper.swap(self, next_id, id); |
2853 | new_matches.insert(next_id, pids); |
2854 | // If we swapped a start state, then update our set. |
2855 | if is_start.contains(&next_id) { |
2856 | is_start.remove(&next_id); |
2857 | is_start.insert(id); |
2858 | } |
2859 | next_id = self.tt.next_state_id(next_id); |
2860 | } |
2861 | matches = new_matches; |
2862 | self.special.max_match = cmp::max( |
2863 | self.special.min_match, |
2864 | self.tt.prev_state_id(next_id), |
2865 | ); |
2866 | } |
2867 | |
2868 | // Shuffle starting states. |
2869 | { |
2870 | let mut next_id = self.to_state_id(2); |
2871 | if self.special.matches() { |
2872 | next_id = self.tt.next_state_id(self.special.max_match); |
2873 | } |
2874 | self.special.min_start = next_id; |
2875 | for id in is_start { |
2876 | remapper.swap(self, next_id, id); |
2877 | next_id = self.tt.next_state_id(next_id); |
2878 | } |
2879 | self.special.max_start = cmp::max( |
2880 | self.special.min_start, |
2881 | self.tt.prev_state_id(next_id), |
2882 | ); |
2883 | } |
2884 | |
2885 | // Finally remap all transitions in our DFA. |
2886 | remapper.remap(self); |
2887 | self.set_pattern_map(&matches)?; |
2888 | self.special.set_max(); |
2889 | self.special.validate().expect("special state ranges should validate" ); |
2890 | self.special |
2891 | .validate_state_len(self.state_len(), self.stride2()) |
2892 | .expect( |
2893 | "special state ranges should be consistent with state length" , |
2894 | ); |
2895 | Ok(()) |
2896 | } |
2897 | |
2898 | /// Checks whether there are universal start states (both anchored and |
2899 | /// unanchored), and if so, sets the relevant fields to the start state |
2900 | /// IDs. |
2901 | /// |
2902 | /// Universal start states occur precisely when the all patterns in the |
2903 | /// DFA have no look-around assertions in their prefix. |
2904 | fn set_universal_starts(&mut self) { |
2905 | assert_eq!(6, Start::len(), "expected 6 start configurations" ); |
2906 | |
2907 | let start_id = |dfa: &mut OwnedDFA, |
2908 | anchored: Anchored, |
2909 | start: Start| { |
2910 | // This OK because we only call 'start' under conditions |
2911 | // in which we know it will succeed. |
2912 | dfa.st.start(anchored, start).expect("valid Input configuration" ) |
2913 | }; |
2914 | if self.start_kind().has_unanchored() { |
2915 | let anchor = Anchored::No; |
2916 | let sid = start_id(self, anchor, Start::NonWordByte); |
2917 | if sid == start_id(self, anchor, Start::WordByte) |
2918 | && sid == start_id(self, anchor, Start::Text) |
2919 | && sid == start_id(self, anchor, Start::LineLF) |
2920 | && sid == start_id(self, anchor, Start::LineCR) |
2921 | && sid == start_id(self, anchor, Start::CustomLineTerminator) |
2922 | { |
2923 | self.st.universal_start_unanchored = Some(sid); |
2924 | } |
2925 | } |
2926 | if self.start_kind().has_anchored() { |
2927 | let anchor = Anchored::Yes; |
2928 | let sid = start_id(self, anchor, Start::NonWordByte); |
2929 | if sid == start_id(self, anchor, Start::WordByte) |
2930 | && sid == start_id(self, anchor, Start::Text) |
2931 | && sid == start_id(self, anchor, Start::LineLF) |
2932 | && sid == start_id(self, anchor, Start::LineCR) |
2933 | && sid == start_id(self, anchor, Start::CustomLineTerminator) |
2934 | { |
2935 | self.st.universal_start_anchored = Some(sid); |
2936 | } |
2937 | } |
2938 | } |
2939 | } |
2940 | |
2941 | // A variety of generic internal methods for accessing DFA internals. |
2942 | impl<T: AsRef<[u32]>> DFA<T> { |
2943 | /// Return the info about special states. |
2944 | pub(crate) fn special(&self) -> &Special { |
2945 | &self.special |
2946 | } |
2947 | |
2948 | /// Return the info about special states as a mutable borrow. |
2949 | #[cfg (feature = "dfa-build" )] |
2950 | pub(crate) fn special_mut(&mut self) -> &mut Special { |
2951 | &mut self.special |
2952 | } |
2953 | |
2954 | /// Returns the quit set (may be empty) used by this DFA. |
2955 | pub(crate) fn quitset(&self) -> &ByteSet { |
2956 | &self.quitset |
2957 | } |
2958 | |
2959 | /// Returns the flags for this DFA. |
2960 | pub(crate) fn flags(&self) -> &Flags { |
2961 | &self.flags |
2962 | } |
2963 | |
2964 | /// Returns an iterator over all states in this DFA. |
2965 | /// |
2966 | /// This iterator yields a tuple for each state. The first element of the |
2967 | /// tuple corresponds to a state's identifier, and the second element |
2968 | /// corresponds to the state itself (comprised of its transitions). |
2969 | pub(crate) fn states(&self) -> StateIter<'_, T> { |
2970 | self.tt.states() |
2971 | } |
2972 | |
2973 | /// Return the total number of states in this DFA. Every DFA has at least |
2974 | /// 1 state, even the empty DFA. |
2975 | pub(crate) fn state_len(&self) -> usize { |
2976 | self.tt.len() |
2977 | } |
2978 | |
2979 | /// Return an iterator over all pattern IDs for the given match state. |
2980 | /// |
2981 | /// If the given state is not a match state, then this panics. |
2982 | #[cfg (feature = "dfa-build" )] |
2983 | pub(crate) fn pattern_id_slice(&self, id: StateID) -> &[PatternID] { |
2984 | assert!(self.is_match_state(id)); |
2985 | self.ms.pattern_id_slice(self.match_state_index(id)) |
2986 | } |
2987 | |
2988 | /// Return the total number of pattern IDs for the given match state. |
2989 | /// |
2990 | /// If the given state is not a match state, then this panics. |
2991 | pub(crate) fn match_pattern_len(&self, id: StateID) -> usize { |
2992 | assert!(self.is_match_state(id)); |
2993 | self.ms.pattern_len(self.match_state_index(id)) |
2994 | } |
2995 | |
2996 | /// Returns the total number of patterns matched by this DFA. |
2997 | pub(crate) fn pattern_len(&self) -> usize { |
2998 | self.ms.pattern_len |
2999 | } |
3000 | |
3001 | /// Returns a map from match state ID to a list of pattern IDs that match |
3002 | /// in that state. |
3003 | #[cfg (feature = "dfa-build" )] |
3004 | pub(crate) fn pattern_map(&self) -> BTreeMap<StateID, Vec<PatternID>> { |
3005 | self.ms.to_map(self) |
3006 | } |
3007 | |
3008 | /// Returns the ID of the quit state for this DFA. |
3009 | #[cfg (feature = "dfa-build" )] |
3010 | pub(crate) fn quit_id(&self) -> StateID { |
3011 | self.to_state_id(1) |
3012 | } |
3013 | |
3014 | /// Convert the given state identifier to the state's index. The state's |
3015 | /// index corresponds to the position in which it appears in the transition |
3016 | /// table. When a DFA is NOT premultiplied, then a state's identifier is |
3017 | /// also its index. When a DFA is premultiplied, then a state's identifier |
3018 | /// is equal to `index * alphabet_len`. This routine reverses that. |
3019 | pub(crate) fn to_index(&self, id: StateID) -> usize { |
3020 | self.tt.to_index(id) |
3021 | } |
3022 | |
3023 | /// Convert an index to a state (in the range 0..self.state_len()) to an |
3024 | /// actual state identifier. |
3025 | /// |
3026 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state |
3027 | /// to some other information (such as a remapped state ID). |
3028 | #[cfg (feature = "dfa-build" )] |
3029 | pub(crate) fn to_state_id(&self, index: usize) -> StateID { |
3030 | self.tt.to_state_id(index) |
3031 | } |
3032 | |
3033 | /// Return the table of state IDs for this DFA's start states. |
3034 | pub(crate) fn starts(&self) -> StartStateIter<'_> { |
3035 | self.st.iter() |
3036 | } |
3037 | |
3038 | /// Returns the index of the match state for the given ID. If the |
3039 | /// given ID does not correspond to a match state, then this may |
3040 | /// panic or produce an incorrect result. |
3041 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3042 | fn match_state_index(&self, id: StateID) -> usize { |
3043 | debug_assert!(self.is_match_state(id)); |
3044 | // This is one of the places where we rely on the fact that match |
3045 | // states are contiguous in the transition table. Namely, that the |
3046 | // first match state ID always corresponds to dfa.special.min_match. |
3047 | // From there, since we know the stride, we can compute the overall |
3048 | // index of any match state given the match state's ID. |
3049 | let min = self.special().min_match.as_usize(); |
3050 | // CORRECTNESS: We're allowed to produce an incorrect result or panic, |
3051 | // so both the subtraction and the unchecked StateID construction is |
3052 | // OK. |
3053 | self.to_index(StateID::new_unchecked(id.as_usize() - min)) |
3054 | } |
3055 | |
3056 | /// Returns the index of the accelerator state for the given ID. If the |
3057 | /// given ID does not correspond to an accelerator state, then this may |
3058 | /// panic or produce an incorrect result. |
3059 | fn accelerator_index(&self, id: StateID) -> usize { |
3060 | let min = self.special().min_accel.as_usize(); |
3061 | // CORRECTNESS: We're allowed to produce an incorrect result or panic, |
3062 | // so both the subtraction and the unchecked StateID construction is |
3063 | // OK. |
3064 | self.to_index(StateID::new_unchecked(id.as_usize() - min)) |
3065 | } |
3066 | |
3067 | /// Return the accelerators for this DFA. |
3068 | fn accels(&self) -> Accels<&[u32]> { |
3069 | self.accels.as_ref() |
3070 | } |
3071 | |
3072 | /// Return this DFA's transition table as a slice. |
3073 | fn trans(&self) -> &[StateID] { |
3074 | self.tt.table() |
3075 | } |
3076 | } |
3077 | |
3078 | impl<T: AsRef<[u32]>> fmt::Debug for DFA<T> { |
3079 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
3080 | writeln!(f, "dense::DFA(" )?; |
3081 | for state in self.states() { |
3082 | fmt_state_indicator(f, self, state.id())?; |
3083 | let id = if f.alternate() { |
3084 | state.id().as_usize() |
3085 | } else { |
3086 | self.to_index(state.id()) |
3087 | }; |
3088 | write!(f, "{:06?}: " , id)?; |
3089 | state.fmt(f)?; |
3090 | write!(f, " \n" )?; |
3091 | } |
3092 | writeln!(f, "" )?; |
3093 | for (i, (start_id, anchored, sty)) in self.starts().enumerate() { |
3094 | let id = if f.alternate() { |
3095 | start_id.as_usize() |
3096 | } else { |
3097 | self.to_index(start_id) |
3098 | }; |
3099 | if i % self.st.stride == 0 { |
3100 | match anchored { |
3101 | Anchored::No => writeln!(f, "START-GROUP(unanchored)" )?, |
3102 | Anchored::Yes => writeln!(f, "START-GROUP(anchored)" )?, |
3103 | Anchored::Pattern(pid) => { |
3104 | writeln!(f, "START_GROUP(pattern: {:?})" , pid)? |
3105 | } |
3106 | } |
3107 | } |
3108 | writeln!(f, " {:?} => {:06?}" , sty, id)?; |
3109 | } |
3110 | if self.pattern_len() > 1 { |
3111 | writeln!(f, "" )?; |
3112 | for i in 0..self.ms.len() { |
3113 | let id = self.ms.match_state_id(self, i); |
3114 | let id = if f.alternate() { |
3115 | id.as_usize() |
3116 | } else { |
3117 | self.to_index(id) |
3118 | }; |
3119 | write!(f, "MATCH({:06?}): " , id)?; |
3120 | for (i, &pid) in self.ms.pattern_id_slice(i).iter().enumerate() |
3121 | { |
3122 | if i > 0 { |
3123 | write!(f, ", " )?; |
3124 | } |
3125 | write!(f, "{:?}" , pid)?; |
3126 | } |
3127 | writeln!(f, "" )?; |
3128 | } |
3129 | } |
3130 | writeln!(f, "state length: {:?}" , self.state_len())?; |
3131 | writeln!(f, "pattern length: {:?}" , self.pattern_len())?; |
3132 | writeln!(f, "flags: {:?}" , self.flags)?; |
3133 | writeln!(f, ")" )?; |
3134 | Ok(()) |
3135 | } |
3136 | } |
3137 | |
3138 | // SAFETY: We assert that our implementation of each method is correct. |
3139 | unsafe impl<T: AsRef<[u32]>> Automaton for DFA<T> { |
3140 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3141 | fn is_special_state(&self, id: StateID) -> bool { |
3142 | self.special.is_special_state(id) |
3143 | } |
3144 | |
3145 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3146 | fn is_dead_state(&self, id: StateID) -> bool { |
3147 | self.special.is_dead_state(id) |
3148 | } |
3149 | |
3150 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3151 | fn is_quit_state(&self, id: StateID) -> bool { |
3152 | self.special.is_quit_state(id) |
3153 | } |
3154 | |
3155 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3156 | fn is_match_state(&self, id: StateID) -> bool { |
3157 | self.special.is_match_state(id) |
3158 | } |
3159 | |
3160 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3161 | fn is_start_state(&self, id: StateID) -> bool { |
3162 | self.special.is_start_state(id) |
3163 | } |
3164 | |
3165 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3166 | fn is_accel_state(&self, id: StateID) -> bool { |
3167 | self.special.is_accel_state(id) |
3168 | } |
3169 | |
3170 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3171 | fn next_state(&self, current: StateID, input: u8) -> StateID { |
3172 | let input = self.byte_classes().get(input); |
3173 | let o = current.as_usize() + usize::from(input); |
3174 | self.trans()[o] |
3175 | } |
3176 | |
3177 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3178 | unsafe fn next_state_unchecked( |
3179 | &self, |
3180 | current: StateID, |
3181 | byte: u8, |
3182 | ) -> StateID { |
3183 | // We don't (or shouldn't) need an unchecked variant for the byte |
3184 | // class mapping, since bound checks should be omitted automatically |
3185 | // by virtue of its representation. If this ends up not being true as |
3186 | // confirmed by codegen, please file an issue. ---AG |
3187 | let class = self.byte_classes().get(byte); |
3188 | let o = current.as_usize() + usize::from(class); |
3189 | let next = *self.trans().get_unchecked(o); |
3190 | next |
3191 | } |
3192 | |
3193 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3194 | fn next_eoi_state(&self, current: StateID) -> StateID { |
3195 | let eoi = self.byte_classes().eoi().as_usize(); |
3196 | let o = current.as_usize() + eoi; |
3197 | self.trans()[o] |
3198 | } |
3199 | |
3200 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3201 | fn pattern_len(&self) -> usize { |
3202 | self.ms.pattern_len |
3203 | } |
3204 | |
3205 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3206 | fn match_len(&self, id: StateID) -> usize { |
3207 | self.match_pattern_len(id) |
3208 | } |
3209 | |
3210 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3211 | fn match_pattern(&self, id: StateID, match_index: usize) -> PatternID { |
3212 | // This is an optimization for the very common case of a DFA with a |
3213 | // single pattern. This conditional avoids a somewhat more costly path |
3214 | // that finds the pattern ID from the state machine, which requires |
3215 | // a bit of slicing/pointer-chasing. This optimization tends to only |
3216 | // matter when matches are frequent. |
3217 | if self.ms.pattern_len == 1 { |
3218 | return PatternID::ZERO; |
3219 | } |
3220 | let state_index = self.match_state_index(id); |
3221 | self.ms.pattern_id(state_index, match_index) |
3222 | } |
3223 | |
3224 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3225 | fn has_empty(&self) -> bool { |
3226 | self.flags.has_empty |
3227 | } |
3228 | |
3229 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3230 | fn is_utf8(&self) -> bool { |
3231 | self.flags.is_utf8 |
3232 | } |
3233 | |
3234 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3235 | fn is_always_start_anchored(&self) -> bool { |
3236 | self.flags.is_always_start_anchored |
3237 | } |
3238 | |
3239 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3240 | fn start_state( |
3241 | &self, |
3242 | config: &start::Config, |
3243 | ) -> Result<StateID, StartError> { |
3244 | let anchored = config.get_anchored(); |
3245 | let start = match config.get_look_behind() { |
3246 | None => Start::Text, |
3247 | Some(byte) => { |
3248 | if !self.quitset.is_empty() && self.quitset.contains(byte) { |
3249 | return Err(StartError::quit(byte)); |
3250 | } |
3251 | self.st.start_map.get(byte) |
3252 | } |
3253 | }; |
3254 | self.st.start(anchored, start) |
3255 | } |
3256 | |
3257 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3258 | fn universal_start_state(&self, mode: Anchored) -> Option<StateID> { |
3259 | match mode { |
3260 | Anchored::No => self.st.universal_start_unanchored, |
3261 | Anchored::Yes => self.st.universal_start_anchored, |
3262 | Anchored::Pattern(_) => None, |
3263 | } |
3264 | } |
3265 | |
3266 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3267 | fn accelerator(&self, id: StateID) -> &[u8] { |
3268 | if !self.is_accel_state(id) { |
3269 | return &[]; |
3270 | } |
3271 | self.accels.needles(self.accelerator_index(id)) |
3272 | } |
3273 | |
3274 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
3275 | fn get_prefilter(&self) -> Option<&Prefilter> { |
3276 | self.pre.as_ref() |
3277 | } |
3278 | } |
3279 | |
3280 | /// The transition table portion of a dense DFA. |
3281 | /// |
3282 | /// The transition table is the core part of the DFA in that it describes how |
3283 | /// to move from one state to another based on the input sequence observed. |
3284 | #[derive(Clone)] |
3285 | pub(crate) struct TransitionTable<T> { |
3286 | /// A contiguous region of memory representing the transition table in |
3287 | /// row-major order. The representation is dense. That is, every state |
3288 | /// has precisely the same number of transitions. The maximum number of |
3289 | /// transitions per state is 257 (256 for each possible byte value, plus 1 |
3290 | /// for the special EOI transition). If a DFA has been instructed to use |
3291 | /// byte classes (the default), then the number of transitions is usually |
3292 | /// substantially fewer. |
3293 | /// |
3294 | /// In practice, T is either `Vec<u32>` or `&[u32]`. |
3295 | table: T, |
3296 | /// A set of equivalence classes, where a single equivalence class |
3297 | /// represents a set of bytes that never discriminate between a match |
3298 | /// and a non-match in the DFA. Each equivalence class corresponds to a |
3299 | /// single character in this DFA's alphabet, where the maximum number of |
3300 | /// characters is 257 (each possible value of a byte plus the special |
3301 | /// EOI transition). Consequently, the number of equivalence classes |
3302 | /// corresponds to the number of transitions for each DFA state. Note |
3303 | /// though that the *space* used by each DFA state in the transition table |
3304 | /// may be larger. The total space used by each DFA state is known as the |
3305 | /// stride. |
3306 | /// |
3307 | /// The only time the number of equivalence classes is fewer than 257 is if |
3308 | /// the DFA's kind uses byte classes (which is the default). Equivalence |
3309 | /// classes should generally only be disabled when debugging, so that |
3310 | /// the transitions themselves aren't obscured. Disabling them has no |
3311 | /// other benefit, since the equivalence class map is always used while |
3312 | /// searching. In the vast majority of cases, the number of equivalence |
3313 | /// classes is substantially smaller than 257, particularly when large |
3314 | /// Unicode classes aren't used. |
3315 | classes: ByteClasses, |
3316 | /// The stride of each DFA state, expressed as a power-of-two exponent. |
3317 | /// |
3318 | /// The stride of a DFA corresponds to the total amount of space used by |
3319 | /// each DFA state in the transition table. This may be bigger than the |
3320 | /// size of a DFA's alphabet, since the stride is always the smallest |
3321 | /// power of two greater than or equal to the alphabet size. |
3322 | /// |
3323 | /// While this wastes space, this avoids the need for integer division |
3324 | /// to convert between premultiplied state IDs and their corresponding |
3325 | /// indices. Instead, we can use simple bit-shifts. |
3326 | /// |
3327 | /// See the docs for the `stride2` method for more details. |
3328 | /// |
3329 | /// The minimum `stride2` value is `1` (corresponding to a stride of `2`) |
3330 | /// while the maximum `stride2` value is `9` (corresponding to a stride of |
3331 | /// `512`). The maximum is not `8` since the maximum alphabet size is `257` |
3332 | /// when accounting for the special EOI transition. However, an alphabet |
3333 | /// length of that size is exceptionally rare since the alphabet is shrunk |
3334 | /// into equivalence classes. |
3335 | stride2: usize, |
3336 | } |
3337 | |
3338 | impl<'a> TransitionTable<&'a [u32]> { |
3339 | /// Deserialize a transition table starting at the beginning of `slice`. |
3340 | /// Upon success, return the total number of bytes read along with the |
3341 | /// transition table. |
3342 | /// |
3343 | /// If there was a problem deserializing any part of the transition table, |
3344 | /// then this returns an error. Notably, if the given slice does not have |
3345 | /// the same alignment as `StateID`, then this will return an error (among |
3346 | /// other possible errors). |
3347 | /// |
3348 | /// This is guaranteed to execute in constant time. |
3349 | /// |
3350 | /// # Safety |
3351 | /// |
3352 | /// This routine is not safe because it does not check the validity of the |
3353 | /// transition table itself. In particular, the transition table can be |
3354 | /// quite large, so checking its validity can be somewhat expensive. An |
3355 | /// invalid transition table is not safe because other code may rely on the |
3356 | /// transition table being correct (such as explicit bounds check elision). |
3357 | /// Therefore, an invalid transition table can lead to undefined behavior. |
3358 | /// |
3359 | /// Callers that use this function must either pass on the safety invariant |
3360 | /// or guarantee that the bytes given contain a valid transition table. |
3361 | /// This guarantee is upheld by the bytes written by `write_to`. |
3362 | unsafe fn from_bytes_unchecked( |
3363 | mut slice: &'a [u8], |
3364 | ) -> Result<(TransitionTable<&'a [u32]>, usize), DeserializeError> { |
3365 | let slice_start = slice.as_ptr().as_usize(); |
3366 | |
3367 | let (state_len, nr) = |
3368 | wire::try_read_u32_as_usize(slice, "state length" )?; |
3369 | slice = &slice[nr..]; |
3370 | |
3371 | let (stride2, nr) = wire::try_read_u32_as_usize(slice, "stride2" )?; |
3372 | slice = &slice[nr..]; |
3373 | |
3374 | let (classes, nr) = ByteClasses::from_bytes(slice)?; |
3375 | slice = &slice[nr..]; |
3376 | |
3377 | // The alphabet length (determined by the byte class map) cannot be |
3378 | // bigger than the stride (total space used by each DFA state). |
3379 | if stride2 > 9 { |
3380 | return Err(DeserializeError::generic( |
3381 | "dense DFA has invalid stride2 (too big)" , |
3382 | )); |
3383 | } |
3384 | // It also cannot be zero, since even a DFA that never matches anything |
3385 | // has a non-zero number of states with at least two equivalence |
3386 | // classes: one for all 256 byte values and another for the EOI |
3387 | // sentinel. |
3388 | if stride2 < 1 { |
3389 | return Err(DeserializeError::generic( |
3390 | "dense DFA has invalid stride2 (too small)" , |
3391 | )); |
3392 | } |
3393 | // This is OK since 1 <= stride2 <= 9. |
3394 | let stride = |
3395 | 1usize.checked_shl(u32::try_from(stride2).unwrap()).unwrap(); |
3396 | if classes.alphabet_len() > stride { |
3397 | return Err(DeserializeError::generic( |
3398 | "alphabet size cannot be bigger than transition table stride" , |
3399 | )); |
3400 | } |
3401 | |
3402 | let trans_len = |
3403 | wire::shl(state_len, stride2, "dense table transition length" )?; |
3404 | let table_bytes_len = wire::mul( |
3405 | trans_len, |
3406 | StateID::SIZE, |
3407 | "dense table state byte length" , |
3408 | )?; |
3409 | wire::check_slice_len(slice, table_bytes_len, "transition table" )?; |
3410 | wire::check_alignment::<StateID>(slice)?; |
3411 | let table_bytes = &slice[..table_bytes_len]; |
3412 | slice = &slice[table_bytes_len..]; |
3413 | // SAFETY: Since StateID is always representable as a u32, all we need |
3414 | // to do is ensure that we have the proper length and alignment. We've |
3415 | // checked both above, so the cast below is safe. |
3416 | // |
3417 | // N.B. This is the only not-safe code in this function. |
3418 | let table = core::slice::from_raw_parts( |
3419 | table_bytes.as_ptr().cast::<u32>(), |
3420 | trans_len, |
3421 | ); |
3422 | let tt = TransitionTable { table, classes, stride2 }; |
3423 | Ok((tt, slice.as_ptr().as_usize() - slice_start)) |
3424 | } |
3425 | } |
3426 | |
3427 | #[cfg (feature = "dfa-build" )] |
3428 | impl TransitionTable<Vec<u32>> { |
3429 | /// Create a minimal transition table with just two states: a dead state |
3430 | /// and a quit state. The alphabet length and stride of the transition |
3431 | /// table is determined by the given set of equivalence classes. |
3432 | fn minimal(classes: ByteClasses) -> TransitionTable<Vec<u32>> { |
3433 | let mut tt = TransitionTable { |
3434 | table: vec![], |
3435 | classes, |
3436 | stride2: classes.stride2(), |
3437 | }; |
3438 | // Two states, regardless of alphabet size, can always fit into u32. |
3439 | tt.add_empty_state().unwrap(); // dead state |
3440 | tt.add_empty_state().unwrap(); // quit state |
3441 | tt |
3442 | } |
3443 | |
3444 | /// Set a transition in this table. Both the `from` and `to` states must |
3445 | /// already exist, otherwise this panics. `unit` should correspond to the |
3446 | /// transition out of `from` to set to `to`. |
3447 | fn set(&mut self, from: StateID, unit: alphabet::Unit, to: StateID) { |
3448 | assert!(self.is_valid(from), "invalid 'from' state" ); |
3449 | assert!(self.is_valid(to), "invalid 'to' state" ); |
3450 | self.table[from.as_usize() + self.classes.get_by_unit(unit)] = |
3451 | to.as_u32(); |
3452 | } |
3453 | |
3454 | /// Add an empty state (a state where all transitions lead to a dead state) |
3455 | /// and return its identifier. The identifier returned is guaranteed to |
3456 | /// not point to any other existing state. |
3457 | /// |
3458 | /// If adding a state would exhaust the state identifier space, then this |
3459 | /// returns an error. |
3460 | fn add_empty_state(&mut self) -> Result<StateID, BuildError> { |
3461 | // Normally, to get a fresh state identifier, we would just |
3462 | // take the index of the next state added to the transition |
3463 | // table. However, we actually perform an optimization here |
3464 | // that premultiplies state IDs by the stride, such that they |
3465 | // point immediately at the beginning of their transitions in |
3466 | // the transition table. This avoids an extra multiplication |
3467 | // instruction for state lookup at search time. |
3468 | // |
3469 | // Premultiplied identifiers means that instead of your matching |
3470 | // loop looking something like this: |
3471 | // |
3472 | // state = dfa.start |
3473 | // for byte in haystack: |
3474 | // next = dfa.transitions[state * stride + byte] |
3475 | // if dfa.is_match(next): |
3476 | // return true |
3477 | // return false |
3478 | // |
3479 | // it can instead look like this: |
3480 | // |
3481 | // state = dfa.start |
3482 | // for byte in haystack: |
3483 | // next = dfa.transitions[state + byte] |
3484 | // if dfa.is_match(next): |
3485 | // return true |
3486 | // return false |
3487 | // |
3488 | // In other words, we save a multiplication instruction in the |
3489 | // critical path. This turns out to be a decent performance win. |
3490 | // The cost of using premultiplied state ids is that they can |
3491 | // require a bigger state id representation. (And they also make |
3492 | // the code a bit more complex, especially during minimization and |
3493 | // when reshuffling states, as one needs to convert back and forth |
3494 | // between state IDs and state indices.) |
3495 | // |
3496 | // To do this, we simply take the index of the state into the |
3497 | // entire transition table, rather than the index of the state |
3498 | // itself. e.g., If the stride is 64, then the ID of the 3rd state |
3499 | // is 192, not 2. |
3500 | let next = self.table.len(); |
3501 | let id = |
3502 | StateID::new(next).map_err(|_| BuildError::too_many_states())?; |
3503 | self.table.extend(iter::repeat(0).take(self.stride())); |
3504 | Ok(id) |
3505 | } |
3506 | |
3507 | /// Swap the two states given in this transition table. |
3508 | /// |
3509 | /// This routine does not do anything to check the correctness of this |
3510 | /// swap. Callers must ensure that other states pointing to id1 and id2 are |
3511 | /// updated appropriately. |
3512 | /// |
3513 | /// Both id1 and id2 must point to valid states, otherwise this panics. |
3514 | fn swap(&mut self, id1: StateID, id2: StateID) { |
3515 | assert!(self.is_valid(id1), "invalid 'id1' state: {:?}" , id1); |
3516 | assert!(self.is_valid(id2), "invalid 'id2' state: {:?}" , id2); |
3517 | // We only need to swap the parts of the state that are used. So if the |
3518 | // stride is 64, but the alphabet length is only 33, then we save a lot |
3519 | // of work. |
3520 | for b in 0..self.classes.alphabet_len() { |
3521 | self.table.swap(id1.as_usize() + b, id2.as_usize() + b); |
3522 | } |
3523 | } |
3524 | |
3525 | /// Remap the transitions for the state given according to the function |
3526 | /// given. This applies the given map function to every transition in the |
3527 | /// given state and changes the transition in place to the result of the |
3528 | /// map function for that transition. |
3529 | fn remap(&mut self, id: StateID, map: impl Fn(StateID) -> StateID) { |
3530 | for byte in 0..self.alphabet_len() { |
3531 | let i = id.as_usize() + byte; |
3532 | let next = self.table()[i]; |
3533 | self.table_mut()[id.as_usize() + byte] = map(next); |
3534 | } |
3535 | } |
3536 | |
3537 | /// Truncate the states in this transition table to the given length. |
3538 | /// |
3539 | /// This routine does not do anything to check the correctness of this |
3540 | /// truncation. Callers must ensure that other states pointing to truncated |
3541 | /// states are updated appropriately. |
3542 | fn truncate(&mut self, len: usize) { |
3543 | self.table.truncate(len << self.stride2); |
3544 | } |
3545 | } |
3546 | |
3547 | impl<T: AsRef<[u32]>> TransitionTable<T> { |
3548 | /// Writes a serialized form of this transition table to the buffer given. |
3549 | /// If the buffer is too small, then an error is returned. To determine |
3550 | /// how big the buffer must be, use `write_to_len`. |
3551 | fn write_to<E: Endian>( |
3552 | &self, |
3553 | mut dst: &mut [u8], |
3554 | ) -> Result<usize, SerializeError> { |
3555 | let nwrite = self.write_to_len(); |
3556 | if dst.len() < nwrite { |
3557 | return Err(SerializeError::buffer_too_small("transition table" )); |
3558 | } |
3559 | dst = &mut dst[..nwrite]; |
3560 | |
3561 | // write state length |
3562 | // Unwrap is OK since number of states is guaranteed to fit in a u32. |
3563 | E::write_u32(u32::try_from(self.len()).unwrap(), dst); |
3564 | dst = &mut dst[size_of::<u32>()..]; |
3565 | |
3566 | // write state stride (as power of 2) |
3567 | // Unwrap is OK since stride2 is guaranteed to be <= 9. |
3568 | E::write_u32(u32::try_from(self.stride2).unwrap(), dst); |
3569 | dst = &mut dst[size_of::<u32>()..]; |
3570 | |
3571 | // write byte class map |
3572 | let n = self.classes.write_to(dst)?; |
3573 | dst = &mut dst[n..]; |
3574 | |
3575 | // write actual transitions |
3576 | for &sid in self.table() { |
3577 | let n = wire::write_state_id::<E>(sid, &mut dst); |
3578 | dst = &mut dst[n..]; |
3579 | } |
3580 | Ok(nwrite) |
3581 | } |
3582 | |
3583 | /// Returns the number of bytes the serialized form of this transition |
3584 | /// table will use. |
3585 | fn write_to_len(&self) -> usize { |
3586 | size_of::<u32>() // state length |
3587 | + size_of::<u32>() // stride2 |
3588 | + self.classes.write_to_len() |
3589 | + (self.table().len() * StateID::SIZE) |
3590 | } |
3591 | |
3592 | /// Validates that every state ID in this transition table is valid. |
3593 | /// |
3594 | /// That is, every state ID can be used to correctly index a state in this |
3595 | /// table. |
3596 | fn validate(&self, sp: &Special) -> Result<(), DeserializeError> { |
3597 | for state in self.states() { |
3598 | // We check that the ID itself is well formed. That is, if it's |
3599 | // a special state then it must actually be a quit, dead, accel, |
3600 | // match or start state. |
3601 | if sp.is_special_state(state.id()) { |
3602 | let is_actually_special = sp.is_dead_state(state.id()) |
3603 | || sp.is_quit_state(state.id()) |
3604 | || sp.is_match_state(state.id()) |
3605 | || sp.is_start_state(state.id()) |
3606 | || sp.is_accel_state(state.id()); |
3607 | if !is_actually_special { |
3608 | // This is kind of a cryptic error message... |
3609 | return Err(DeserializeError::generic( |
3610 | "found dense state tagged as special but \ |
3611 | wasn't actually special" , |
3612 | )); |
3613 | } |
3614 | } |
3615 | for (_, to) in state.transitions() { |
3616 | if !self.is_valid(to) { |
3617 | return Err(DeserializeError::generic( |
3618 | "found invalid state ID in transition table" , |
3619 | )); |
3620 | } |
3621 | } |
3622 | } |
3623 | Ok(()) |
3624 | } |
3625 | |
3626 | /// Converts this transition table to a borrowed value. |
3627 | fn as_ref(&self) -> TransitionTable<&'_ [u32]> { |
3628 | TransitionTable { |
3629 | table: self.table.as_ref(), |
3630 | classes: self.classes.clone(), |
3631 | stride2: self.stride2, |
3632 | } |
3633 | } |
3634 | |
3635 | /// Converts this transition table to an owned value. |
3636 | #[cfg (feature = "alloc" )] |
3637 | fn to_owned(&self) -> TransitionTable<alloc::vec::Vec<u32>> { |
3638 | TransitionTable { |
3639 | table: self.table.as_ref().to_vec(), |
3640 | classes: self.classes.clone(), |
3641 | stride2: self.stride2, |
3642 | } |
3643 | } |
3644 | |
3645 | /// Return the state for the given ID. If the given ID is not valid, then |
3646 | /// this panics. |
3647 | fn state(&self, id: StateID) -> State<'_> { |
3648 | assert!(self.is_valid(id)); |
3649 | |
3650 | let i = id.as_usize(); |
3651 | State { |
3652 | id, |
3653 | stride2: self.stride2, |
3654 | transitions: &self.table()[i..i + self.alphabet_len()], |
3655 | } |
3656 | } |
3657 | |
3658 | /// Returns an iterator over all states in this transition table. |
3659 | /// |
3660 | /// This iterator yields a tuple for each state. The first element of the |
3661 | /// tuple corresponds to a state's identifier, and the second element |
3662 | /// corresponds to the state itself (comprised of its transitions). |
3663 | fn states(&self) -> StateIter<'_, T> { |
3664 | StateIter { |
3665 | tt: self, |
3666 | it: self.table().chunks(self.stride()).enumerate(), |
3667 | } |
3668 | } |
3669 | |
3670 | /// Convert a state identifier to an index to a state (in the range |
3671 | /// 0..self.len()). |
3672 | /// |
3673 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state |
3674 | /// to some other information (such as a remapped state ID). |
3675 | /// |
3676 | /// If the given ID is not valid, then this may panic or produce an |
3677 | /// incorrect index. |
3678 | fn to_index(&self, id: StateID) -> usize { |
3679 | id.as_usize() >> self.stride2 |
3680 | } |
3681 | |
3682 | /// Convert an index to a state (in the range 0..self.len()) to an actual |
3683 | /// state identifier. |
3684 | /// |
3685 | /// This is useful when using a `Vec<T>` as an efficient map keyed by state |
3686 | /// to some other information (such as a remapped state ID). |
3687 | /// |
3688 | /// If the given index is not in the specified range, then this may panic |
3689 | /// or produce an incorrect state ID. |
3690 | fn to_state_id(&self, index: usize) -> StateID { |
3691 | // CORRECTNESS: If the given index is not valid, then it is not |
3692 | // required for this to panic or return a valid state ID. |
3693 | StateID::new_unchecked(index << self.stride2) |
3694 | } |
3695 | |
3696 | /// Returns the state ID for the state immediately following the one given. |
3697 | /// |
3698 | /// This does not check whether the state ID returned is invalid. In fact, |
3699 | /// if the state ID given is the last state in this DFA, then the state ID |
3700 | /// returned is guaranteed to be invalid. |
3701 | #[cfg (feature = "dfa-build" )] |
3702 | fn next_state_id(&self, id: StateID) -> StateID { |
3703 | self.to_state_id(self.to_index(id).checked_add(1).unwrap()) |
3704 | } |
3705 | |
3706 | /// Returns the state ID for the state immediately preceding the one given. |
3707 | /// |
3708 | /// If the dead ID given (which is zero), then this panics. |
3709 | #[cfg (feature = "dfa-build" )] |
3710 | fn prev_state_id(&self, id: StateID) -> StateID { |
3711 | self.to_state_id(self.to_index(id).checked_sub(1).unwrap()) |
3712 | } |
3713 | |
3714 | /// Returns the table as a slice of state IDs. |
3715 | fn table(&self) -> &[StateID] { |
3716 | wire::u32s_to_state_ids(self.table.as_ref()) |
3717 | } |
3718 | |
3719 | /// Returns the total number of states in this transition table. |
3720 | /// |
3721 | /// Note that a DFA always has at least two states: the dead and quit |
3722 | /// states. In particular, the dead state always has ID 0 and is |
3723 | /// correspondingly always the first state. The dead state is never a match |
3724 | /// state. |
3725 | fn len(&self) -> usize { |
3726 | self.table().len() >> self.stride2 |
3727 | } |
3728 | |
3729 | /// Returns the total stride for every state in this DFA. This corresponds |
3730 | /// to the total number of transitions used by each state in this DFA's |
3731 | /// transition table. |
3732 | fn stride(&self) -> usize { |
3733 | 1 << self.stride2 |
3734 | } |
3735 | |
3736 | /// Returns the total number of elements in the alphabet for this |
3737 | /// transition table. This is always less than or equal to `self.stride()`. |
3738 | /// It is only equal when the alphabet length is a power of 2. Otherwise, |
3739 | /// it is always strictly less. |
3740 | fn alphabet_len(&self) -> usize { |
3741 | self.classes.alphabet_len() |
3742 | } |
3743 | |
3744 | /// Returns true if and only if the given state ID is valid for this |
3745 | /// transition table. Validity in this context means that the given ID can |
3746 | /// be used as a valid offset with `self.stride()` to index this transition |
3747 | /// table. |
3748 | fn is_valid(&self, id: StateID) -> bool { |
3749 | let id = id.as_usize(); |
3750 | id < self.table().len() && id % self.stride() == 0 |
3751 | } |
3752 | |
3753 | /// Return the memory usage, in bytes, of this transition table. |
3754 | /// |
3755 | /// This does not include the size of a `TransitionTable` value itself. |
3756 | fn memory_usage(&self) -> usize { |
3757 | self.table().len() * StateID::SIZE |
3758 | } |
3759 | } |
3760 | |
3761 | #[cfg (feature = "dfa-build" )] |
3762 | impl<T: AsMut<[u32]>> TransitionTable<T> { |
3763 | /// Returns the table as a slice of state IDs. |
3764 | fn table_mut(&mut self) -> &mut [StateID] { |
3765 | wire::u32s_to_state_ids_mut(self.table.as_mut()) |
3766 | } |
3767 | } |
3768 | |
3769 | /// The set of all possible starting states in a DFA. |
3770 | /// |
3771 | /// The set of starting states corresponds to the possible choices one can make |
3772 | /// in terms of starting a DFA. That is, before following the first transition, |
3773 | /// you first need to select the state that you start in. |
3774 | /// |
3775 | /// Normally, a DFA converted from an NFA that has a single starting state |
3776 | /// would itself just have one starting state. However, our support for look |
3777 | /// around generally requires more starting states. The correct starting state |
3778 | /// is chosen based on certain properties of the position at which we begin |
3779 | /// our search. |
3780 | /// |
3781 | /// Before listing those properties, we first must define two terms: |
3782 | /// |
3783 | /// * `haystack` - The bytes to execute the search. The search always starts |
3784 | /// at the beginning of `haystack` and ends before or at the end of |
3785 | /// `haystack`. |
3786 | /// * `context` - The (possibly empty) bytes surrounding `haystack`. `haystack` |
3787 | /// must be contained within `context` such that `context` is at least as big |
3788 | /// as `haystack`. |
3789 | /// |
3790 | /// This split is crucial for dealing with look-around. For example, consider |
3791 | /// the context `foobarbaz`, the haystack `bar` and the regex `^bar$`. This |
3792 | /// regex should _not_ match the haystack since `bar` does not appear at the |
3793 | /// beginning of the input. Similarly, the regex `\Bbar\B` should match the |
3794 | /// haystack because `bar` is not surrounded by word boundaries. But a search |
3795 | /// that does not take context into account would not permit `\B` to match |
3796 | /// since the beginning of any string matches a word boundary. Similarly, a |
3797 | /// search that does not take context into account when searching `^bar$` in |
3798 | /// the haystack `bar` would produce a match when it shouldn't. |
3799 | /// |
3800 | /// Thus, it follows that the starting state is chosen based on the following |
3801 | /// criteria, derived from the position at which the search starts in the |
3802 | /// `context` (corresponding to the start of `haystack`): |
3803 | /// |
3804 | /// 1. If the search starts at the beginning of `context`, then the `Text` |
3805 | /// start state is used. (Since `^` corresponds to |
3806 | /// `hir::Anchor::Start`.) |
3807 | /// 2. If the search starts at a position immediately following a line |
3808 | /// terminator, then the `Line` start state is used. (Since `(?m:^)` |
3809 | /// corresponds to `hir::Anchor::StartLF`.) |
3810 | /// 3. If the search starts at a position immediately following a byte |
3811 | /// classified as a "word" character (`[_0-9a-zA-Z]`), then the `WordByte` |
3812 | /// start state is used. (Since `(?-u:\b)` corresponds to a word boundary.) |
3813 | /// 4. Otherwise, if the search starts at a position immediately following |
3814 | /// a byte that is not classified as a "word" character (`[^_0-9a-zA-Z]`), |
3815 | /// then the `NonWordByte` start state is used. (Since `(?-u:\B)` |
3816 | /// corresponds to a not-word-boundary.) |
3817 | /// |
3818 | /// (N.B. Unicode word boundaries are not supported by the DFA because they |
3819 | /// require multi-byte look-around and this is difficult to support in a DFA.) |
3820 | /// |
3821 | /// To further complicate things, we also support constructing individual |
3822 | /// anchored start states for each pattern in the DFA. (Which is required to |
3823 | /// implement overlapping regexes correctly, but is also generally useful.) |
3824 | /// Thus, when individual start states for each pattern are enabled, then the |
3825 | /// total number of start states represented is `4 + (4 * #patterns)`, where |
3826 | /// the 4 comes from each of the 4 possibilities above. The first 4 represents |
3827 | /// the starting states for the entire DFA, which support searching for |
3828 | /// multiple patterns simultaneously (possibly unanchored). |
3829 | /// |
3830 | /// If individual start states are disabled, then this will only store 4 |
3831 | /// start states. Typically, individual start states are only enabled when |
3832 | /// constructing the reverse DFA for regex matching. But they are also useful |
3833 | /// for building DFAs that can search for a specific pattern or even to support |
3834 | /// both anchored and unanchored searches with the same DFA. |
3835 | /// |
3836 | /// Note though that while the start table always has either `4` or |
3837 | /// `4 + (4 * #patterns)` starting state *ids*, the total number of states |
3838 | /// might be considerably smaller. That is, many of the IDs may be duplicative. |
3839 | /// (For example, if a regex doesn't have a `\b` sub-pattern, then there's no |
3840 | /// reason to generate a unique starting state for handling word boundaries. |
3841 | /// Similarly for start/end anchors.) |
3842 | #[derive(Clone)] |
3843 | pub(crate) struct StartTable<T> { |
3844 | /// The initial start state IDs. |
3845 | /// |
3846 | /// In practice, T is either `Vec<u32>` or `&[u32]`. |
3847 | /// |
3848 | /// The first `2 * stride` (currently always 8) entries always correspond |
3849 | /// to the starts states for the entire DFA, with the first 4 entries being |
3850 | /// for unanchored searches and the second 4 entries being for anchored |
3851 | /// searches. To keep things simple, we always use 8 entries even if the |
3852 | /// `StartKind` is not both. |
3853 | /// |
3854 | /// After that, there are `stride * patterns` state IDs, where `patterns` |
3855 | /// may be zero in the case of a DFA with no patterns or in the case where |
3856 | /// the DFA was built without enabling starting states for each pattern. |
3857 | table: T, |
3858 | /// The starting state configuration supported. When 'both', both |
3859 | /// unanchored and anchored searches work. When 'unanchored', anchored |
3860 | /// searches panic. When 'anchored', unanchored searches panic. |
3861 | kind: StartKind, |
3862 | /// The start state configuration for every possible byte. |
3863 | start_map: StartByteMap, |
3864 | /// The number of starting state IDs per pattern. |
3865 | stride: usize, |
3866 | /// The total number of patterns for which starting states are encoded. |
3867 | /// This is `None` for DFAs that were built without start states for each |
3868 | /// pattern. Thus, one cannot use this field to say how many patterns |
3869 | /// are in the DFA in all cases. It is specific to how many patterns are |
3870 | /// represented in this start table. |
3871 | pattern_len: Option<usize>, |
3872 | /// The universal starting state for unanchored searches. This is only |
3873 | /// present when the DFA supports unanchored searches and when all starting |
3874 | /// state IDs for an unanchored search are equivalent. |
3875 | universal_start_unanchored: Option<StateID>, |
3876 | /// The universal starting state for anchored searches. This is only |
3877 | /// present when the DFA supports anchored searches and when all starting |
3878 | /// state IDs for an anchored search are equivalent. |
3879 | universal_start_anchored: Option<StateID>, |
3880 | } |
3881 | |
3882 | #[cfg (feature = "dfa-build" )] |
3883 | impl StartTable<Vec<u32>> { |
3884 | /// Create a valid set of start states all pointing to the dead state. |
3885 | /// |
3886 | /// When the corresponding DFA is constructed with start states for each |
3887 | /// pattern, then `patterns` should be the number of patterns. Otherwise, |
3888 | /// it should be zero. |
3889 | /// |
3890 | /// If the total table size could exceed the allocatable limit, then this |
3891 | /// returns an error. In practice, this is unlikely to be able to occur, |
3892 | /// since it's likely that allocation would have failed long before it got |
3893 | /// to this point. |
3894 | fn dead( |
3895 | kind: StartKind, |
3896 | lookm: &LookMatcher, |
3897 | pattern_len: Option<usize>, |
3898 | ) -> Result<StartTable<Vec<u32>>, BuildError> { |
3899 | if let Some(len) = pattern_len { |
3900 | assert!(len <= PatternID::LIMIT); |
3901 | } |
3902 | let stride = Start::len(); |
3903 | // OK because 2*4 is never going to overflow anything. |
3904 | let starts_len = stride.checked_mul(2).unwrap(); |
3905 | let pattern_starts_len = |
3906 | match stride.checked_mul(pattern_len.unwrap_or(0)) { |
3907 | Some(x) => x, |
3908 | None => return Err(BuildError::too_many_start_states()), |
3909 | }; |
3910 | let table_len = match starts_len.checked_add(pattern_starts_len) { |
3911 | Some(x) => x, |
3912 | None => return Err(BuildError::too_many_start_states()), |
3913 | }; |
3914 | if let Err(_) = isize::try_from(table_len) { |
3915 | return Err(BuildError::too_many_start_states()); |
3916 | } |
3917 | let table = vec![DEAD.as_u32(); table_len]; |
3918 | let start_map = StartByteMap::new(lookm); |
3919 | Ok(StartTable { |
3920 | table, |
3921 | kind, |
3922 | start_map, |
3923 | stride, |
3924 | pattern_len, |
3925 | universal_start_unanchored: None, |
3926 | universal_start_anchored: None, |
3927 | }) |
3928 | } |
3929 | } |
3930 | |
3931 | impl<'a> StartTable<&'a [u32]> { |
3932 | /// Deserialize a table of start state IDs starting at the beginning of |
3933 | /// `slice`. Upon success, return the total number of bytes read along with |
3934 | /// the table of starting state IDs. |
3935 | /// |
3936 | /// If there was a problem deserializing any part of the starting IDs, |
3937 | /// then this returns an error. Notably, if the given slice does not have |
3938 | /// the same alignment as `StateID`, then this will return an error (among |
3939 | /// other possible errors). |
3940 | /// |
3941 | /// This is guaranteed to execute in constant time. |
3942 | /// |
3943 | /// # Safety |
3944 | /// |
3945 | /// This routine is not safe because it does not check the validity of the |
3946 | /// starting state IDs themselves. In particular, the number of starting |
3947 | /// IDs can be of variable length, so it's possible that checking their |
3948 | /// validity cannot be done in constant time. An invalid starting state |
3949 | /// ID is not safe because other code may rely on the starting IDs being |
3950 | /// correct (such as explicit bounds check elision). Therefore, an invalid |
3951 | /// start ID can lead to undefined behavior. |
3952 | /// |
3953 | /// Callers that use this function must either pass on the safety invariant |
3954 | /// or guarantee that the bytes given contain valid starting state IDs. |
3955 | /// This guarantee is upheld by the bytes written by `write_to`. |
3956 | unsafe fn from_bytes_unchecked( |
3957 | mut slice: &'a [u8], |
3958 | ) -> Result<(StartTable<&'a [u32]>, usize), DeserializeError> { |
3959 | let slice_start = slice.as_ptr().as_usize(); |
3960 | |
3961 | let (kind, nr) = StartKind::from_bytes(slice)?; |
3962 | slice = &slice[nr..]; |
3963 | |
3964 | let (start_map, nr) = StartByteMap::from_bytes(slice)?; |
3965 | slice = &slice[nr..]; |
3966 | |
3967 | let (stride, nr) = |
3968 | wire::try_read_u32_as_usize(slice, "start table stride" )?; |
3969 | slice = &slice[nr..]; |
3970 | if stride != Start::len() { |
3971 | return Err(DeserializeError::generic( |
3972 | "invalid starting table stride" , |
3973 | )); |
3974 | } |
3975 | |
3976 | let (maybe_pattern_len, nr) = |
3977 | wire::try_read_u32_as_usize(slice, "start table patterns" )?; |
3978 | slice = &slice[nr..]; |
3979 | let pattern_len = if maybe_pattern_len.as_u32() == u32::MAX { |
3980 | None |
3981 | } else { |
3982 | Some(maybe_pattern_len) |
3983 | }; |
3984 | if pattern_len.map_or(false, |len| len > PatternID::LIMIT) { |
3985 | return Err(DeserializeError::generic( |
3986 | "invalid number of patterns" , |
3987 | )); |
3988 | } |
3989 | |
3990 | let (universal_unanchored, nr) = |
3991 | wire::try_read_u32(slice, "universal unanchored start" )?; |
3992 | slice = &slice[nr..]; |
3993 | let universal_start_unanchored = if universal_unanchored == u32::MAX { |
3994 | None |
3995 | } else { |
3996 | Some(StateID::try_from(universal_unanchored).map_err(|e| { |
3997 | DeserializeError::state_id_error( |
3998 | e, |
3999 | "universal unanchored start" , |
4000 | ) |
4001 | })?) |
4002 | }; |
4003 | |
4004 | let (universal_anchored, nr) = |
4005 | wire::try_read_u32(slice, "universal anchored start" )?; |
4006 | slice = &slice[nr..]; |
4007 | let universal_start_anchored = if universal_anchored == u32::MAX { |
4008 | None |
4009 | } else { |
4010 | Some(StateID::try_from(universal_anchored).map_err(|e| { |
4011 | DeserializeError::state_id_error(e, "universal anchored start" ) |
4012 | })?) |
4013 | }; |
4014 | |
4015 | let pattern_table_size = wire::mul( |
4016 | stride, |
4017 | pattern_len.unwrap_or(0), |
4018 | "invalid pattern length" , |
4019 | )?; |
4020 | // Our start states always start with a two stride of start states for |
4021 | // the entire automaton. The first stride is for unanchored starting |
4022 | // states and the second stride is for anchored starting states. What |
4023 | // follows it are an optional set of start states for each pattern. |
4024 | let start_state_len = wire::add( |
4025 | wire::mul(2, stride, "start state stride too big" )?, |
4026 | pattern_table_size, |
4027 | "invalid 'any' pattern starts size" , |
4028 | )?; |
4029 | let table_bytes_len = wire::mul( |
4030 | start_state_len, |
4031 | StateID::SIZE, |
4032 | "pattern table bytes length" , |
4033 | )?; |
4034 | wire::check_slice_len(slice, table_bytes_len, "start ID table" )?; |
4035 | wire::check_alignment::<StateID>(slice)?; |
4036 | let table_bytes = &slice[..table_bytes_len]; |
4037 | slice = &slice[table_bytes_len..]; |
4038 | // SAFETY: Since StateID is always representable as a u32, all we need |
4039 | // to do is ensure that we have the proper length and alignment. We've |
4040 | // checked both above, so the cast below is safe. |
4041 | // |
4042 | // N.B. This is the only not-safe code in this function. |
4043 | let table = core::slice::from_raw_parts( |
4044 | table_bytes.as_ptr().cast::<u32>(), |
4045 | start_state_len, |
4046 | ); |
4047 | let st = StartTable { |
4048 | table, |
4049 | kind, |
4050 | start_map, |
4051 | stride, |
4052 | pattern_len, |
4053 | universal_start_unanchored, |
4054 | universal_start_anchored, |
4055 | }; |
4056 | Ok((st, slice.as_ptr().as_usize() - slice_start)) |
4057 | } |
4058 | } |
4059 | |
4060 | impl<T: AsRef<[u32]>> StartTable<T> { |
4061 | /// Writes a serialized form of this start table to the buffer given. If |
4062 | /// the buffer is too small, then an error is returned. To determine how |
4063 | /// big the buffer must be, use `write_to_len`. |
4064 | fn write_to<E: Endian>( |
4065 | &self, |
4066 | mut dst: &mut [u8], |
4067 | ) -> Result<usize, SerializeError> { |
4068 | let nwrite = self.write_to_len(); |
4069 | if dst.len() < nwrite { |
4070 | return Err(SerializeError::buffer_too_small( |
4071 | "starting table ids" , |
4072 | )); |
4073 | } |
4074 | dst = &mut dst[..nwrite]; |
4075 | |
4076 | // write start kind |
4077 | let nw = self.kind.write_to::<E>(dst)?; |
4078 | dst = &mut dst[nw..]; |
4079 | // write start byte map |
4080 | let nw = self.start_map.write_to(dst)?; |
4081 | dst = &mut dst[nw..]; |
4082 | // write stride |
4083 | // Unwrap is OK since the stride is always 4 (currently). |
4084 | E::write_u32(u32::try_from(self.stride).unwrap(), dst); |
4085 | dst = &mut dst[size_of::<u32>()..]; |
4086 | // write pattern length |
4087 | // Unwrap is OK since number of patterns is guaranteed to fit in a u32. |
4088 | E::write_u32( |
4089 | u32::try_from(self.pattern_len.unwrap_or(0xFFFF_FFFF)).unwrap(), |
4090 | dst, |
4091 | ); |
4092 | dst = &mut dst[size_of::<u32>()..]; |
4093 | // write universal start unanchored state id, u32::MAX if absent |
4094 | E::write_u32( |
4095 | self.universal_start_unanchored |
4096 | .map_or(u32::MAX, |sid| sid.as_u32()), |
4097 | dst, |
4098 | ); |
4099 | dst = &mut dst[size_of::<u32>()..]; |
4100 | // write universal start anchored state id, u32::MAX if absent |
4101 | E::write_u32( |
4102 | self.universal_start_anchored.map_or(u32::MAX, |sid| sid.as_u32()), |
4103 | dst, |
4104 | ); |
4105 | dst = &mut dst[size_of::<u32>()..]; |
4106 | // write start IDs |
4107 | for &sid in self.table() { |
4108 | let n = wire::write_state_id::<E>(sid, &mut dst); |
4109 | dst = &mut dst[n..]; |
4110 | } |
4111 | Ok(nwrite) |
4112 | } |
4113 | |
4114 | /// Returns the number of bytes the serialized form of this start ID table |
4115 | /// will use. |
4116 | fn write_to_len(&self) -> usize { |
4117 | self.kind.write_to_len() |
4118 | + self.start_map.write_to_len() |
4119 | + size_of::<u32>() // stride |
4120 | + size_of::<u32>() // # patterns |
4121 | + size_of::<u32>() // universal unanchored start |
4122 | + size_of::<u32>() // universal anchored start |
4123 | + (self.table().len() * StateID::SIZE) |
4124 | } |
4125 | |
4126 | /// Validates that every state ID in this start table is valid by checking |
4127 | /// it against the given transition table (which must be for the same DFA). |
4128 | /// |
4129 | /// That is, every state ID can be used to correctly index a state. |
4130 | fn validate( |
4131 | &self, |
4132 | tt: &TransitionTable<T>, |
4133 | ) -> Result<(), DeserializeError> { |
4134 | if !self.universal_start_unanchored.map_or(true, |s| tt.is_valid(s)) { |
4135 | return Err(DeserializeError::generic( |
4136 | "found invalid universal unanchored starting state ID" , |
4137 | )); |
4138 | } |
4139 | if !self.universal_start_anchored.map_or(true, |s| tt.is_valid(s)) { |
4140 | return Err(DeserializeError::generic( |
4141 | "found invalid universal anchored starting state ID" , |
4142 | )); |
4143 | } |
4144 | for &id in self.table() { |
4145 | if !tt.is_valid(id) { |
4146 | return Err(DeserializeError::generic( |
4147 | "found invalid starting state ID" , |
4148 | )); |
4149 | } |
4150 | } |
4151 | Ok(()) |
4152 | } |
4153 | |
4154 | /// Converts this start list to a borrowed value. |
4155 | fn as_ref(&self) -> StartTable<&'_ [u32]> { |
4156 | StartTable { |
4157 | table: self.table.as_ref(), |
4158 | kind: self.kind, |
4159 | start_map: self.start_map.clone(), |
4160 | stride: self.stride, |
4161 | pattern_len: self.pattern_len, |
4162 | universal_start_unanchored: self.universal_start_unanchored, |
4163 | universal_start_anchored: self.universal_start_anchored, |
4164 | } |
4165 | } |
4166 | |
4167 | /// Converts this start list to an owned value. |
4168 | #[cfg (feature = "alloc" )] |
4169 | fn to_owned(&self) -> StartTable<alloc::vec::Vec<u32>> { |
4170 | StartTable { |
4171 | table: self.table.as_ref().to_vec(), |
4172 | kind: self.kind, |
4173 | start_map: self.start_map.clone(), |
4174 | stride: self.stride, |
4175 | pattern_len: self.pattern_len, |
4176 | universal_start_unanchored: self.universal_start_unanchored, |
4177 | universal_start_anchored: self.universal_start_anchored, |
4178 | } |
4179 | } |
4180 | |
4181 | /// Return the start state for the given input and starting configuration. |
4182 | /// This returns an error if the input configuration is not supported by |
4183 | /// this DFA. For example, requesting an unanchored search when the DFA was |
4184 | /// not built with unanchored starting states. Or asking for an anchored |
4185 | /// pattern search with an invalid pattern ID or on a DFA that was not |
4186 | /// built with start states for each pattern. |
4187 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4188 | fn start( |
4189 | &self, |
4190 | anchored: Anchored, |
4191 | start: Start, |
4192 | ) -> Result<StateID, StartError> { |
4193 | let start_index = start.as_usize(); |
4194 | let index = match anchored { |
4195 | Anchored::No => { |
4196 | if !self.kind.has_unanchored() { |
4197 | return Err(StartError::unsupported_anchored(anchored)); |
4198 | } |
4199 | start_index |
4200 | } |
4201 | Anchored::Yes => { |
4202 | if !self.kind.has_anchored() { |
4203 | return Err(StartError::unsupported_anchored(anchored)); |
4204 | } |
4205 | self.stride + start_index |
4206 | } |
4207 | Anchored::Pattern(pid) => { |
4208 | let len = match self.pattern_len { |
4209 | None => { |
4210 | return Err(StartError::unsupported_anchored(anchored)) |
4211 | } |
4212 | Some(len) => len, |
4213 | }; |
4214 | if pid.as_usize() >= len { |
4215 | return Ok(DEAD); |
4216 | } |
4217 | (2 * self.stride) |
4218 | + (self.stride * pid.as_usize()) |
4219 | + start_index |
4220 | } |
4221 | }; |
4222 | Ok(self.table()[index]) |
4223 | } |
4224 | |
4225 | /// Returns an iterator over all start state IDs in this table. |
4226 | /// |
4227 | /// Each item is a triple of: start state ID, the start state type and the |
4228 | /// pattern ID (if any). |
4229 | fn iter(&self) -> StartStateIter<'_> { |
4230 | StartStateIter { st: self.as_ref(), i: 0 } |
4231 | } |
4232 | |
4233 | /// Returns the table as a slice of state IDs. |
4234 | fn table(&self) -> &[StateID] { |
4235 | wire::u32s_to_state_ids(self.table.as_ref()) |
4236 | } |
4237 | |
4238 | /// Return the memory usage, in bytes, of this start list. |
4239 | /// |
4240 | /// This does not include the size of a `StartList` value itself. |
4241 | fn memory_usage(&self) -> usize { |
4242 | self.table().len() * StateID::SIZE |
4243 | } |
4244 | } |
4245 | |
4246 | #[cfg (feature = "dfa-build" )] |
4247 | impl<T: AsMut<[u32]>> StartTable<T> { |
4248 | /// Set the start state for the given index and pattern. |
4249 | /// |
4250 | /// If the pattern ID or state ID are not valid, then this will panic. |
4251 | fn set_start(&mut self, anchored: Anchored, start: Start, id: StateID) { |
4252 | let start_index = start.as_usize(); |
4253 | let index = match anchored { |
4254 | Anchored::No => start_index, |
4255 | Anchored::Yes => self.stride + start_index, |
4256 | Anchored::Pattern(pid) => { |
4257 | let pid = pid.as_usize(); |
4258 | let len = self |
4259 | .pattern_len |
4260 | .expect("start states for each pattern enabled" ); |
4261 | assert!(pid < len, "invalid pattern ID {:?}" , pid); |
4262 | self.stride |
4263 | .checked_mul(pid) |
4264 | .unwrap() |
4265 | .checked_add(self.stride.checked_mul(2).unwrap()) |
4266 | .unwrap() |
4267 | .checked_add(start_index) |
4268 | .unwrap() |
4269 | } |
4270 | }; |
4271 | self.table_mut()[index] = id; |
4272 | } |
4273 | |
4274 | /// Returns the table as a mutable slice of state IDs. |
4275 | fn table_mut(&mut self) -> &mut [StateID] { |
4276 | wire::u32s_to_state_ids_mut(self.table.as_mut()) |
4277 | } |
4278 | } |
4279 | |
4280 | /// An iterator over start state IDs. |
4281 | /// |
4282 | /// This iterator yields a triple of start state ID, the anchored mode and the |
4283 | /// start state type. If a pattern ID is relevant, then the anchored mode will |
4284 | /// contain it. Start states with an anchored mode containing a pattern ID will |
4285 | /// only occur when the DFA was compiled with start states for each pattern |
4286 | /// (which is disabled by default). |
4287 | pub(crate) struct StartStateIter<'a> { |
4288 | st: StartTable<&'a [u32]>, |
4289 | i: usize, |
4290 | } |
4291 | |
4292 | impl<'a> Iterator for StartStateIter<'a> { |
4293 | type Item = (StateID, Anchored, Start); |
4294 | |
4295 | fn next(&mut self) -> Option<(StateID, Anchored, Start)> { |
4296 | let i = self.i; |
4297 | let table = self.st.table(); |
4298 | if i >= table.len() { |
4299 | return None; |
4300 | } |
4301 | self.i += 1; |
4302 | |
4303 | // This unwrap is okay since the stride of the starting state table |
4304 | // must always match the number of start state types. |
4305 | let start_type = Start::from_usize(i % self.st.stride).unwrap(); |
4306 | let anchored = if i < self.st.stride { |
4307 | Anchored::No |
4308 | } else if i < (2 * self.st.stride) { |
4309 | Anchored::Yes |
4310 | } else { |
4311 | let pid = (i - (2 * self.st.stride)) / self.st.stride; |
4312 | Anchored::Pattern(PatternID::new(pid).unwrap()) |
4313 | }; |
4314 | Some((table[i], anchored, start_type)) |
4315 | } |
4316 | } |
4317 | |
4318 | /// This type represents that patterns that should be reported whenever a DFA |
4319 | /// enters a match state. This structure exists to support DFAs that search for |
4320 | /// matches for multiple regexes. |
4321 | /// |
4322 | /// This structure relies on the fact that all match states in a DFA occur |
4323 | /// contiguously in the DFA's transition table. (See dfa/special.rs for a more |
4324 | /// detailed breakdown of the representation.) Namely, when a match occurs, we |
4325 | /// know its state ID. Since we know the start and end of the contiguous region |
4326 | /// of match states, we can use that to compute the position at which the match |
4327 | /// state occurs. That in turn is used as an offset into this structure. |
4328 | #[derive(Clone, Debug)] |
4329 | struct MatchStates<T> { |
4330 | /// Slices is a flattened sequence of pairs, where each pair points to a |
4331 | /// sub-slice of pattern_ids. The first element of the pair is an offset |
4332 | /// into pattern_ids and the second element of the pair is the number |
4333 | /// of 32-bit pattern IDs starting at that position. That is, each pair |
4334 | /// corresponds to a single DFA match state and its corresponding match |
4335 | /// IDs. The number of pairs always corresponds to the number of distinct |
4336 | /// DFA match states. |
4337 | /// |
4338 | /// In practice, T is either Vec<u32> or &[u32]. |
4339 | slices: T, |
4340 | /// A flattened sequence of pattern IDs for each DFA match state. The only |
4341 | /// way to correctly read this sequence is indirectly via `slices`. |
4342 | /// |
4343 | /// In practice, T is either Vec<u32> or &[u32]. |
4344 | pattern_ids: T, |
4345 | /// The total number of unique patterns represented by these match states. |
4346 | pattern_len: usize, |
4347 | } |
4348 | |
4349 | impl<'a> MatchStates<&'a [u32]> { |
4350 | unsafe fn from_bytes_unchecked( |
4351 | mut slice: &'a [u8], |
4352 | ) -> Result<(MatchStates<&'a [u32]>, usize), DeserializeError> { |
4353 | let slice_start = slice.as_ptr().as_usize(); |
4354 | |
4355 | // Read the total number of match states. |
4356 | let (state_len, nr) = |
4357 | wire::try_read_u32_as_usize(slice, "match state length" )?; |
4358 | slice = &slice[nr..]; |
4359 | |
4360 | // Read the slice start/length pairs. |
4361 | let pair_len = wire::mul(2, state_len, "match state offset pairs" )?; |
4362 | let slices_bytes_len = wire::mul( |
4363 | pair_len, |
4364 | PatternID::SIZE, |
4365 | "match state slice offset byte length" , |
4366 | )?; |
4367 | wire::check_slice_len(slice, slices_bytes_len, "match state slices" )?; |
4368 | wire::check_alignment::<PatternID>(slice)?; |
4369 | let slices_bytes = &slice[..slices_bytes_len]; |
4370 | slice = &slice[slices_bytes_len..]; |
4371 | // SAFETY: Since PatternID is always representable as a u32, all we |
4372 | // need to do is ensure that we have the proper length and alignment. |
4373 | // We've checked both above, so the cast below is safe. |
4374 | // |
4375 | // N.B. This is one of the few not-safe snippets in this function, |
4376 | // so we mark it explicitly to call it out. |
4377 | let slices = core::slice::from_raw_parts( |
4378 | slices_bytes.as_ptr().cast::<u32>(), |
4379 | pair_len, |
4380 | ); |
4381 | |
4382 | // Read the total number of unique pattern IDs (which is always 1 more |
4383 | // than the maximum pattern ID in this automaton, since pattern IDs are |
4384 | // handed out contiguously starting at 0). |
4385 | let (pattern_len, nr) = |
4386 | wire::try_read_u32_as_usize(slice, "pattern length" )?; |
4387 | slice = &slice[nr..]; |
4388 | |
4389 | // Now read the pattern ID length. We don't need to store this |
4390 | // explicitly, but we need it to know how many pattern IDs to read. |
4391 | let (idlen, nr) = |
4392 | wire::try_read_u32_as_usize(slice, "pattern ID length" )?; |
4393 | slice = &slice[nr..]; |
4394 | |
4395 | // Read the actual pattern IDs. |
4396 | let pattern_ids_len = |
4397 | wire::mul(idlen, PatternID::SIZE, "pattern ID byte length" )?; |
4398 | wire::check_slice_len(slice, pattern_ids_len, "match pattern IDs" )?; |
4399 | wire::check_alignment::<PatternID>(slice)?; |
4400 | let pattern_ids_bytes = &slice[..pattern_ids_len]; |
4401 | slice = &slice[pattern_ids_len..]; |
4402 | // SAFETY: Since PatternID is always representable as a u32, all we |
4403 | // need to do is ensure that we have the proper length and alignment. |
4404 | // We've checked both above, so the cast below is safe. |
4405 | // |
4406 | // N.B. This is one of the few not-safe snippets in this function, |
4407 | // so we mark it explicitly to call it out. |
4408 | let pattern_ids = core::slice::from_raw_parts( |
4409 | pattern_ids_bytes.as_ptr().cast::<u32>(), |
4410 | idlen, |
4411 | ); |
4412 | |
4413 | let ms = MatchStates { slices, pattern_ids, pattern_len }; |
4414 | Ok((ms, slice.as_ptr().as_usize() - slice_start)) |
4415 | } |
4416 | } |
4417 | |
4418 | #[cfg (feature = "dfa-build" )] |
4419 | impl MatchStates<Vec<u32>> { |
4420 | fn empty(pattern_len: usize) -> MatchStates<Vec<u32>> { |
4421 | assert!(pattern_len <= PatternID::LIMIT); |
4422 | MatchStates { slices: vec![], pattern_ids: vec![], pattern_len } |
4423 | } |
4424 | |
4425 | fn new( |
4426 | matches: &BTreeMap<StateID, Vec<PatternID>>, |
4427 | pattern_len: usize, |
4428 | ) -> Result<MatchStates<Vec<u32>>, BuildError> { |
4429 | let mut m = MatchStates::empty(pattern_len); |
4430 | for (_, pids) in matches.iter() { |
4431 | let start = PatternID::new(m.pattern_ids.len()) |
4432 | .map_err(|_| BuildError::too_many_match_pattern_ids())?; |
4433 | m.slices.push(start.as_u32()); |
4434 | // This is always correct since the number of patterns in a single |
4435 | // match state can never exceed maximum number of allowable |
4436 | // patterns. Why? Because a pattern can only appear once in a |
4437 | // particular match state, by construction. (And since our pattern |
4438 | // ID limit is one less than u32::MAX, we're guaranteed that the |
4439 | // length fits in a u32.) |
4440 | m.slices.push(u32::try_from(pids.len()).unwrap()); |
4441 | for &pid in pids { |
4442 | m.pattern_ids.push(pid.as_u32()); |
4443 | } |
4444 | } |
4445 | m.pattern_len = pattern_len; |
4446 | Ok(m) |
4447 | } |
4448 | |
4449 | fn new_with_map( |
4450 | &self, |
4451 | matches: &BTreeMap<StateID, Vec<PatternID>>, |
4452 | ) -> Result<MatchStates<Vec<u32>>, BuildError> { |
4453 | MatchStates::new(matches, self.pattern_len) |
4454 | } |
4455 | } |
4456 | |
4457 | impl<T: AsRef<[u32]>> MatchStates<T> { |
4458 | /// Writes a serialized form of these match states to the buffer given. If |
4459 | /// the buffer is too small, then an error is returned. To determine how |
4460 | /// big the buffer must be, use `write_to_len`. |
4461 | fn write_to<E: Endian>( |
4462 | &self, |
4463 | mut dst: &mut [u8], |
4464 | ) -> Result<usize, SerializeError> { |
4465 | let nwrite = self.write_to_len(); |
4466 | if dst.len() < nwrite { |
4467 | return Err(SerializeError::buffer_too_small("match states" )); |
4468 | } |
4469 | dst = &mut dst[..nwrite]; |
4470 | |
4471 | // write state ID length |
4472 | // Unwrap is OK since number of states is guaranteed to fit in a u32. |
4473 | E::write_u32(u32::try_from(self.len()).unwrap(), dst); |
4474 | dst = &mut dst[size_of::<u32>()..]; |
4475 | |
4476 | // write slice offset pairs |
4477 | for &pid in self.slices() { |
4478 | let n = wire::write_pattern_id::<E>(pid, &mut dst); |
4479 | dst = &mut dst[n..]; |
4480 | } |
4481 | |
4482 | // write unique pattern ID length |
4483 | // Unwrap is OK since number of patterns is guaranteed to fit in a u32. |
4484 | E::write_u32(u32::try_from(self.pattern_len).unwrap(), dst); |
4485 | dst = &mut dst[size_of::<u32>()..]; |
4486 | |
4487 | // write pattern ID length |
4488 | // Unwrap is OK since we check at construction (and deserialization) |
4489 | // that the number of patterns is representable as a u32. |
4490 | E::write_u32(u32::try_from(self.pattern_ids().len()).unwrap(), dst); |
4491 | dst = &mut dst[size_of::<u32>()..]; |
4492 | |
4493 | // write pattern IDs |
4494 | for &pid in self.pattern_ids() { |
4495 | let n = wire::write_pattern_id::<E>(pid, &mut dst); |
4496 | dst = &mut dst[n..]; |
4497 | } |
4498 | |
4499 | Ok(nwrite) |
4500 | } |
4501 | |
4502 | /// Returns the number of bytes the serialized form of these match states |
4503 | /// will use. |
4504 | fn write_to_len(&self) -> usize { |
4505 | size_of::<u32>() // match state length |
4506 | + (self.slices().len() * PatternID::SIZE) |
4507 | + size_of::<u32>() // unique pattern ID length |
4508 | + size_of::<u32>() // pattern ID length |
4509 | + (self.pattern_ids().len() * PatternID::SIZE) |
4510 | } |
4511 | |
4512 | /// Valides that the match state info is itself internally consistent and |
4513 | /// consistent with the recorded match state region in the given DFA. |
4514 | fn validate(&self, dfa: &DFA<T>) -> Result<(), DeserializeError> { |
4515 | if self.len() != dfa.special.match_len(dfa.stride()) { |
4516 | return Err(DeserializeError::generic( |
4517 | "match state length mismatch" , |
4518 | )); |
4519 | } |
4520 | for si in 0..self.len() { |
4521 | let start = self.slices()[si * 2].as_usize(); |
4522 | let len = self.slices()[si * 2 + 1].as_usize(); |
4523 | if start >= self.pattern_ids().len() { |
4524 | return Err(DeserializeError::generic( |
4525 | "invalid pattern ID start offset" , |
4526 | )); |
4527 | } |
4528 | if start + len > self.pattern_ids().len() { |
4529 | return Err(DeserializeError::generic( |
4530 | "invalid pattern ID length" , |
4531 | )); |
4532 | } |
4533 | for mi in 0..len { |
4534 | let pid = self.pattern_id(si, mi); |
4535 | if pid.as_usize() >= self.pattern_len { |
4536 | return Err(DeserializeError::generic( |
4537 | "invalid pattern ID" , |
4538 | )); |
4539 | } |
4540 | } |
4541 | } |
4542 | Ok(()) |
4543 | } |
4544 | |
4545 | /// Converts these match states back into their map form. This is useful |
4546 | /// when shuffling states, as the normal MatchStates representation is not |
4547 | /// amenable to easy state swapping. But with this map, to swap id1 and |
4548 | /// id2, all you need to do is: |
4549 | /// |
4550 | /// if let Some(pids) = map.remove(&id1) { |
4551 | /// map.insert(id2, pids); |
4552 | /// } |
4553 | /// |
4554 | /// Once shuffling is done, use MatchStates::new to convert back. |
4555 | #[cfg (feature = "dfa-build" )] |
4556 | fn to_map(&self, dfa: &DFA<T>) -> BTreeMap<StateID, Vec<PatternID>> { |
4557 | let mut map = BTreeMap::new(); |
4558 | for i in 0..self.len() { |
4559 | let mut pids = vec![]; |
4560 | for j in 0..self.pattern_len(i) { |
4561 | pids.push(self.pattern_id(i, j)); |
4562 | } |
4563 | map.insert(self.match_state_id(dfa, i), pids); |
4564 | } |
4565 | map |
4566 | } |
4567 | |
4568 | /// Converts these match states to a borrowed value. |
4569 | fn as_ref(&self) -> MatchStates<&'_ [u32]> { |
4570 | MatchStates { |
4571 | slices: self.slices.as_ref(), |
4572 | pattern_ids: self.pattern_ids.as_ref(), |
4573 | pattern_len: self.pattern_len, |
4574 | } |
4575 | } |
4576 | |
4577 | /// Converts these match states to an owned value. |
4578 | #[cfg (feature = "alloc" )] |
4579 | fn to_owned(&self) -> MatchStates<alloc::vec::Vec<u32>> { |
4580 | MatchStates { |
4581 | slices: self.slices.as_ref().to_vec(), |
4582 | pattern_ids: self.pattern_ids.as_ref().to_vec(), |
4583 | pattern_len: self.pattern_len, |
4584 | } |
4585 | } |
4586 | |
4587 | /// Returns the match state ID given the match state index. (Where the |
4588 | /// first match state corresponds to index 0.) |
4589 | /// |
4590 | /// This panics if there is no match state at the given index. |
4591 | fn match_state_id(&self, dfa: &DFA<T>, index: usize) -> StateID { |
4592 | assert!(dfa.special.matches(), "no match states to index" ); |
4593 | // This is one of the places where we rely on the fact that match |
4594 | // states are contiguous in the transition table. Namely, that the |
4595 | // first match state ID always corresponds to dfa.special.min_start. |
4596 | // From there, since we know the stride, we can compute the ID of any |
4597 | // match state given its index. |
4598 | let stride2 = u32::try_from(dfa.stride2()).unwrap(); |
4599 | let offset = index.checked_shl(stride2).unwrap(); |
4600 | let id = dfa.special.min_match.as_usize().checked_add(offset).unwrap(); |
4601 | let sid = StateID::new(id).unwrap(); |
4602 | assert!(dfa.is_match_state(sid)); |
4603 | sid |
4604 | } |
4605 | |
4606 | /// Returns the pattern ID at the given match index for the given match |
4607 | /// state. |
4608 | /// |
4609 | /// The match state index is the state index minus the state index of the |
4610 | /// first match state in the DFA. |
4611 | /// |
4612 | /// The match index is the index of the pattern ID for the given state. |
4613 | /// The index must be less than `self.pattern_len(state_index)`. |
4614 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4615 | fn pattern_id(&self, state_index: usize, match_index: usize) -> PatternID { |
4616 | self.pattern_id_slice(state_index)[match_index] |
4617 | } |
4618 | |
4619 | /// Returns the number of patterns in the given match state. |
4620 | /// |
4621 | /// The match state index is the state index minus the state index of the |
4622 | /// first match state in the DFA. |
4623 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4624 | fn pattern_len(&self, state_index: usize) -> usize { |
4625 | self.slices()[state_index * 2 + 1].as_usize() |
4626 | } |
4627 | |
4628 | /// Returns all of the pattern IDs for the given match state index. |
4629 | /// |
4630 | /// The match state index is the state index minus the state index of the |
4631 | /// first match state in the DFA. |
4632 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4633 | fn pattern_id_slice(&self, state_index: usize) -> &[PatternID] { |
4634 | let start = self.slices()[state_index * 2].as_usize(); |
4635 | let len = self.pattern_len(state_index); |
4636 | &self.pattern_ids()[start..start + len] |
4637 | } |
4638 | |
4639 | /// Returns the pattern ID offset slice of u32 as a slice of PatternID. |
4640 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4641 | fn slices(&self) -> &[PatternID] { |
4642 | wire::u32s_to_pattern_ids(self.slices.as_ref()) |
4643 | } |
4644 | |
4645 | /// Returns the total number of match states. |
4646 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4647 | fn len(&self) -> usize { |
4648 | assert_eq!(0, self.slices().len() % 2); |
4649 | self.slices().len() / 2 |
4650 | } |
4651 | |
4652 | /// Returns the pattern ID slice of u32 as a slice of PatternID. |
4653 | #[cfg_attr (feature = "perf-inline" , inline(always))] |
4654 | fn pattern_ids(&self) -> &[PatternID] { |
4655 | wire::u32s_to_pattern_ids(self.pattern_ids.as_ref()) |
4656 | } |
4657 | |
4658 | /// Return the memory usage, in bytes, of these match pairs. |
4659 | fn memory_usage(&self) -> usize { |
4660 | (self.slices().len() + self.pattern_ids().len()) * PatternID::SIZE |
4661 | } |
4662 | } |
4663 | |
4664 | /// A common set of flags for both dense and sparse DFAs. This primarily |
4665 | /// centralizes the serialization format of these flags at a bitset. |
4666 | #[derive(Clone, Copy, Debug)] |
4667 | pub(crate) struct Flags { |
4668 | /// Whether the DFA can match the empty string. When this is false, all |
4669 | /// matches returned by this DFA are guaranteed to have non-zero length. |
4670 | pub(crate) has_empty: bool, |
4671 | /// Whether the DFA should only produce matches with spans that correspond |
4672 | /// to valid UTF-8. This also includes omitting any zero-width matches that |
4673 | /// split the UTF-8 encoding of a codepoint. |
4674 | pub(crate) is_utf8: bool, |
4675 | /// Whether the DFA is always anchored or not, regardless of `Input` |
4676 | /// configuration. This is useful for avoiding a reverse scan even when |
4677 | /// executing unanchored searches. |
4678 | pub(crate) is_always_start_anchored: bool, |
4679 | } |
4680 | |
4681 | impl Flags { |
4682 | /// Creates a set of flags for a DFA from an NFA. |
4683 | /// |
4684 | /// N.B. This constructor was defined at the time of writing because all |
4685 | /// of the flags are derived directly from the NFA. If this changes in the |
4686 | /// future, we might be more thoughtful about how the `Flags` value is |
4687 | /// itself built. |
4688 | #[cfg (feature = "dfa-build" )] |
4689 | fn from_nfa(nfa: &thompson::NFA) -> Flags { |
4690 | Flags { |
4691 | has_empty: nfa.has_empty(), |
4692 | is_utf8: nfa.is_utf8(), |
4693 | is_always_start_anchored: nfa.is_always_start_anchored(), |
4694 | } |
4695 | } |
4696 | |
4697 | /// Deserializes the flags from the given slice. On success, this also |
4698 | /// returns the number of bytes read from the slice. |
4699 | pub(crate) fn from_bytes( |
4700 | slice: &[u8], |
4701 | ) -> Result<(Flags, usize), DeserializeError> { |
4702 | let (bits, nread) = wire::try_read_u32(slice, "flag bitset" )?; |
4703 | let flags = Flags { |
4704 | has_empty: bits & (1 << 0) != 0, |
4705 | is_utf8: bits & (1 << 1) != 0, |
4706 | is_always_start_anchored: bits & (1 << 2) != 0, |
4707 | }; |
4708 | Ok((flags, nread)) |
4709 | } |
4710 | |
4711 | /// Writes these flags to the given byte slice. If the buffer is too small, |
4712 | /// then an error is returned. To determine how big the buffer must be, |
4713 | /// use `write_to_len`. |
4714 | pub(crate) fn write_to<E: Endian>( |
4715 | &self, |
4716 | dst: &mut [u8], |
4717 | ) -> Result<usize, SerializeError> { |
4718 | fn bool_to_int(b: bool) -> u32 { |
4719 | if b { |
4720 | 1 |
4721 | } else { |
4722 | 0 |
4723 | } |
4724 | } |
4725 | |
4726 | let nwrite = self.write_to_len(); |
4727 | if dst.len() < nwrite { |
4728 | return Err(SerializeError::buffer_too_small("flag bitset" )); |
4729 | } |
4730 | let bits = (bool_to_int(self.has_empty) << 0) |
4731 | | (bool_to_int(self.is_utf8) << 1) |
4732 | | (bool_to_int(self.is_always_start_anchored) << 2); |
4733 | E::write_u32(bits, dst); |
4734 | Ok(nwrite) |
4735 | } |
4736 | |
4737 | /// Returns the number of bytes the serialized form of these flags |
4738 | /// will use. |
4739 | pub(crate) fn write_to_len(&self) -> usize { |
4740 | size_of::<u32>() |
4741 | } |
4742 | } |
4743 | |
4744 | /// An iterator over all states in a DFA. |
4745 | /// |
4746 | /// This iterator yields a tuple for each state. The first element of the |
4747 | /// tuple corresponds to a state's identifier, and the second element |
4748 | /// corresponds to the state itself (comprised of its transitions). |
4749 | /// |
4750 | /// `'a` corresponding to the lifetime of original DFA, `T` corresponds to |
4751 | /// the type of the transition table itself. |
4752 | pub(crate) struct StateIter<'a, T> { |
4753 | tt: &'a TransitionTable<T>, |
4754 | it: iter::Enumerate<slice::Chunks<'a, StateID>>, |
4755 | } |
4756 | |
4757 | impl<'a, T: AsRef<[u32]>> Iterator for StateIter<'a, T> { |
4758 | type Item = State<'a>; |
4759 | |
4760 | fn next(&mut self) -> Option<State<'a>> { |
4761 | self.it.next().map(|(index, _)| { |
4762 | let id = self.tt.to_state_id(index); |
4763 | self.tt.state(id) |
4764 | }) |
4765 | } |
4766 | } |
4767 | |
4768 | /// An immutable representation of a single DFA state. |
4769 | /// |
4770 | /// `'a` correspondings to the lifetime of a DFA's transition table. |
4771 | pub(crate) struct State<'a> { |
4772 | id: StateID, |
4773 | stride2: usize, |
4774 | transitions: &'a [StateID], |
4775 | } |
4776 | |
4777 | impl<'a> State<'a> { |
4778 | /// Return an iterator over all transitions in this state. This yields |
4779 | /// a number of transitions equivalent to the alphabet length of the |
4780 | /// corresponding DFA. |
4781 | /// |
4782 | /// Each transition is represented by a tuple. The first element is |
4783 | /// the input byte for that transition and the second element is the |
4784 | /// transitions itself. |
4785 | pub(crate) fn transitions(&self) -> StateTransitionIter<'_> { |
4786 | StateTransitionIter { |
4787 | len: self.transitions.len(), |
4788 | it: self.transitions.iter().enumerate(), |
4789 | } |
4790 | } |
4791 | |
4792 | /// Return an iterator over a sparse representation of the transitions in |
4793 | /// this state. Only non-dead transitions are returned. |
4794 | /// |
4795 | /// The "sparse" representation in this case corresponds to a sequence of |
4796 | /// triples. The first two elements of the triple comprise an inclusive |
4797 | /// byte range while the last element corresponds to the transition taken |
4798 | /// for all bytes in the range. |
4799 | /// |
4800 | /// This is somewhat more condensed than the classical sparse |
4801 | /// representation (where you have an element for every non-dead |
4802 | /// transition), but in practice, checking if a byte is in a range is very |
4803 | /// cheap and using ranges tends to conserve quite a bit more space. |
4804 | pub(crate) fn sparse_transitions(&self) -> StateSparseTransitionIter<'_> { |
4805 | StateSparseTransitionIter { dense: self.transitions(), cur: None } |
4806 | } |
4807 | |
4808 | /// Returns the identifier for this state. |
4809 | pub(crate) fn id(&self) -> StateID { |
4810 | self.id |
4811 | } |
4812 | |
4813 | /// Analyzes this state to determine whether it can be accelerated. If so, |
4814 | /// it returns an accelerator that contains at least one byte. |
4815 | #[cfg (feature = "dfa-build" )] |
4816 | fn accelerate(&self, classes: &ByteClasses) -> Option<Accel> { |
4817 | // We just try to add bytes to our accelerator. Once adding fails |
4818 | // (because we've added too many bytes), then give up. |
4819 | let mut accel = Accel::new(); |
4820 | for (class, id) in self.transitions() { |
4821 | if id == self.id() { |
4822 | continue; |
4823 | } |
4824 | for unit in classes.elements(class) { |
4825 | if let Some(byte) = unit.as_u8() { |
4826 | if !accel.add(byte) { |
4827 | return None; |
4828 | } |
4829 | } |
4830 | } |
4831 | } |
4832 | if accel.is_empty() { |
4833 | None |
4834 | } else { |
4835 | Some(accel) |
4836 | } |
4837 | } |
4838 | } |
4839 | |
4840 | impl<'a> fmt::Debug for State<'a> { |
4841 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
4842 | for (i, (start, end, sid)) in self.sparse_transitions().enumerate() { |
4843 | let id = if f.alternate() { |
4844 | sid.as_usize() |
4845 | } else { |
4846 | sid.as_usize() >> self.stride2 |
4847 | }; |
4848 | if i > 0 { |
4849 | write!(f, ", " )?; |
4850 | } |
4851 | if start == end { |
4852 | write!(f, "{:?} => {:?}" , start, id)?; |
4853 | } else { |
4854 | write!(f, "{:?}-{:?} => {:?}" , start, end, id)?; |
4855 | } |
4856 | } |
4857 | Ok(()) |
4858 | } |
4859 | } |
4860 | |
4861 | /// An iterator over all transitions in a single DFA state. This yields |
4862 | /// a number of transitions equivalent to the alphabet length of the |
4863 | /// corresponding DFA. |
4864 | /// |
4865 | /// Each transition is represented by a tuple. The first element is the input |
4866 | /// byte for that transition and the second element is the transition itself. |
4867 | #[derive(Debug)] |
4868 | pub(crate) struct StateTransitionIter<'a> { |
4869 | len: usize, |
4870 | it: iter::Enumerate<slice::Iter<'a, StateID>>, |
4871 | } |
4872 | |
4873 | impl<'a> Iterator for StateTransitionIter<'a> { |
4874 | type Item = (alphabet::Unit, StateID); |
4875 | |
4876 | fn next(&mut self) -> Option<(alphabet::Unit, StateID)> { |
4877 | self.it.next().map(|(i, &id)| { |
4878 | let unit = if i + 1 == self.len { |
4879 | alphabet::Unit::eoi(i) |
4880 | } else { |
4881 | let b = u8::try_from(i) |
4882 | .expect("raw byte alphabet is never exceeded" ); |
4883 | alphabet::Unit::u8(b) |
4884 | }; |
4885 | (unit, id) |
4886 | }) |
4887 | } |
4888 | } |
4889 | |
4890 | /// An iterator over all non-DEAD transitions in a single DFA state using a |
4891 | /// sparse representation. |
4892 | /// |
4893 | /// Each transition is represented by a triple. The first two elements of the |
4894 | /// triple comprise an inclusive byte range while the last element corresponds |
4895 | /// to the transition taken for all bytes in the range. |
4896 | /// |
4897 | /// As a convenience, this always returns `alphabet::Unit` values of the same |
4898 | /// type. That is, you'll never get a (byte, EOI) or a (EOI, byte). Only (byte, |
4899 | /// byte) and (EOI, EOI) values are yielded. |
4900 | #[derive(Debug)] |
4901 | pub(crate) struct StateSparseTransitionIter<'a> { |
4902 | dense: StateTransitionIter<'a>, |
4903 | cur: Option<(alphabet::Unit, alphabet::Unit, StateID)>, |
4904 | } |
4905 | |
4906 | impl<'a> Iterator for StateSparseTransitionIter<'a> { |
4907 | type Item = (alphabet::Unit, alphabet::Unit, StateID); |
4908 | |
4909 | fn next(&mut self) -> Option<(alphabet::Unit, alphabet::Unit, StateID)> { |
4910 | while let Some((unit, next)) = self.dense.next() { |
4911 | let (prev_start, prev_end, prev_next) = match self.cur { |
4912 | Some(t) => t, |
4913 | None => { |
4914 | self.cur = Some((unit, unit, next)); |
4915 | continue; |
4916 | } |
4917 | }; |
4918 | if prev_next == next && !unit.is_eoi() { |
4919 | self.cur = Some((prev_start, unit, prev_next)); |
4920 | } else { |
4921 | self.cur = Some((unit, unit, next)); |
4922 | if prev_next != DEAD { |
4923 | return Some((prev_start, prev_end, prev_next)); |
4924 | } |
4925 | } |
4926 | } |
4927 | if let Some((start, end, next)) = self.cur.take() { |
4928 | if next != DEAD { |
4929 | return Some((start, end, next)); |
4930 | } |
4931 | } |
4932 | None |
4933 | } |
4934 | } |
4935 | |
4936 | /// An error that occurred during the construction of a DFA. |
4937 | /// |
4938 | /// This error does not provide many introspection capabilities. There are |
4939 | /// generally only two things you can do with it: |
4940 | /// |
4941 | /// * Obtain a human readable message via its `std::fmt::Display` impl. |
4942 | /// * Access an underlying [`nfa::thompson::BuildError`](thompson::BuildError) |
4943 | /// type from its `source` method via the `std::error::Error` trait. This error |
4944 | /// only occurs when using convenience routines for building a DFA directly |
4945 | /// from a pattern string. |
4946 | /// |
4947 | /// When the `std` feature is enabled, this implements the `std::error::Error` |
4948 | /// trait. |
4949 | #[cfg (feature = "dfa-build" )] |
4950 | #[derive(Clone, Debug)] |
4951 | pub struct BuildError { |
4952 | kind: BuildErrorKind, |
4953 | } |
4954 | |
4955 | /// The kind of error that occurred during the construction of a DFA. |
4956 | /// |
4957 | /// Note that this error is non-exhaustive. Adding new variants is not |
4958 | /// considered a breaking change. |
4959 | #[cfg (feature = "dfa-build" )] |
4960 | #[derive(Clone, Debug)] |
4961 | enum BuildErrorKind { |
4962 | /// An error that occurred while constructing an NFA as a precursor step |
4963 | /// before a DFA is compiled. |
4964 | NFA(thompson::BuildError), |
4965 | /// An error that occurred because an unsupported regex feature was used. |
4966 | /// The message string describes which unsupported feature was used. |
4967 | /// |
4968 | /// The primary regex feature that is unsupported by DFAs is the Unicode |
4969 | /// word boundary look-around assertion (`\b`). This can be worked around |
4970 | /// by either using an ASCII word boundary (`(?-u:\b)`) or by enabling |
4971 | /// Unicode word boundaries when building a DFA. |
4972 | Unsupported(&'static str), |
4973 | /// An error that occurs if too many states are produced while building a |
4974 | /// DFA. |
4975 | TooManyStates, |
4976 | /// An error that occurs if too many start states are needed while building |
4977 | /// a DFA. |
4978 | /// |
4979 | /// This is a kind of oddball error that occurs when building a DFA with |
4980 | /// start states enabled for each pattern and enough patterns to cause |
4981 | /// the table of start states to overflow `usize`. |
4982 | TooManyStartStates, |
4983 | /// This is another oddball error that can occur if there are too many |
4984 | /// patterns spread out across too many match states. |
4985 | TooManyMatchPatternIDs, |
4986 | /// An error that occurs if the DFA got too big during determinization. |
4987 | DFAExceededSizeLimit { limit: usize }, |
4988 | /// An error that occurs if auxiliary storage (not the DFA) used during |
4989 | /// determinization got too big. |
4990 | DeterminizeExceededSizeLimit { limit: usize }, |
4991 | } |
4992 | |
4993 | #[cfg (feature = "dfa-build" )] |
4994 | impl BuildError { |
4995 | /// Return the kind of this error. |
4996 | fn kind(&self) -> &BuildErrorKind { |
4997 | &self.kind |
4998 | } |
4999 | |
5000 | pub(crate) fn nfa(err: thompson::BuildError) -> BuildError { |
5001 | BuildError { kind: BuildErrorKind::NFA(err) } |
5002 | } |
5003 | |
5004 | pub(crate) fn unsupported_dfa_word_boundary_unicode() -> BuildError { |
5005 | let msg = "cannot build DFAs for regexes with Unicode word \ |
5006 | boundaries; switch to ASCII word boundaries, or \ |
5007 | heuristically enable Unicode word boundaries or use a \ |
5008 | different regex engine" ; |
5009 | BuildError { kind: BuildErrorKind::Unsupported(msg) } |
5010 | } |
5011 | |
5012 | pub(crate) fn too_many_states() -> BuildError { |
5013 | BuildError { kind: BuildErrorKind::TooManyStates } |
5014 | } |
5015 | |
5016 | pub(crate) fn too_many_start_states() -> BuildError { |
5017 | BuildError { kind: BuildErrorKind::TooManyStartStates } |
5018 | } |
5019 | |
5020 | pub(crate) fn too_many_match_pattern_ids() -> BuildError { |
5021 | BuildError { kind: BuildErrorKind::TooManyMatchPatternIDs } |
5022 | } |
5023 | |
5024 | pub(crate) fn dfa_exceeded_size_limit(limit: usize) -> BuildError { |
5025 | BuildError { kind: BuildErrorKind::DFAExceededSizeLimit { limit } } |
5026 | } |
5027 | |
5028 | pub(crate) fn determinize_exceeded_size_limit(limit: usize) -> BuildError { |
5029 | BuildError { |
5030 | kind: BuildErrorKind::DeterminizeExceededSizeLimit { limit }, |
5031 | } |
5032 | } |
5033 | } |
5034 | |
5035 | #[cfg (all(feature = "std" , feature = "dfa-build" ))] |
5036 | impl std::error::Error for BuildError { |
5037 | fn source(&self) -> Option<&(dyn std::error::Error + 'static)> { |
5038 | match self.kind() { |
5039 | BuildErrorKind::NFA(ref err) => Some(err), |
5040 | _ => None, |
5041 | } |
5042 | } |
5043 | } |
5044 | |
5045 | #[cfg (feature = "dfa-build" )] |
5046 | impl core::fmt::Display for BuildError { |
5047 | fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { |
5048 | match self.kind() { |
5049 | BuildErrorKind::NFA(_) => write!(f, "error building NFA" ), |
5050 | BuildErrorKind::Unsupported(ref msg) => { |
5051 | write!(f, "unsupported regex feature for DFAs: {}" , msg) |
5052 | } |
5053 | BuildErrorKind::TooManyStates => write!( |
5054 | f, |
5055 | "number of DFA states exceeds limit of {}" , |
5056 | StateID::LIMIT, |
5057 | ), |
5058 | BuildErrorKind::TooManyStartStates => { |
5059 | let stride = Start::len(); |
5060 | // The start table has `stride` entries for starting states for |
5061 | // the entire DFA, and then `stride` entries for each pattern |
5062 | // if start states for each pattern are enabled (which is the |
5063 | // only way this error can occur). Thus, the total number of |
5064 | // patterns that can fit in the table is `stride` less than |
5065 | // what we can allocate. |
5066 | let max = usize::try_from(core::isize::MAX).unwrap(); |
5067 | let limit = (max - stride) / stride; |
5068 | write!( |
5069 | f, |
5070 | "compiling DFA with start states exceeds pattern \ |
5071 | pattern limit of {}" , |
5072 | limit, |
5073 | ) |
5074 | } |
5075 | BuildErrorKind::TooManyMatchPatternIDs => write!( |
5076 | f, |
5077 | "compiling DFA with total patterns in all match states \ |
5078 | exceeds limit of {}" , |
5079 | PatternID::LIMIT, |
5080 | ), |
5081 | BuildErrorKind::DFAExceededSizeLimit { limit } => write!( |
5082 | f, |
5083 | "DFA exceeded size limit of {:?} during determinization" , |
5084 | limit, |
5085 | ), |
5086 | BuildErrorKind::DeterminizeExceededSizeLimit { limit } => { |
5087 | write!(f, "determinization exceeded size limit of {:?}" , limit) |
5088 | } |
5089 | } |
5090 | } |
5091 | } |
5092 | |
5093 | #[cfg (all(test, feature = "syntax" , feature = "dfa-build" ))] |
5094 | mod tests { |
5095 | use crate::{Input, MatchError}; |
5096 | |
5097 | use super::*; |
5098 | |
5099 | #[test] |
5100 | fn errors_with_unicode_word_boundary() { |
5101 | let pattern = r"\b" ; |
5102 | assert!(Builder::new().build(pattern).is_err()); |
5103 | } |
5104 | |
5105 | #[test] |
5106 | fn roundtrip_never_match() { |
5107 | let dfa = DFA::never_match().unwrap(); |
5108 | let (buf, _) = dfa.to_bytes_native_endian(); |
5109 | let dfa: DFA<&[u32]> = DFA::from_bytes(&buf).unwrap().0; |
5110 | |
5111 | assert_eq!(None, dfa.try_search_fwd(&Input::new("foo12345" )).unwrap()); |
5112 | } |
5113 | |
5114 | #[test] |
5115 | fn roundtrip_always_match() { |
5116 | use crate::HalfMatch; |
5117 | |
5118 | let dfa = DFA::always_match().unwrap(); |
5119 | let (buf, _) = dfa.to_bytes_native_endian(); |
5120 | let dfa: DFA<&[u32]> = DFA::from_bytes(&buf).unwrap().0; |
5121 | |
5122 | assert_eq!( |
5123 | Some(HalfMatch::must(0, 0)), |
5124 | dfa.try_search_fwd(&Input::new("foo12345" )).unwrap() |
5125 | ); |
5126 | } |
5127 | |
5128 | // See the analogous test in src/hybrid/dfa.rs. |
5129 | #[test] |
5130 | fn heuristic_unicode_reverse() { |
5131 | let dfa = DFA::builder() |
5132 | .configure(DFA::config().unicode_word_boundary(true)) |
5133 | .thompson(thompson::Config::new().reverse(true)) |
5134 | .build(r"\b[0-9]+\b" ) |
5135 | .unwrap(); |
5136 | |
5137 | let input = Input::new("β123" ).range(2..); |
5138 | let expected = MatchError::quit(0xB2, 1); |
5139 | let got = dfa.try_search_rev(&input); |
5140 | assert_eq!(Err(expected), got); |
5141 | |
5142 | let input = Input::new("123β" ).range(..3); |
5143 | let expected = MatchError::quit(0xCE, 3); |
5144 | let got = dfa.try_search_rev(&input); |
5145 | assert_eq!(Err(expected), got); |
5146 | } |
5147 | } |
5148 | |