1 | #[cfg (feature = "std" )] |
2 | use core::fmt; |
3 | #[cfg (feature = "std" )] |
4 | use core::iter; |
5 | use core::mem; |
6 | use core::slice; |
7 | |
8 | #[cfg (feature = "std" )] |
9 | use byteorder::{BigEndian, LittleEndian}; |
10 | use byteorder::{ByteOrder, NativeEndian}; |
11 | #[cfg (feature = "std" )] |
12 | use regex_syntax::ParserBuilder; |
13 | |
14 | use classes::ByteClasses; |
15 | #[cfg (feature = "std" )] |
16 | use determinize::Determinizer; |
17 | use dfa::DFA; |
18 | #[cfg (feature = "std" )] |
19 | use error::{Error, Result}; |
20 | #[cfg (feature = "std" )] |
21 | use minimize::Minimizer; |
22 | #[cfg (feature = "std" )] |
23 | use nfa::{self, NFA}; |
24 | #[cfg (feature = "std" )] |
25 | use sparse::SparseDFA; |
26 | use state_id::{dead_id, StateID}; |
27 | #[cfg (feature = "std" )] |
28 | use state_id::{ |
29 | next_state_id, premultiply_overflow_error, write_state_id_bytes, |
30 | }; |
31 | |
32 | /// The size of the alphabet in a standard DFA. |
33 | /// |
34 | /// Specifically, this length controls the number of transitions present in |
35 | /// each DFA state. However, when the byte class optimization is enabled, |
36 | /// then each DFA maps the space of all possible 256 byte values to at most |
37 | /// 256 distinct equivalence classes. In this case, the number of distinct |
38 | /// equivalence classes corresponds to the internal alphabet of the DFA, in the |
39 | /// sense that each DFA state has a number of transitions equal to the number |
40 | /// of equivalence classes despite supporting matching on all possible byte |
41 | /// values. |
42 | const ALPHABET_LEN: usize = 256; |
43 | |
44 | /// Masks used in serialization of DFAs. |
45 | pub(crate) const MASK_PREMULTIPLIED: u16 = 0b0000_0000_0000_0001; |
46 | pub(crate) const MASK_ANCHORED: u16 = 0b0000_0000_0000_0010; |
47 | |
48 | /// A dense table-based deterministic finite automaton (DFA). |
49 | /// |
50 | /// A dense DFA represents the core matching primitive in this crate. That is, |
51 | /// logically, all DFAs have a single start state, one or more match states |
52 | /// and a transition table that maps the current state and the current byte of |
53 | /// input to the next state. A DFA can use this information to implement fast |
54 | /// searching. In particular, the use of a dense DFA generally makes the trade |
55 | /// off that match speed is the most valuable characteristic, even if building |
56 | /// the regex may take significant time *and* space. As such, the processing |
57 | /// of every byte of input is done with a small constant number of operations |
58 | /// that does not vary with the pattern, its size or the size of the alphabet. |
59 | /// If your needs don't line up with this trade off, then a dense DFA may not |
60 | /// be an adequate solution to your problem. |
61 | /// |
62 | /// In contrast, a [sparse DFA](enum.SparseDFA.html) makes the opposite |
63 | /// trade off: it uses less space but will execute a variable number of |
64 | /// instructions per byte at match time, which makes it slower for matching. |
65 | /// |
66 | /// A DFA can be built using the default configuration via the |
67 | /// [`DenseDFA::new`](enum.DenseDFA.html#method.new) constructor. Otherwise, |
68 | /// one can configure various aspects via the |
69 | /// [`dense::Builder`](dense/struct.Builder.html). |
70 | /// |
71 | /// A single DFA fundamentally supports the following operations: |
72 | /// |
73 | /// 1. Detection of a match. |
74 | /// 2. Location of the end of the first possible match. |
75 | /// 3. Location of the end of the leftmost-first match. |
76 | /// |
77 | /// A notable absence from the above list of capabilities is the location of |
78 | /// the *start* of a match. In order to provide both the start and end of a |
79 | /// match, *two* DFAs are required. This functionality is provided by a |
80 | /// [`Regex`](struct.Regex.html), which can be built with its basic |
81 | /// constructor, [`Regex::new`](struct.Regex.html#method.new), or with |
82 | /// a [`RegexBuilder`](struct.RegexBuilder.html). |
83 | /// |
84 | /// # State size |
85 | /// |
86 | /// A `DenseDFA` has two type parameters, `T` and `S`. `T` corresponds to |
87 | /// the type of the DFA's transition table while `S` corresponds to the |
88 | /// representation used for the DFA's state identifiers as described by the |
89 | /// [`StateID`](trait.StateID.html) trait. This type parameter is typically |
90 | /// `usize`, but other valid choices provided by this crate include `u8`, |
91 | /// `u16`, `u32` and `u64`. The primary reason for choosing a different state |
92 | /// identifier representation than the default is to reduce the amount of |
93 | /// memory used by a DFA. Note though, that if the chosen representation cannot |
94 | /// accommodate the size of your DFA, then building the DFA will fail and |
95 | /// return an error. |
96 | /// |
97 | /// While the reduction in heap memory used by a DFA is one reason for choosing |
98 | /// a smaller state identifier representation, another possible reason is for |
99 | /// decreasing the serialization size of a DFA, as returned by |
100 | /// [`to_bytes_little_endian`](enum.DenseDFA.html#method.to_bytes_little_endian), |
101 | /// [`to_bytes_big_endian`](enum.DenseDFA.html#method.to_bytes_big_endian) |
102 | /// or |
103 | /// [`to_bytes_native_endian`](enum.DenseDFA.html#method.to_bytes_native_endian). |
104 | /// |
105 | /// The type of the transition table is typically either `Vec<S>` or `&[S]`, |
106 | /// depending on where the transition table is stored. |
107 | /// |
108 | /// # Variants |
109 | /// |
110 | /// This DFA is defined as a non-exhaustive enumeration of different types of |
111 | /// dense DFAs. All of these dense DFAs use the same internal representation |
112 | /// for the transition table, but they vary in how the transition table is |
113 | /// read. A DFA's specific variant depends on the configuration options set via |
114 | /// [`dense::Builder`](dense/struct.Builder.html). The default variant is |
115 | /// `PremultipliedByteClass`. |
116 | /// |
117 | /// # The `DFA` trait |
118 | /// |
119 | /// This type implements the [`DFA`](trait.DFA.html) trait, which means it |
120 | /// can be used for searching. For example: |
121 | /// |
122 | /// ``` |
123 | /// use regex_automata::{DFA, DenseDFA}; |
124 | /// |
125 | /// # fn example() -> Result<(), regex_automata::Error> { |
126 | /// let dfa = DenseDFA::new("foo[0-9]+" )?; |
127 | /// assert_eq!(Some(8), dfa.find(b"foo12345" )); |
128 | /// # Ok(()) }; example().unwrap() |
129 | /// ``` |
130 | /// |
131 | /// The `DFA` trait also provides an assortment of other lower level methods |
132 | /// for DFAs, such as `start_state` and `next_state`. While these are correctly |
133 | /// implemented, it is an anti-pattern to use them in performance sensitive |
134 | /// code on the `DenseDFA` type directly. Namely, each implementation requires |
135 | /// a branch to determine which type of dense DFA is being used. Instead, |
136 | /// this branch should be pushed up a layer in the code since walking the |
137 | /// transitions of a DFA is usually a hot path. If you do need to use these |
138 | /// lower level methods in performance critical code, then you should match on |
139 | /// the variants of this DFA and use each variant's implementation of the `DFA` |
140 | /// trait directly. |
141 | #[derive (Clone, Debug)] |
142 | pub enum DenseDFA<T: AsRef<[S]>, S: StateID> { |
143 | /// A standard DFA that does not use premultiplication or byte classes. |
144 | Standard(Standard<T, S>), |
145 | /// A DFA that shrinks its alphabet to a set of equivalence classes instead |
146 | /// of using all possible byte values. Any two bytes belong to the same |
147 | /// equivalence class if and only if they can be used interchangeably |
148 | /// anywhere in the DFA while never discriminating between a match and a |
149 | /// non-match. |
150 | /// |
151 | /// This type of DFA can result in significant space reduction with a very |
152 | /// small match time performance penalty. |
153 | ByteClass(ByteClass<T, S>), |
154 | /// A DFA that premultiplies all of its state identifiers in its |
155 | /// transition table. This saves an instruction per byte at match time |
156 | /// which improves search performance. |
157 | /// |
158 | /// The only downside of premultiplication is that it may prevent one from |
159 | /// using a smaller state identifier representation than you otherwise |
160 | /// could. |
161 | Premultiplied(Premultiplied<T, S>), |
162 | /// The default configuration of a DFA, which uses byte classes and |
163 | /// premultiplies its state identifiers. |
164 | PremultipliedByteClass(PremultipliedByteClass<T, S>), |
165 | /// Hints that destructuring should not be exhaustive. |
166 | /// |
167 | /// This enum may grow additional variants, so this makes sure clients |
168 | /// don't count on exhaustive matching. (Otherwise, adding a new variant |
169 | /// could break existing code.) |
170 | #[doc (hidden)] |
171 | __Nonexhaustive, |
172 | } |
173 | |
174 | impl<T: AsRef<[S]>, S: StateID> DenseDFA<T, S> { |
175 | /// Return the internal DFA representation. |
176 | /// |
177 | /// All variants share the same internal representation. |
178 | fn repr(&self) -> &Repr<T, S> { |
179 | match *self { |
180 | DenseDFA::Standard(ref r: &Standard) => &r.0, |
181 | DenseDFA::ByteClass(ref r: &ByteClass) => &r.0, |
182 | DenseDFA::Premultiplied(ref r: &Premultiplied) => &r.0, |
183 | DenseDFA::PremultipliedByteClass(ref r: &PremultipliedByteClass) => &r.0, |
184 | DenseDFA::__Nonexhaustive => unreachable!(), |
185 | } |
186 | } |
187 | } |
188 | |
189 | #[cfg (feature = "std" )] |
190 | impl DenseDFA<Vec<usize>, usize> { |
191 | /// Parse the given regular expression using a default configuration and |
192 | /// return the corresponding DFA. |
193 | /// |
194 | /// The default configuration uses `usize` for state IDs, premultiplies |
195 | /// them and reduces the alphabet size by splitting bytes into equivalence |
196 | /// classes. The DFA is *not* minimized. |
197 | /// |
198 | /// If you want a non-default configuration, then use the |
199 | /// [`dense::Builder`](dense/struct.Builder.html) |
200 | /// to set your own configuration. |
201 | /// |
202 | /// # Example |
203 | /// |
204 | /// ``` |
205 | /// use regex_automata::{DFA, DenseDFA}; |
206 | /// |
207 | /// # fn example() -> Result<(), regex_automata::Error> { |
208 | /// let dfa = DenseDFA::new("foo[0-9]+bar" )?; |
209 | /// assert_eq!(Some(11), dfa.find(b"foo12345bar" )); |
210 | /// # Ok(()) }; example().unwrap() |
211 | /// ``` |
212 | pub fn new(pattern: &str) -> Result<DenseDFA<Vec<usize>, usize>> { |
213 | Builder::new().build(pattern) |
214 | } |
215 | } |
216 | |
217 | #[cfg (feature = "std" )] |
218 | impl<S: StateID> DenseDFA<Vec<S>, S> { |
219 | /// Create a new empty DFA that never matches any input. |
220 | /// |
221 | /// # Example |
222 | /// |
223 | /// In order to build an empty DFA, callers must provide a type hint |
224 | /// indicating their choice of state identifier representation. |
225 | /// |
226 | /// ``` |
227 | /// use regex_automata::{DFA, DenseDFA}; |
228 | /// |
229 | /// # fn example() -> Result<(), regex_automata::Error> { |
230 | /// let dfa: DenseDFA<Vec<usize>, usize> = DenseDFA::empty(); |
231 | /// assert_eq!(None, dfa.find(b"" )); |
232 | /// assert_eq!(None, dfa.find(b"foo" )); |
233 | /// # Ok(()) }; example().unwrap() |
234 | /// ``` |
235 | pub fn empty() -> DenseDFA<Vec<S>, S> { |
236 | Repr::empty().into_dense_dfa() |
237 | } |
238 | } |
239 | |
240 | impl<T: AsRef<[S]>, S: StateID> DenseDFA<T, S> { |
241 | /// Cheaply return a borrowed version of this dense DFA. Specifically, the |
242 | /// DFA returned always uses `&[S]` for its transition table while keeping |
243 | /// the same state identifier representation. |
244 | pub fn as_ref<'a>(&'a self) -> DenseDFA<&'a [S], S> { |
245 | match *self { |
246 | DenseDFA::Standard(ref r) => { |
247 | DenseDFA::Standard(Standard(r.0.as_ref())) |
248 | } |
249 | DenseDFA::ByteClass(ref r) => { |
250 | DenseDFA::ByteClass(ByteClass(r.0.as_ref())) |
251 | } |
252 | DenseDFA::Premultiplied(ref r) => { |
253 | DenseDFA::Premultiplied(Premultiplied(r.0.as_ref())) |
254 | } |
255 | DenseDFA::PremultipliedByteClass(ref r) => { |
256 | let inner = PremultipliedByteClass(r.0.as_ref()); |
257 | DenseDFA::PremultipliedByteClass(inner) |
258 | } |
259 | DenseDFA::__Nonexhaustive => unreachable!(), |
260 | } |
261 | } |
262 | |
263 | /// Return an owned version of this sparse DFA. Specifically, the DFA |
264 | /// returned always uses `Vec<u8>` for its transition table while keeping |
265 | /// the same state identifier representation. |
266 | /// |
267 | /// Effectively, this returns a sparse DFA whose transition table lives |
268 | /// on the heap. |
269 | #[cfg (feature = "std" )] |
270 | pub fn to_owned(&self) -> DenseDFA<Vec<S>, S> { |
271 | match *self { |
272 | DenseDFA::Standard(ref r) => { |
273 | DenseDFA::Standard(Standard(r.0.to_owned())) |
274 | } |
275 | DenseDFA::ByteClass(ref r) => { |
276 | DenseDFA::ByteClass(ByteClass(r.0.to_owned())) |
277 | } |
278 | DenseDFA::Premultiplied(ref r) => { |
279 | DenseDFA::Premultiplied(Premultiplied(r.0.to_owned())) |
280 | } |
281 | DenseDFA::PremultipliedByteClass(ref r) => { |
282 | let inner = PremultipliedByteClass(r.0.to_owned()); |
283 | DenseDFA::PremultipliedByteClass(inner) |
284 | } |
285 | DenseDFA::__Nonexhaustive => unreachable!(), |
286 | } |
287 | } |
288 | |
289 | /// Returns the memory usage, in bytes, of this DFA. |
290 | /// |
291 | /// The memory usage is computed based on the number of bytes used to |
292 | /// represent this DFA's transition table. This corresponds to heap memory |
293 | /// usage. |
294 | /// |
295 | /// This does **not** include the stack size used up by this DFA. To |
296 | /// compute that, used `std::mem::size_of::<DenseDFA>()`. |
297 | pub fn memory_usage(&self) -> usize { |
298 | self.repr().memory_usage() |
299 | } |
300 | } |
301 | |
302 | /// Routines for converting a dense DFA to other representations, such as |
303 | /// sparse DFAs, smaller state identifiers or raw bytes suitable for persistent |
304 | /// storage. |
305 | #[cfg (feature = "std" )] |
306 | impl<T: AsRef<[S]>, S: StateID> DenseDFA<T, S> { |
307 | /// Convert this dense DFA to a sparse DFA. |
308 | /// |
309 | /// This is a convenience routine for `to_sparse_sized` that fixes the |
310 | /// state identifier representation of the sparse DFA to the same |
311 | /// representation used for this dense DFA. |
312 | /// |
313 | /// If the chosen state identifier representation is too small to represent |
314 | /// all states in the sparse DFA, then this returns an error. In most |
315 | /// cases, if a dense DFA is constructable with `S` then a sparse DFA will |
316 | /// be as well. However, it is not guaranteed. |
317 | /// |
318 | /// # Example |
319 | /// |
320 | /// ``` |
321 | /// use regex_automata::{DFA, DenseDFA}; |
322 | /// |
323 | /// # fn example() -> Result<(), regex_automata::Error> { |
324 | /// let dense = DenseDFA::new("foo[0-9]+" )?; |
325 | /// let sparse = dense.to_sparse()?; |
326 | /// assert_eq!(Some(8), sparse.find(b"foo12345" )); |
327 | /// # Ok(()) }; example().unwrap() |
328 | /// ``` |
329 | pub fn to_sparse(&self) -> Result<SparseDFA<Vec<u8>, S>> { |
330 | self.to_sparse_sized() |
331 | } |
332 | |
333 | /// Convert this dense DFA to a sparse DFA. |
334 | /// |
335 | /// Using this routine requires supplying a type hint to choose the state |
336 | /// identifier representation for the resulting sparse DFA. |
337 | /// |
338 | /// If the chosen state identifier representation is too small to represent |
339 | /// all states in the sparse DFA, then this returns an error. |
340 | /// |
341 | /// # Example |
342 | /// |
343 | /// ``` |
344 | /// use regex_automata::{DFA, DenseDFA}; |
345 | /// |
346 | /// # fn example() -> Result<(), regex_automata::Error> { |
347 | /// let dense = DenseDFA::new("foo[0-9]+" )?; |
348 | /// let sparse = dense.to_sparse_sized::<u8>()?; |
349 | /// assert_eq!(Some(8), sparse.find(b"foo12345" )); |
350 | /// # Ok(()) }; example().unwrap() |
351 | /// ``` |
352 | pub fn to_sparse_sized<A: StateID>( |
353 | &self, |
354 | ) -> Result<SparseDFA<Vec<u8>, A>> { |
355 | self.repr().to_sparse_sized() |
356 | } |
357 | |
358 | /// Create a new DFA whose match semantics are equivalent to this DFA, |
359 | /// but attempt to use `u8` for the representation of state identifiers. |
360 | /// If `u8` is insufficient to represent all state identifiers in this |
361 | /// DFA, then this returns an error. |
362 | /// |
363 | /// This is a convenience routine for `to_sized::<u8>()`. |
364 | pub fn to_u8(&self) -> Result<DenseDFA<Vec<u8>, u8>> { |
365 | self.to_sized() |
366 | } |
367 | |
368 | /// Create a new DFA whose match semantics are equivalent to this DFA, |
369 | /// but attempt to use `u16` for the representation of state identifiers. |
370 | /// If `u16` is insufficient to represent all state identifiers in this |
371 | /// DFA, then this returns an error. |
372 | /// |
373 | /// This is a convenience routine for `to_sized::<u16>()`. |
374 | pub fn to_u16(&self) -> Result<DenseDFA<Vec<u16>, u16>> { |
375 | self.to_sized() |
376 | } |
377 | |
378 | /// Create a new DFA whose match semantics are equivalent to this DFA, |
379 | /// but attempt to use `u32` for the representation of state identifiers. |
380 | /// If `u32` is insufficient to represent all state identifiers in this |
381 | /// DFA, then this returns an error. |
382 | /// |
383 | /// This is a convenience routine for `to_sized::<u32>()`. |
384 | #[cfg (any(target_pointer_width = "32" , target_pointer_width = "64" ))] |
385 | pub fn to_u32(&self) -> Result<DenseDFA<Vec<u32>, u32>> { |
386 | self.to_sized() |
387 | } |
388 | |
389 | /// Create a new DFA whose match semantics are equivalent to this DFA, |
390 | /// but attempt to use `u64` for the representation of state identifiers. |
391 | /// If `u64` is insufficient to represent all state identifiers in this |
392 | /// DFA, then this returns an error. |
393 | /// |
394 | /// This is a convenience routine for `to_sized::<u64>()`. |
395 | #[cfg (target_pointer_width = "64" )] |
396 | pub fn to_u64(&self) -> Result<DenseDFA<Vec<u64>, u64>> { |
397 | self.to_sized() |
398 | } |
399 | |
400 | /// Create a new DFA whose match semantics are equivalent to this DFA, but |
401 | /// attempt to use `A` for the representation of state identifiers. If `A` |
402 | /// is insufficient to represent all state identifiers in this DFA, then |
403 | /// this returns an error. |
404 | /// |
405 | /// An alternative way to construct such a DFA is to use |
406 | /// [`dense::Builder::build_with_size`](dense/struct.Builder.html#method.build_with_size). |
407 | /// In general, using the builder is preferred since it will use the given |
408 | /// state identifier representation throughout determinization (and |
409 | /// minimization, if done), and thereby using less memory throughout the |
410 | /// entire construction process. However, these routines are necessary |
411 | /// in cases where, say, a minimized DFA could fit in a smaller state |
412 | /// identifier representation, but the initial determinized DFA would not. |
413 | pub fn to_sized<A: StateID>(&self) -> Result<DenseDFA<Vec<A>, A>> { |
414 | self.repr().to_sized().map(|r| r.into_dense_dfa()) |
415 | } |
416 | |
417 | /// Serialize a DFA to raw bytes, aligned to an 8 byte boundary, in little |
418 | /// endian format. |
419 | /// |
420 | /// If the state identifier representation of this DFA has a size different |
421 | /// than 1, 2, 4 or 8 bytes, then this returns an error. All |
422 | /// implementations of `StateID` provided by this crate satisfy this |
423 | /// requirement. |
424 | pub fn to_bytes_little_endian(&self) -> Result<Vec<u8>> { |
425 | self.repr().to_bytes::<LittleEndian>() |
426 | } |
427 | |
428 | /// Serialize a DFA to raw bytes, aligned to an 8 byte boundary, in big |
429 | /// endian format. |
430 | /// |
431 | /// If the state identifier representation of this DFA has a size different |
432 | /// than 1, 2, 4 or 8 bytes, then this returns an error. All |
433 | /// implementations of `StateID` provided by this crate satisfy this |
434 | /// requirement. |
435 | pub fn to_bytes_big_endian(&self) -> Result<Vec<u8>> { |
436 | self.repr().to_bytes::<BigEndian>() |
437 | } |
438 | |
439 | /// Serialize a DFA to raw bytes, aligned to an 8 byte boundary, in native |
440 | /// endian format. Generally, it is better to pick an explicit endianness |
441 | /// using either `to_bytes_little_endian` or `to_bytes_big_endian`. This |
442 | /// routine is useful in tests where the DFA is serialized and deserialized |
443 | /// on the same platform. |
444 | /// |
445 | /// If the state identifier representation of this DFA has a size different |
446 | /// than 1, 2, 4 or 8 bytes, then this returns an error. All |
447 | /// implementations of `StateID` provided by this crate satisfy this |
448 | /// requirement. |
449 | pub fn to_bytes_native_endian(&self) -> Result<Vec<u8>> { |
450 | self.repr().to_bytes::<NativeEndian>() |
451 | } |
452 | } |
453 | |
454 | impl<'a, S: StateID> DenseDFA<&'a [S], S> { |
455 | /// Deserialize a DFA with a specific state identifier representation. |
456 | /// |
457 | /// Deserializing a DFA using this routine will never allocate heap memory. |
458 | /// This is also guaranteed to be a constant time operation that does not |
459 | /// vary with the size of the DFA. |
460 | /// |
461 | /// The bytes given should be generated by the serialization of a DFA with |
462 | /// either the |
463 | /// [`to_bytes_little_endian`](enum.DenseDFA.html#method.to_bytes_little_endian) |
464 | /// method or the |
465 | /// [`to_bytes_big_endian`](enum.DenseDFA.html#method.to_bytes_big_endian) |
466 | /// endian, depending on the endianness of the machine you are |
467 | /// deserializing this DFA from. |
468 | /// |
469 | /// If the state identifier representation is `usize`, then deserialization |
470 | /// is dependent on the pointer size. For this reason, it is best to |
471 | /// serialize DFAs using a fixed size representation for your state |
472 | /// identifiers, such as `u8`, `u16`, `u32` or `u64`. |
473 | /// |
474 | /// # Panics |
475 | /// |
476 | /// The bytes given should be *trusted*. In particular, if the bytes |
477 | /// are not a valid serialization of a DFA, or if the given bytes are |
478 | /// not aligned to an 8 byte boundary, or if the endianness of the |
479 | /// serialized bytes is different than the endianness of the machine that |
480 | /// is deserializing the DFA, then this routine will panic. Moreover, it is |
481 | /// possible for this deserialization routine to succeed even if the given |
482 | /// bytes do not represent a valid serialized dense DFA. |
483 | /// |
484 | /// # Safety |
485 | /// |
486 | /// This routine is unsafe because it permits callers to provide an |
487 | /// arbitrary transition table with possibly incorrect transitions. While |
488 | /// the various serialization routines will never return an incorrect |
489 | /// transition table, there is no guarantee that the bytes provided here |
490 | /// are correct. While deserialization does many checks (as documented |
491 | /// above in the panic conditions), this routine does not check that the |
492 | /// transition table is correct. Given an incorrect transition table, it is |
493 | /// possible for the search routines to access out-of-bounds memory because |
494 | /// of explicit bounds check elision. |
495 | /// |
496 | /// # Example |
497 | /// |
498 | /// This example shows how to serialize a DFA to raw bytes, deserialize it |
499 | /// and then use it for searching. Note that we first convert the DFA to |
500 | /// using `u16` for its state identifier representation before serializing |
501 | /// it. While this isn't strictly necessary, it's good practice in order to |
502 | /// decrease the size of the DFA and to avoid platform specific pitfalls |
503 | /// such as differing pointer sizes. |
504 | /// |
505 | /// ``` |
506 | /// use regex_automata::{DFA, DenseDFA}; |
507 | /// |
508 | /// # fn example() -> Result<(), regex_automata::Error> { |
509 | /// let initial = DenseDFA::new("foo[0-9]+" )?; |
510 | /// let bytes = initial.to_u16()?.to_bytes_native_endian()?; |
511 | /// let dfa: DenseDFA<&[u16], u16> = unsafe { |
512 | /// DenseDFA::from_bytes(&bytes) |
513 | /// }; |
514 | /// |
515 | /// assert_eq!(Some(8), dfa.find(b"foo12345" )); |
516 | /// # Ok(()) }; example().unwrap() |
517 | /// ``` |
518 | pub unsafe fn from_bytes(buf: &'a [u8]) -> DenseDFA<&'a [S], S> { |
519 | Repr::from_bytes(buf).into_dense_dfa() |
520 | } |
521 | } |
522 | |
523 | #[cfg (feature = "std" )] |
524 | impl<S: StateID> DenseDFA<Vec<S>, S> { |
525 | /// Minimize this DFA in place. |
526 | /// |
527 | /// This is not part of the public API. It is only exposed to allow for |
528 | /// more granular external benchmarking. |
529 | #[doc (hidden)] |
530 | pub fn minimize(&mut self) { |
531 | self.repr_mut().minimize(); |
532 | } |
533 | |
534 | /// Return a mutable reference to the internal DFA representation. |
535 | fn repr_mut(&mut self) -> &mut Repr<Vec<S>, S> { |
536 | match *self { |
537 | DenseDFA::Standard(ref mut r: &mut Standard, S>) => &mut r.0, |
538 | DenseDFA::ByteClass(ref mut r: &mut ByteClass, S>) => &mut r.0, |
539 | DenseDFA::Premultiplied(ref mut r: &mut Premultiplied, …>) => &mut r.0, |
540 | DenseDFA::PremultipliedByteClass(ref mut r: &mut PremultipliedByteClass<…, …>) => &mut r.0, |
541 | DenseDFA::__Nonexhaustive => unreachable!(), |
542 | } |
543 | } |
544 | } |
545 | |
546 | impl<T: AsRef<[S]>, S: StateID> DFA for DenseDFA<T, S> { |
547 | type ID = S; |
548 | |
549 | #[inline ] |
550 | fn start_state(&self) -> S { |
551 | self.repr().start_state() |
552 | } |
553 | |
554 | #[inline ] |
555 | fn is_match_state(&self, id: S) -> bool { |
556 | self.repr().is_match_state(id) |
557 | } |
558 | |
559 | #[inline ] |
560 | fn is_dead_state(&self, id: S) -> bool { |
561 | self.repr().is_dead_state(id) |
562 | } |
563 | |
564 | #[inline ] |
565 | fn is_match_or_dead_state(&self, id: S) -> bool { |
566 | self.repr().is_match_or_dead_state(id) |
567 | } |
568 | |
569 | #[inline ] |
570 | fn is_anchored(&self) -> bool { |
571 | self.repr().is_anchored() |
572 | } |
573 | |
574 | #[inline ] |
575 | fn next_state(&self, current: S, input: u8) -> S { |
576 | match *self { |
577 | DenseDFA::Standard(ref r) => r.next_state(current, input), |
578 | DenseDFA::ByteClass(ref r) => r.next_state(current, input), |
579 | DenseDFA::Premultiplied(ref r) => r.next_state(current, input), |
580 | DenseDFA::PremultipliedByteClass(ref r) => { |
581 | r.next_state(current, input) |
582 | } |
583 | DenseDFA::__Nonexhaustive => unreachable!(), |
584 | } |
585 | } |
586 | |
587 | #[inline ] |
588 | unsafe fn next_state_unchecked(&self, current: S, input: u8) -> S { |
589 | match *self { |
590 | DenseDFA::Standard(ref r) => { |
591 | r.next_state_unchecked(current, input) |
592 | } |
593 | DenseDFA::ByteClass(ref r) => { |
594 | r.next_state_unchecked(current, input) |
595 | } |
596 | DenseDFA::Premultiplied(ref r) => { |
597 | r.next_state_unchecked(current, input) |
598 | } |
599 | DenseDFA::PremultipliedByteClass(ref r) => { |
600 | r.next_state_unchecked(current, input) |
601 | } |
602 | DenseDFA::__Nonexhaustive => unreachable!(), |
603 | } |
604 | } |
605 | |
606 | // We specialize the following methods because it lets us lift the |
607 | // case analysis between the different types of dense DFAs. Instead of |
608 | // doing the case analysis for every transition, we do it once before |
609 | // searching. |
610 | |
611 | #[inline ] |
612 | fn is_match_at(&self, bytes: &[u8], start: usize) -> bool { |
613 | match *self { |
614 | DenseDFA::Standard(ref r) => r.is_match_at(bytes, start), |
615 | DenseDFA::ByteClass(ref r) => r.is_match_at(bytes, start), |
616 | DenseDFA::Premultiplied(ref r) => r.is_match_at(bytes, start), |
617 | DenseDFA::PremultipliedByteClass(ref r) => { |
618 | r.is_match_at(bytes, start) |
619 | } |
620 | DenseDFA::__Nonexhaustive => unreachable!(), |
621 | } |
622 | } |
623 | |
624 | #[inline ] |
625 | fn shortest_match_at(&self, bytes: &[u8], start: usize) -> Option<usize> { |
626 | match *self { |
627 | DenseDFA::Standard(ref r) => r.shortest_match_at(bytes, start), |
628 | DenseDFA::ByteClass(ref r) => r.shortest_match_at(bytes, start), |
629 | DenseDFA::Premultiplied(ref r) => { |
630 | r.shortest_match_at(bytes, start) |
631 | } |
632 | DenseDFA::PremultipliedByteClass(ref r) => { |
633 | r.shortest_match_at(bytes, start) |
634 | } |
635 | DenseDFA::__Nonexhaustive => unreachable!(), |
636 | } |
637 | } |
638 | |
639 | #[inline ] |
640 | fn find_at(&self, bytes: &[u8], start: usize) -> Option<usize> { |
641 | match *self { |
642 | DenseDFA::Standard(ref r) => r.find_at(bytes, start), |
643 | DenseDFA::ByteClass(ref r) => r.find_at(bytes, start), |
644 | DenseDFA::Premultiplied(ref r) => r.find_at(bytes, start), |
645 | DenseDFA::PremultipliedByteClass(ref r) => r.find_at(bytes, start), |
646 | DenseDFA::__Nonexhaustive => unreachable!(), |
647 | } |
648 | } |
649 | |
650 | #[inline ] |
651 | fn rfind_at(&self, bytes: &[u8], start: usize) -> Option<usize> { |
652 | match *self { |
653 | DenseDFA::Standard(ref r) => r.rfind_at(bytes, start), |
654 | DenseDFA::ByteClass(ref r) => r.rfind_at(bytes, start), |
655 | DenseDFA::Premultiplied(ref r) => r.rfind_at(bytes, start), |
656 | DenseDFA::PremultipliedByteClass(ref r) => { |
657 | r.rfind_at(bytes, start) |
658 | } |
659 | DenseDFA::__Nonexhaustive => unreachable!(), |
660 | } |
661 | } |
662 | } |
663 | |
664 | /// A standard dense DFA that does not use premultiplication or byte classes. |
665 | /// |
666 | /// Generally, it isn't necessary to use this type directly, since a `DenseDFA` |
667 | /// can be used for searching directly. One possible reason why one might want |
668 | /// to use this type directly is if you are implementing your own search |
669 | /// routines by walking a DFA's transitions directly. In that case, you'll want |
670 | /// to use this type (or any of the other DFA variant types) directly, since |
671 | /// they implement `next_state` more efficiently. |
672 | #[derive (Clone, Debug)] |
673 | pub struct Standard<T: AsRef<[S]>, S: StateID>(Repr<T, S>); |
674 | |
675 | impl<T: AsRef<[S]>, S: StateID> DFA for Standard<T, S> { |
676 | type ID = S; |
677 | |
678 | #[inline ] |
679 | fn start_state(&self) -> S { |
680 | self.0.start_state() |
681 | } |
682 | |
683 | #[inline ] |
684 | fn is_match_state(&self, id: S) -> bool { |
685 | self.0.is_match_state(id) |
686 | } |
687 | |
688 | #[inline ] |
689 | fn is_dead_state(&self, id: S) -> bool { |
690 | self.0.is_dead_state(id) |
691 | } |
692 | |
693 | #[inline ] |
694 | fn is_match_or_dead_state(&self, id: S) -> bool { |
695 | self.0.is_match_or_dead_state(id) |
696 | } |
697 | |
698 | #[inline ] |
699 | fn is_anchored(&self) -> bool { |
700 | self.0.is_anchored() |
701 | } |
702 | |
703 | #[inline ] |
704 | fn next_state(&self, current: S, input: u8) -> S { |
705 | let o = current.to_usize() * ALPHABET_LEN + input as usize; |
706 | self.0.trans()[o] |
707 | } |
708 | |
709 | #[inline ] |
710 | unsafe fn next_state_unchecked(&self, current: S, input: u8) -> S { |
711 | let o = current.to_usize() * ALPHABET_LEN + input as usize; |
712 | *self.0.trans().get_unchecked(o) |
713 | } |
714 | } |
715 | |
716 | /// A dense DFA that shrinks its alphabet. |
717 | /// |
718 | /// Alphabet shrinking is achieved by using a set of equivalence classes |
719 | /// instead of using all possible byte values. Any two bytes belong to the same |
720 | /// equivalence class if and only if they can be used interchangeably anywhere |
721 | /// in the DFA while never discriminating between a match and a non-match. |
722 | /// |
723 | /// This type of DFA can result in significant space reduction with a very |
724 | /// small match time performance penalty. |
725 | /// |
726 | /// Generally, it isn't necessary to use this type directly, since a `DenseDFA` |
727 | /// can be used for searching directly. One possible reason why one might want |
728 | /// to use this type directly is if you are implementing your own search |
729 | /// routines by walking a DFA's transitions directly. In that case, you'll want |
730 | /// to use this type (or any of the other DFA variant types) directly, since |
731 | /// they implement `next_state` more efficiently. |
732 | #[derive (Clone, Debug)] |
733 | pub struct ByteClass<T: AsRef<[S]>, S: StateID>(Repr<T, S>); |
734 | |
735 | impl<T: AsRef<[S]>, S: StateID> DFA for ByteClass<T, S> { |
736 | type ID = S; |
737 | |
738 | #[inline ] |
739 | fn start_state(&self) -> S { |
740 | self.0.start_state() |
741 | } |
742 | |
743 | #[inline ] |
744 | fn is_match_state(&self, id: S) -> bool { |
745 | self.0.is_match_state(id) |
746 | } |
747 | |
748 | #[inline ] |
749 | fn is_dead_state(&self, id: S) -> bool { |
750 | self.0.is_dead_state(id) |
751 | } |
752 | |
753 | #[inline ] |
754 | fn is_match_or_dead_state(&self, id: S) -> bool { |
755 | self.0.is_match_or_dead_state(id) |
756 | } |
757 | |
758 | #[inline ] |
759 | fn is_anchored(&self) -> bool { |
760 | self.0.is_anchored() |
761 | } |
762 | |
763 | #[inline ] |
764 | fn next_state(&self, current: S, input: u8) -> S { |
765 | let input = self.0.byte_classes().get(input); |
766 | let o = current.to_usize() * self.0.alphabet_len() + input as usize; |
767 | self.0.trans()[o] |
768 | } |
769 | |
770 | #[inline ] |
771 | unsafe fn next_state_unchecked(&self, current: S, input: u8) -> S { |
772 | let input = self.0.byte_classes().get_unchecked(input); |
773 | let o = current.to_usize() * self.0.alphabet_len() + input as usize; |
774 | *self.0.trans().get_unchecked(o) |
775 | } |
776 | } |
777 | |
778 | /// A dense DFA that premultiplies all of its state identifiers in its |
779 | /// transition table. |
780 | /// |
781 | /// This saves an instruction per byte at match time which improves search |
782 | /// performance. |
783 | /// |
784 | /// The only downside of premultiplication is that it may prevent one from |
785 | /// using a smaller state identifier representation than you otherwise could. |
786 | /// |
787 | /// Generally, it isn't necessary to use this type directly, since a `DenseDFA` |
788 | /// can be used for searching directly. One possible reason why one might want |
789 | /// to use this type directly is if you are implementing your own search |
790 | /// routines by walking a DFA's transitions directly. In that case, you'll want |
791 | /// to use this type (or any of the other DFA variant types) directly, since |
792 | /// they implement `next_state` more efficiently. |
793 | #[derive (Clone, Debug)] |
794 | pub struct Premultiplied<T: AsRef<[S]>, S: StateID>(Repr<T, S>); |
795 | |
796 | impl<T: AsRef<[S]>, S: StateID> DFA for Premultiplied<T, S> { |
797 | type ID = S; |
798 | |
799 | #[inline ] |
800 | fn start_state(&self) -> S { |
801 | self.0.start_state() |
802 | } |
803 | |
804 | #[inline ] |
805 | fn is_match_state(&self, id: S) -> bool { |
806 | self.0.is_match_state(id) |
807 | } |
808 | |
809 | #[inline ] |
810 | fn is_dead_state(&self, id: S) -> bool { |
811 | self.0.is_dead_state(id) |
812 | } |
813 | |
814 | #[inline ] |
815 | fn is_match_or_dead_state(&self, id: S) -> bool { |
816 | self.0.is_match_or_dead_state(id) |
817 | } |
818 | |
819 | #[inline ] |
820 | fn is_anchored(&self) -> bool { |
821 | self.0.is_anchored() |
822 | } |
823 | |
824 | #[inline ] |
825 | fn next_state(&self, current: S, input: u8) -> S { |
826 | let o = current.to_usize() + input as usize; |
827 | self.0.trans()[o] |
828 | } |
829 | |
830 | #[inline ] |
831 | unsafe fn next_state_unchecked(&self, current: S, input: u8) -> S { |
832 | let o = current.to_usize() + input as usize; |
833 | *self.0.trans().get_unchecked(o) |
834 | } |
835 | } |
836 | |
837 | /// The default configuration of a dense DFA, which uses byte classes and |
838 | /// premultiplies its state identifiers. |
839 | /// |
840 | /// Generally, it isn't necessary to use this type directly, since a `DenseDFA` |
841 | /// can be used for searching directly. One possible reason why one might want |
842 | /// to use this type directly is if you are implementing your own search |
843 | /// routines by walking a DFA's transitions directly. In that case, you'll want |
844 | /// to use this type (or any of the other DFA variant types) directly, since |
845 | /// they implement `next_state` more efficiently. |
846 | #[derive (Clone, Debug)] |
847 | pub struct PremultipliedByteClass<T: AsRef<[S]>, S: StateID>(Repr<T, S>); |
848 | |
849 | impl<T: AsRef<[S]>, S: StateID> DFA for PremultipliedByteClass<T, S> { |
850 | type ID = S; |
851 | |
852 | #[inline ] |
853 | fn start_state(&self) -> S { |
854 | self.0.start_state() |
855 | } |
856 | |
857 | #[inline ] |
858 | fn is_match_state(&self, id: S) -> bool { |
859 | self.0.is_match_state(id) |
860 | } |
861 | |
862 | #[inline ] |
863 | fn is_dead_state(&self, id: S) -> bool { |
864 | self.0.is_dead_state(id) |
865 | } |
866 | |
867 | #[inline ] |
868 | fn is_match_or_dead_state(&self, id: S) -> bool { |
869 | self.0.is_match_or_dead_state(id) |
870 | } |
871 | |
872 | #[inline ] |
873 | fn is_anchored(&self) -> bool { |
874 | self.0.is_anchored() |
875 | } |
876 | |
877 | #[inline ] |
878 | fn next_state(&self, current: S, input: u8) -> S { |
879 | let input = self.0.byte_classes().get(input); |
880 | let o = current.to_usize() + input as usize; |
881 | self.0.trans()[o] |
882 | } |
883 | |
884 | #[inline ] |
885 | unsafe fn next_state_unchecked(&self, current: S, input: u8) -> S { |
886 | let input = self.0.byte_classes().get_unchecked(input); |
887 | let o = current.to_usize() + input as usize; |
888 | *self.0.trans().get_unchecked(o) |
889 | } |
890 | } |
891 | |
892 | /// The internal representation of a dense DFA. |
893 | /// |
894 | /// This representation is shared by all DFA variants. |
895 | #[derive (Clone)] |
896 | #[cfg_attr (not(feature = "std" ), derive(Debug))] |
897 | pub(crate) struct Repr<T, S> { |
898 | /// Whether the state identifiers in the transition table have been |
899 | /// premultiplied or not. |
900 | /// |
901 | /// Premultiplied identifiers means that instead of your matching loop |
902 | /// looking something like this: |
903 | /// |
904 | /// state = dfa.start |
905 | /// for byte in haystack: |
906 | /// next = dfa.transitions[state * len(alphabet) + byte] |
907 | /// if dfa.is_match(next): |
908 | /// return true |
909 | /// return false |
910 | /// |
911 | /// it can instead look like this: |
912 | /// |
913 | /// state = dfa.start |
914 | /// for byte in haystack: |
915 | /// next = dfa.transitions[state + byte] |
916 | /// if dfa.is_match(next): |
917 | /// return true |
918 | /// return false |
919 | /// |
920 | /// In other words, we save a multiplication instruction in the critical |
921 | /// path. This turns out to be a decent performance win. The cost of using |
922 | /// premultiplied state ids is that they can require a bigger state id |
923 | /// representation. |
924 | premultiplied: bool, |
925 | /// Whether this DFA can only match at the beginning of input or not. |
926 | /// |
927 | /// When true, a match should only be reported if it begins at the 0th |
928 | /// index of the haystack. |
929 | anchored: bool, |
930 | /// The initial start state ID. |
931 | start: S, |
932 | /// The total number of states in this DFA. Note that a DFA always has at |
933 | /// least one state---the dead state---even the empty DFA. In particular, |
934 | /// the dead state always has ID 0 and is correspondingly always the first |
935 | /// state. The dead state is never a match state. |
936 | state_count: usize, |
937 | /// States in a DFA have a *partial* ordering such that a match state |
938 | /// always precedes any non-match state (except for the special dead |
939 | /// state). |
940 | /// |
941 | /// `max_match` corresponds to the last state that is a match state. This |
942 | /// encoding has two critical benefits. Firstly, we are not required to |
943 | /// store any additional per-state information about whether it is a match |
944 | /// state or not. Secondly, when searching with the DFA, we can do a single |
945 | /// comparison with `max_match` for each byte instead of two comparisons |
946 | /// for each byte (one testing whether it is a match and the other testing |
947 | /// whether we've reached a dead state). Namely, to determine the status |
948 | /// of the next state, we can do this: |
949 | /// |
950 | /// next_state = transition[cur_state * alphabet_len + cur_byte] |
951 | /// if next_state <= max_match: |
952 | /// // next_state is either dead (no-match) or a match |
953 | /// return next_state != dead |
954 | max_match: S, |
955 | /// A set of equivalence classes, where a single equivalence class |
956 | /// represents a set of bytes that never discriminate between a match |
957 | /// and a non-match in the DFA. Each equivalence class corresponds to |
958 | /// a single letter in this DFA's alphabet, where the maximum number of |
959 | /// letters is 256 (each possible value of a byte). Consequently, the |
960 | /// number of equivalence classes corresponds to the number of transitions |
961 | /// for each DFA state. |
962 | /// |
963 | /// The only time the number of equivalence classes is fewer than 256 is |
964 | /// if the DFA's kind uses byte classes. If the DFA doesn't use byte |
965 | /// classes, then this vector is empty. |
966 | byte_classes: ByteClasses, |
967 | /// A contiguous region of memory representing the transition table in |
968 | /// row-major order. The representation is dense. That is, every state has |
969 | /// precisely the same number of transitions. The maximum number of |
970 | /// transitions is 256. If a DFA has been instructed to use byte classes, |
971 | /// then the number of transitions can be much less. |
972 | /// |
973 | /// In practice, T is either Vec<S> or &[S]. |
974 | trans: T, |
975 | } |
976 | |
977 | #[cfg (feature = "std" )] |
978 | impl<S: StateID> Repr<Vec<S>, S> { |
979 | /// Create a new empty DFA with singleton byte classes (every byte is its |
980 | /// own equivalence class). |
981 | pub fn empty() -> Repr<Vec<S>, S> { |
982 | Repr::empty_with_byte_classes(ByteClasses::singletons()) |
983 | } |
984 | |
985 | /// Create a new empty DFA with the given set of byte equivalence classes. |
986 | /// An empty DFA never matches any input. |
987 | pub fn empty_with_byte_classes( |
988 | byte_classes: ByteClasses, |
989 | ) -> Repr<Vec<S>, S> { |
990 | let mut dfa = Repr { |
991 | premultiplied: false, |
992 | anchored: true, |
993 | start: dead_id(), |
994 | state_count: 0, |
995 | max_match: S::from_usize(0), |
996 | byte_classes, |
997 | trans: vec![], |
998 | }; |
999 | // Every state ID repr must be able to fit at least one state. |
1000 | dfa.add_empty_state().unwrap(); |
1001 | dfa |
1002 | } |
1003 | |
1004 | /// Sets whether this DFA is anchored or not. |
1005 | pub fn anchored(mut self, yes: bool) -> Repr<Vec<S>, S> { |
1006 | self.anchored = yes; |
1007 | self |
1008 | } |
1009 | } |
1010 | |
1011 | impl<T: AsRef<[S]>, S: StateID> Repr<T, S> { |
1012 | /// Convert this internal DFA representation to a DenseDFA based on its |
1013 | /// transition table access pattern. |
1014 | pub fn into_dense_dfa(self) -> DenseDFA<T, S> { |
1015 | match (self.premultiplied, self.byte_classes().is_singleton()) { |
1016 | // no premultiplication, no byte classes |
1017 | (false, true) => DenseDFA::Standard(Standard(self)), |
1018 | // no premultiplication, yes byte classes |
1019 | (false, false) => DenseDFA::ByteClass(ByteClass(self)), |
1020 | // yes premultiplication, no byte classes |
1021 | (true, true) => DenseDFA::Premultiplied(Premultiplied(self)), |
1022 | // yes premultiplication, yes byte classes |
1023 | (true, false) => { |
1024 | DenseDFA::PremultipliedByteClass(PremultipliedByteClass(self)) |
1025 | } |
1026 | } |
1027 | } |
1028 | |
1029 | fn as_ref<'a>(&'a self) -> Repr<&'a [S], S> { |
1030 | Repr { |
1031 | premultiplied: self.premultiplied, |
1032 | anchored: self.anchored, |
1033 | start: self.start, |
1034 | state_count: self.state_count, |
1035 | max_match: self.max_match, |
1036 | byte_classes: self.byte_classes().clone(), |
1037 | trans: self.trans(), |
1038 | } |
1039 | } |
1040 | |
1041 | #[cfg (feature = "std" )] |
1042 | fn to_owned(&self) -> Repr<Vec<S>, S> { |
1043 | Repr { |
1044 | premultiplied: self.premultiplied, |
1045 | anchored: self.anchored, |
1046 | start: self.start, |
1047 | state_count: self.state_count, |
1048 | max_match: self.max_match, |
1049 | byte_classes: self.byte_classes().clone(), |
1050 | trans: self.trans().to_vec(), |
1051 | } |
1052 | } |
1053 | |
1054 | /// Return the starting state of this DFA. |
1055 | /// |
1056 | /// All searches using this DFA must begin at this state. There is exactly |
1057 | /// one starting state for every DFA. A starting state may be a dead state |
1058 | /// or a matching state or neither. |
1059 | pub fn start_state(&self) -> S { |
1060 | self.start |
1061 | } |
1062 | |
1063 | /// Returns true if and only if the given identifier corresponds to a match |
1064 | /// state. |
1065 | pub fn is_match_state(&self, id: S) -> bool { |
1066 | id <= self.max_match && id != dead_id() |
1067 | } |
1068 | |
1069 | /// Returns true if and only if the given identifier corresponds to a dead |
1070 | /// state. |
1071 | pub fn is_dead_state(&self, id: S) -> bool { |
1072 | id == dead_id() |
1073 | } |
1074 | |
1075 | /// Returns true if and only if the given identifier could correspond to |
1076 | /// either a match state or a dead state. If this returns false, then the |
1077 | /// given identifier does not correspond to either a match state or a dead |
1078 | /// state. |
1079 | pub fn is_match_or_dead_state(&self, id: S) -> bool { |
1080 | id <= self.max_match_state() |
1081 | } |
1082 | |
1083 | /// Returns the maximum identifier for which a match state can exist. |
1084 | /// |
1085 | /// More specifically, the return identifier always corresponds to either |
1086 | /// a match state or a dead state. Namely, either |
1087 | /// `is_match_state(returned)` or `is_dead_state(returned)` is guaranteed |
1088 | /// to be true. |
1089 | pub fn max_match_state(&self) -> S { |
1090 | self.max_match |
1091 | } |
1092 | |
1093 | /// Returns true if and only if this DFA is anchored. |
1094 | pub fn is_anchored(&self) -> bool { |
1095 | self.anchored |
1096 | } |
1097 | |
1098 | /// Return the byte classes used by this DFA. |
1099 | pub fn byte_classes(&self) -> &ByteClasses { |
1100 | &self.byte_classes |
1101 | } |
1102 | |
1103 | /// Returns an iterator over all states in this DFA. |
1104 | /// |
1105 | /// This iterator yields a tuple for each state. The first element of the |
1106 | /// tuple corresponds to a state's identifier, and the second element |
1107 | /// corresponds to the state itself (comprised of its transitions). |
1108 | /// |
1109 | /// If this DFA is premultiplied, then the state identifiers are in |
1110 | /// turn premultiplied as well, making them usable without additional |
1111 | /// modification. |
1112 | #[cfg (feature = "std" )] |
1113 | pub fn states(&self) -> StateIter<T, S> { |
1114 | let it = self.trans().chunks(self.alphabet_len()); |
1115 | StateIter { dfa: self, it: it.enumerate() } |
1116 | } |
1117 | |
1118 | /// Return the total number of states in this DFA. Every DFA has at least |
1119 | /// 1 state, even the empty DFA. |
1120 | #[cfg (feature = "std" )] |
1121 | pub fn state_count(&self) -> usize { |
1122 | self.state_count |
1123 | } |
1124 | |
1125 | /// Return the number of elements in this DFA's alphabet. |
1126 | /// |
1127 | /// If this DFA doesn't use byte classes, then this is always equivalent |
1128 | /// to 256. Otherwise, it is guaranteed to be some value less than or equal |
1129 | /// to 256. |
1130 | pub fn alphabet_len(&self) -> usize { |
1131 | self.byte_classes().alphabet_len() |
1132 | } |
1133 | |
1134 | /// Returns the memory usage, in bytes, of this DFA. |
1135 | pub fn memory_usage(&self) -> usize { |
1136 | self.trans().len() * mem::size_of::<S>() |
1137 | } |
1138 | |
1139 | /// Convert the given state identifier to the state's index. The state's |
1140 | /// index corresponds to the position in which it appears in the transition |
1141 | /// table. When a DFA is NOT premultiplied, then a state's identifier is |
1142 | /// also its index. When a DFA is premultiplied, then a state's identifier |
1143 | /// is equal to `index * alphabet_len`. This routine reverses that. |
1144 | #[cfg (feature = "std" )] |
1145 | pub fn state_id_to_index(&self, id: S) -> usize { |
1146 | if self.premultiplied { |
1147 | id.to_usize() / self.alphabet_len() |
1148 | } else { |
1149 | id.to_usize() |
1150 | } |
1151 | } |
1152 | |
1153 | /// Return this DFA's transition table as a slice. |
1154 | fn trans(&self) -> &[S] { |
1155 | self.trans.as_ref() |
1156 | } |
1157 | |
1158 | /// Create a sparse DFA from the internal representation of a dense DFA. |
1159 | #[cfg (feature = "std" )] |
1160 | pub fn to_sparse_sized<A: StateID>( |
1161 | &self, |
1162 | ) -> Result<SparseDFA<Vec<u8>, A>> { |
1163 | SparseDFA::from_dense_sized(self) |
1164 | } |
1165 | |
1166 | /// Create a new DFA whose match semantics are equivalent to this DFA, but |
1167 | /// attempt to use `A` for the representation of state identifiers. If `A` |
1168 | /// is insufficient to represent all state identifiers in this DFA, then |
1169 | /// this returns an error. |
1170 | #[cfg (feature = "std" )] |
1171 | pub fn to_sized<A: StateID>(&self) -> Result<Repr<Vec<A>, A>> { |
1172 | // Check that this DFA can fit into A's representation. |
1173 | let mut last_state_id = self.state_count - 1; |
1174 | if self.premultiplied { |
1175 | last_state_id *= self.alphabet_len(); |
1176 | } |
1177 | if last_state_id > A::max_id() { |
1178 | return Err(Error::state_id_overflow(A::max_id())); |
1179 | } |
1180 | |
1181 | // We're off to the races. The new DFA is the same as the old one, |
1182 | // but its transition table is truncated. |
1183 | let mut new = Repr { |
1184 | premultiplied: self.premultiplied, |
1185 | anchored: self.anchored, |
1186 | start: A::from_usize(self.start.to_usize()), |
1187 | state_count: self.state_count, |
1188 | max_match: A::from_usize(self.max_match.to_usize()), |
1189 | byte_classes: self.byte_classes().clone(), |
1190 | trans: vec![dead_id::<A>(); self.trans().len()], |
1191 | }; |
1192 | for (i, id) in new.trans.iter_mut().enumerate() { |
1193 | *id = A::from_usize(self.trans()[i].to_usize()); |
1194 | } |
1195 | Ok(new) |
1196 | } |
1197 | |
1198 | /// Serialize a DFA to raw bytes, aligned to an 8 byte boundary. |
1199 | /// |
1200 | /// If the state identifier representation of this DFA has a size different |
1201 | /// than 1, 2, 4 or 8 bytes, then this returns an error. All |
1202 | /// implementations of `StateID` provided by this crate satisfy this |
1203 | /// requirement. |
1204 | #[cfg (feature = "std" )] |
1205 | pub(crate) fn to_bytes<A: ByteOrder>(&self) -> Result<Vec<u8>> { |
1206 | let label = b"rust-regex-automata-dfa \x00" ; |
1207 | assert_eq!(24, label.len()); |
1208 | |
1209 | let trans_size = mem::size_of::<S>() * self.trans().len(); |
1210 | let size = |
1211 | // For human readable label. |
1212 | label.len() |
1213 | // endiannes check, must be equal to 0xFEFF for native endian |
1214 | + 2 |
1215 | // For version number. |
1216 | + 2 |
1217 | // Size of state ID representation, in bytes. |
1218 | // Must be 1, 2, 4 or 8. |
1219 | + 2 |
1220 | // For DFA misc options. |
1221 | + 2 |
1222 | // For start state. |
1223 | + 8 |
1224 | // For state count. |
1225 | + 8 |
1226 | // For max match state. |
1227 | + 8 |
1228 | // For byte class map. |
1229 | + 256 |
1230 | // For transition table. |
1231 | + trans_size; |
1232 | // sanity check, this can be updated if need be |
1233 | assert_eq!(312 + trans_size, size); |
1234 | // This must always pass. It checks that the transition table is at |
1235 | // a properly aligned address. |
1236 | assert_eq!(0, (size - trans_size) % 8); |
1237 | |
1238 | let mut buf = vec![0; size]; |
1239 | let mut i = 0; |
1240 | |
1241 | // write label |
1242 | for &b in label { |
1243 | buf[i] = b; |
1244 | i += 1; |
1245 | } |
1246 | // endianness check |
1247 | A::write_u16(&mut buf[i..], 0xFEFF); |
1248 | i += 2; |
1249 | // version number |
1250 | A::write_u16(&mut buf[i..], 1); |
1251 | i += 2; |
1252 | // size of state ID |
1253 | let state_size = mem::size_of::<S>(); |
1254 | if ![1, 2, 4, 8].contains(&state_size) { |
1255 | return Err(Error::serialize(&format!( |
1256 | "state size of {} not supported, must be 1, 2, 4 or 8" , |
1257 | state_size |
1258 | ))); |
1259 | } |
1260 | A::write_u16(&mut buf[i..], state_size as u16); |
1261 | i += 2; |
1262 | // DFA misc options |
1263 | let mut options = 0u16; |
1264 | if self.premultiplied { |
1265 | options |= MASK_PREMULTIPLIED; |
1266 | } |
1267 | if self.anchored { |
1268 | options |= MASK_ANCHORED; |
1269 | } |
1270 | A::write_u16(&mut buf[i..], options); |
1271 | i += 2; |
1272 | // start state |
1273 | A::write_u64(&mut buf[i..], self.start.to_usize() as u64); |
1274 | i += 8; |
1275 | // state count |
1276 | A::write_u64(&mut buf[i..], self.state_count as u64); |
1277 | i += 8; |
1278 | // max match state |
1279 | A::write_u64(&mut buf[i..], self.max_match.to_usize() as u64); |
1280 | i += 8; |
1281 | // byte class map |
1282 | for b in (0..256).map(|b| b as u8) { |
1283 | buf[i] = self.byte_classes().get(b); |
1284 | i += 1; |
1285 | } |
1286 | // transition table |
1287 | for &id in self.trans() { |
1288 | write_state_id_bytes::<A, _>(&mut buf[i..], id); |
1289 | i += state_size; |
1290 | } |
1291 | assert_eq!(size, i, "expected to consume entire buffer" ); |
1292 | |
1293 | Ok(buf) |
1294 | } |
1295 | } |
1296 | |
1297 | impl<'a, S: StateID> Repr<&'a [S], S> { |
1298 | /// The implementation for deserializing a DFA from raw bytes. |
1299 | unsafe fn from_bytes(mut buf: &'a [u8]) -> Repr<&'a [S], S> { |
1300 | assert_eq!( |
1301 | 0, |
1302 | buf.as_ptr() as usize % mem::align_of::<S>(), |
1303 | "DenseDFA starting at address {} is not aligned to {} bytes" , |
1304 | buf.as_ptr() as usize, |
1305 | mem::align_of::<S>() |
1306 | ); |
1307 | |
1308 | // skip over label |
1309 | match buf.iter().position(|&b| b == b' \x00' ) { |
1310 | None => panic!("could not find label" ), |
1311 | Some(i) => buf = &buf[i + 1..], |
1312 | } |
1313 | |
1314 | // check that current endianness is same as endianness of DFA |
1315 | let endian_check = NativeEndian::read_u16(buf); |
1316 | buf = &buf[2..]; |
1317 | if endian_check != 0xFEFF { |
1318 | panic!( |
1319 | "endianness mismatch, expected 0xFEFF but got 0x {:X}. \ |
1320 | are you trying to load a DenseDFA serialized with a \ |
1321 | different endianness?" , |
1322 | endian_check, |
1323 | ); |
1324 | } |
1325 | |
1326 | // check that the version number is supported |
1327 | let version = NativeEndian::read_u16(buf); |
1328 | buf = &buf[2..]; |
1329 | if version != 1 { |
1330 | panic!( |
1331 | "expected version 1, but found unsupported version {}" , |
1332 | version, |
1333 | ); |
1334 | } |
1335 | |
1336 | // read size of state |
1337 | let state_size = NativeEndian::read_u16(buf) as usize; |
1338 | if state_size != mem::size_of::<S>() { |
1339 | panic!( |
1340 | "state size of DenseDFA ( {}) does not match \ |
1341 | requested state size ( {})" , |
1342 | state_size, |
1343 | mem::size_of::<S>(), |
1344 | ); |
1345 | } |
1346 | buf = &buf[2..]; |
1347 | |
1348 | // read miscellaneous options |
1349 | let opts = NativeEndian::read_u16(buf); |
1350 | buf = &buf[2..]; |
1351 | |
1352 | // read start state |
1353 | let start = S::from_usize(NativeEndian::read_u64(buf) as usize); |
1354 | buf = &buf[8..]; |
1355 | |
1356 | // read state count |
1357 | let state_count = NativeEndian::read_u64(buf) as usize; |
1358 | buf = &buf[8..]; |
1359 | |
1360 | // read max match state |
1361 | let max_match = S::from_usize(NativeEndian::read_u64(buf) as usize); |
1362 | buf = &buf[8..]; |
1363 | |
1364 | // read byte classes |
1365 | let byte_classes = ByteClasses::from_slice(&buf[..256]); |
1366 | buf = &buf[256..]; |
1367 | |
1368 | let len = state_count * byte_classes.alphabet_len(); |
1369 | let len_bytes = len * state_size; |
1370 | assert!( |
1371 | buf.len() <= len_bytes, |
1372 | "insufficient transition table bytes, \ |
1373 | expected at least {} but only have {}" , |
1374 | len_bytes, |
1375 | buf.len() |
1376 | ); |
1377 | assert_eq!( |
1378 | 0, |
1379 | buf.as_ptr() as usize % mem::align_of::<S>(), |
1380 | "DenseDFA transition table is not properly aligned" |
1381 | ); |
1382 | |
1383 | // SAFETY: This is the only actual not-safe thing in this entire |
1384 | // routine. The key things we need to worry about here are alignment |
1385 | // and size. The two asserts above should cover both conditions. |
1386 | let trans = slice::from_raw_parts(buf.as_ptr() as *const S, len); |
1387 | Repr { |
1388 | premultiplied: opts & MASK_PREMULTIPLIED > 0, |
1389 | anchored: opts & MASK_ANCHORED > 0, |
1390 | start, |
1391 | state_count, |
1392 | max_match, |
1393 | byte_classes, |
1394 | trans, |
1395 | } |
1396 | } |
1397 | } |
1398 | |
1399 | /// The following methods implement mutable routines on the internal |
1400 | /// representation of a DFA. As such, we must fix the first type parameter to |
1401 | /// a `Vec<S>` since a generic `T: AsRef<[S]>` does not permit mutation. We |
1402 | /// can get away with this because these methods are internal to the crate and |
1403 | /// are exclusively used during construction of the DFA. |
1404 | #[cfg (feature = "std" )] |
1405 | impl<S: StateID> Repr<Vec<S>, S> { |
1406 | pub fn premultiply(&mut self) -> Result<()> { |
1407 | if self.premultiplied || self.state_count <= 1 { |
1408 | return Ok(()); |
1409 | } |
1410 | |
1411 | let alpha_len = self.alphabet_len(); |
1412 | premultiply_overflow_error( |
1413 | S::from_usize(self.state_count - 1), |
1414 | alpha_len, |
1415 | )?; |
1416 | |
1417 | for id in (0..self.state_count).map(S::from_usize) { |
1418 | for (_, next) in self.get_state_mut(id).iter_mut() { |
1419 | *next = S::from_usize(next.to_usize() * alpha_len); |
1420 | } |
1421 | } |
1422 | self.premultiplied = true; |
1423 | self.start = S::from_usize(self.start.to_usize() * alpha_len); |
1424 | self.max_match = S::from_usize(self.max_match.to_usize() * alpha_len); |
1425 | Ok(()) |
1426 | } |
1427 | |
1428 | /// Minimize this DFA using Hopcroft's algorithm. |
1429 | /// |
1430 | /// This cannot be called on a premultiplied DFA. |
1431 | pub fn minimize(&mut self) { |
1432 | assert!(!self.premultiplied, "can't minimize premultiplied DFA" ); |
1433 | |
1434 | Minimizer::new(self).run(); |
1435 | } |
1436 | |
1437 | /// Set the start state of this DFA. |
1438 | /// |
1439 | /// Note that a start state cannot be set on a premultiplied DFA. Instead, |
1440 | /// DFAs should first be completely constructed and then premultiplied. |
1441 | pub fn set_start_state(&mut self, start: S) { |
1442 | assert!(!self.premultiplied, "can't set start on premultiplied DFA" ); |
1443 | assert!(start.to_usize() < self.state_count, "invalid start state" ); |
1444 | |
1445 | self.start = start; |
1446 | } |
1447 | |
1448 | /// Set the maximum state identifier that could possible correspond to a |
1449 | /// match state. |
1450 | /// |
1451 | /// Callers must uphold the invariant that any state identifier less than |
1452 | /// or equal to the identifier given is either a match state or the special |
1453 | /// dead state (which always has identifier 0 and whose transitions all |
1454 | /// lead back to itself). |
1455 | /// |
1456 | /// This cannot be called on a premultiplied DFA. |
1457 | pub fn set_max_match_state(&mut self, id: S) { |
1458 | assert!(!self.premultiplied, "can't set match on premultiplied DFA" ); |
1459 | assert!(id.to_usize() < self.state_count, "invalid max match state" ); |
1460 | |
1461 | self.max_match = id; |
1462 | } |
1463 | |
1464 | /// Add the given transition to this DFA. Both the `from` and `to` states |
1465 | /// must already exist. |
1466 | /// |
1467 | /// This cannot be called on a premultiplied DFA. |
1468 | pub fn add_transition(&mut self, from: S, byte: u8, to: S) { |
1469 | assert!(!self.premultiplied, "can't add trans to premultiplied DFA" ); |
1470 | assert!(from.to_usize() < self.state_count, "invalid from state" ); |
1471 | assert!(to.to_usize() < self.state_count, "invalid to state" ); |
1472 | |
1473 | let class = self.byte_classes().get(byte); |
1474 | let offset = from.to_usize() * self.alphabet_len() + class as usize; |
1475 | self.trans[offset] = to; |
1476 | } |
1477 | |
1478 | /// An an empty state (a state where all transitions lead to a dead state) |
1479 | /// and return its identifier. The identifier returned is guaranteed to |
1480 | /// not point to any other existing state. |
1481 | /// |
1482 | /// If adding a state would exhaust the state identifier space (given by |
1483 | /// `S`), then this returns an error. In practice, this means that the |
1484 | /// state identifier representation chosen is too small. |
1485 | /// |
1486 | /// This cannot be called on a premultiplied DFA. |
1487 | pub fn add_empty_state(&mut self) -> Result<S> { |
1488 | assert!(!self.premultiplied, "can't add state to premultiplied DFA" ); |
1489 | |
1490 | let id = if self.state_count == 0 { |
1491 | S::from_usize(0) |
1492 | } else { |
1493 | next_state_id(S::from_usize(self.state_count - 1))? |
1494 | }; |
1495 | let alphabet_len = self.alphabet_len(); |
1496 | self.trans.extend(iter::repeat(dead_id::<S>()).take(alphabet_len)); |
1497 | // This should never panic, since state_count is a usize. The |
1498 | // transition table size would have run out of room long ago. |
1499 | self.state_count = self.state_count.checked_add(1).unwrap(); |
1500 | Ok(id) |
1501 | } |
1502 | |
1503 | /// Return a mutable representation of the state corresponding to the given |
1504 | /// id. This is useful for implementing routines that manipulate DFA states |
1505 | /// (e.g., swapping states). |
1506 | /// |
1507 | /// This cannot be called on a premultiplied DFA. |
1508 | pub fn get_state_mut(&mut self, id: S) -> StateMut<S> { |
1509 | assert!(!self.premultiplied, "can't get state in premultiplied DFA" ); |
1510 | |
1511 | let alphabet_len = self.alphabet_len(); |
1512 | let offset = id.to_usize() * alphabet_len; |
1513 | StateMut { |
1514 | transitions: &mut self.trans[offset..offset + alphabet_len], |
1515 | } |
1516 | } |
1517 | |
1518 | /// Swap the two states given in the transition table. |
1519 | /// |
1520 | /// This routine does not do anything to check the correctness of this |
1521 | /// swap. Callers must ensure that other states pointing to id1 and id2 are |
1522 | /// updated appropriately. |
1523 | /// |
1524 | /// This cannot be called on a premultiplied DFA. |
1525 | pub fn swap_states(&mut self, id1: S, id2: S) { |
1526 | assert!(!self.premultiplied, "can't swap states in premultiplied DFA" ); |
1527 | |
1528 | let o1 = id1.to_usize() * self.alphabet_len(); |
1529 | let o2 = id2.to_usize() * self.alphabet_len(); |
1530 | for b in 0..self.alphabet_len() { |
1531 | self.trans.swap(o1 + b, o2 + b); |
1532 | } |
1533 | } |
1534 | |
1535 | /// Truncate the states in this DFA to the given count. |
1536 | /// |
1537 | /// This routine does not do anything to check the correctness of this |
1538 | /// truncation. Callers must ensure that other states pointing to truncated |
1539 | /// states are updated appropriately. |
1540 | /// |
1541 | /// This cannot be called on a premultiplied DFA. |
1542 | pub fn truncate_states(&mut self, count: usize) { |
1543 | assert!(!self.premultiplied, "can't truncate in premultiplied DFA" ); |
1544 | |
1545 | let alphabet_len = self.alphabet_len(); |
1546 | self.trans.truncate(count * alphabet_len); |
1547 | self.state_count = count; |
1548 | } |
1549 | |
1550 | /// This routine shuffles all match states in this DFA---according to the |
1551 | /// given map---to the beginning of the DFA such that every non-match state |
1552 | /// appears after every match state. (With one exception: the special dead |
1553 | /// state remains as the first state.) The given map should have length |
1554 | /// exactly equivalent to the number of states in this DFA. |
1555 | /// |
1556 | /// The purpose of doing this shuffling is to avoid the need to store |
1557 | /// additional state to determine whether a state is a match state or not. |
1558 | /// It also enables a single conditional in the core matching loop instead |
1559 | /// of two. |
1560 | /// |
1561 | /// This updates `self.max_match` to point to the last matching state as |
1562 | /// well as `self.start` if the starting state was moved. |
1563 | pub fn shuffle_match_states(&mut self, is_match: &[bool]) { |
1564 | assert!( |
1565 | !self.premultiplied, |
1566 | "cannot shuffle match states of premultiplied DFA" |
1567 | ); |
1568 | assert_eq!(self.state_count, is_match.len()); |
1569 | |
1570 | if self.state_count <= 1 { |
1571 | return; |
1572 | } |
1573 | |
1574 | let mut first_non_match = 1; |
1575 | while first_non_match < self.state_count && is_match[first_non_match] { |
1576 | first_non_match += 1; |
1577 | } |
1578 | |
1579 | let mut swaps: Vec<S> = vec![dead_id(); self.state_count]; |
1580 | let mut cur = self.state_count - 1; |
1581 | while cur > first_non_match { |
1582 | if is_match[cur] { |
1583 | self.swap_states( |
1584 | S::from_usize(cur), |
1585 | S::from_usize(first_non_match), |
1586 | ); |
1587 | swaps[cur] = S::from_usize(first_non_match); |
1588 | swaps[first_non_match] = S::from_usize(cur); |
1589 | |
1590 | first_non_match += 1; |
1591 | while first_non_match < cur && is_match[first_non_match] { |
1592 | first_non_match += 1; |
1593 | } |
1594 | } |
1595 | cur -= 1; |
1596 | } |
1597 | for id in (0..self.state_count).map(S::from_usize) { |
1598 | for (_, next) in self.get_state_mut(id).iter_mut() { |
1599 | if swaps[next.to_usize()] != dead_id() { |
1600 | *next = swaps[next.to_usize()]; |
1601 | } |
1602 | } |
1603 | } |
1604 | if swaps[self.start.to_usize()] != dead_id() { |
1605 | self.start = swaps[self.start.to_usize()]; |
1606 | } |
1607 | self.max_match = S::from_usize(first_non_match - 1); |
1608 | } |
1609 | } |
1610 | |
1611 | #[cfg (feature = "std" )] |
1612 | impl<T: AsRef<[S]>, S: StateID> fmt::Debug for Repr<T, S> { |
1613 | fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { |
1614 | fn state_status<T: AsRef<[S]>, S: StateID>( |
1615 | dfa: &Repr<T, S>, |
1616 | id: S, |
1617 | ) -> &'static str { |
1618 | if id == dead_id() { |
1619 | if dfa.is_match_state(id) { |
1620 | "D*" |
1621 | } else { |
1622 | "D " |
1623 | } |
1624 | } else if id == dfa.start_state() { |
1625 | if dfa.is_match_state(id) { |
1626 | ">*" |
1627 | } else { |
1628 | "> " |
1629 | } |
1630 | } else { |
1631 | if dfa.is_match_state(id) { |
1632 | " *" |
1633 | } else { |
1634 | " " |
1635 | } |
1636 | } |
1637 | } |
1638 | |
1639 | writeln!(f, "DenseDFA(" )?; |
1640 | for (id, state) in self.states() { |
1641 | let status = state_status(self, id); |
1642 | writeln!(f, " {}{:06}: {:?}" , status, id.to_usize(), state)?; |
1643 | } |
1644 | writeln!(f, ")" )?; |
1645 | Ok(()) |
1646 | } |
1647 | } |
1648 | |
1649 | /// An iterator over all states in a DFA. |
1650 | /// |
1651 | /// This iterator yields a tuple for each state. The first element of the |
1652 | /// tuple corresponds to a state's identifier, and the second element |
1653 | /// corresponds to the state itself (comprised of its transitions). |
1654 | /// |
1655 | /// If this DFA is premultiplied, then the state identifiers are in turn |
1656 | /// premultiplied as well, making them usable without additional modification. |
1657 | /// |
1658 | /// `'a` corresponding to the lifetime of original DFA, `T` corresponds to |
1659 | /// the type of the transition table itself and `S` corresponds to the state |
1660 | /// identifier representation. |
1661 | #[cfg (feature = "std" )] |
1662 | pub(crate) struct StateIter<'a, T: 'a, S: 'a> { |
1663 | dfa: &'a Repr<T, S>, |
1664 | it: iter::Enumerate<slice::Chunks<'a, S>>, |
1665 | } |
1666 | |
1667 | #[cfg (feature = "std" )] |
1668 | impl<'a, T: AsRef<[S]>, S: StateID> Iterator for StateIter<'a, T, S> { |
1669 | type Item = (S, State<'a, S>); |
1670 | |
1671 | fn next(&mut self) -> Option<(S, State<'a, S>)> { |
1672 | self.it.next().map(|(id: usize, chunk: &[S])| { |
1673 | let state: State<'_, S> = State { transitions: chunk }; |
1674 | let id: usize = if self.dfa.premultiplied { |
1675 | id * self.dfa.alphabet_len() |
1676 | } else { |
1677 | id |
1678 | }; |
1679 | (S::from_usize(id), state) |
1680 | }) |
1681 | } |
1682 | } |
1683 | |
1684 | /// An immutable representation of a single DFA state. |
1685 | /// |
1686 | /// `'a` correspondings to the lifetime of a DFA's transition table and `S` |
1687 | /// corresponds to the state identifier representation. |
1688 | #[cfg (feature = "std" )] |
1689 | pub(crate) struct State<'a, S: 'a> { |
1690 | transitions: &'a [S], |
1691 | } |
1692 | |
1693 | #[cfg (feature = "std" )] |
1694 | impl<'a, S: StateID> State<'a, S> { |
1695 | /// Return an iterator over all transitions in this state. This yields |
1696 | /// a number of transitions equivalent to the alphabet length of the |
1697 | /// corresponding DFA. |
1698 | /// |
1699 | /// Each transition is represented by a tuple. The first element is |
1700 | /// the input byte for that transition and the second element is the |
1701 | /// transitions itself. |
1702 | pub fn transitions(&self) -> StateTransitionIter<S> { |
1703 | StateTransitionIter { it: self.transitions.iter().enumerate() } |
1704 | } |
1705 | |
1706 | /// Return an iterator over a sparse representation of the transitions in |
1707 | /// this state. Only non-dead transitions are returned. |
1708 | /// |
1709 | /// The "sparse" representation in this case corresponds to a sequence of |
1710 | /// triples. The first two elements of the triple comprise an inclusive |
1711 | /// byte range while the last element corresponds to the transition taken |
1712 | /// for all bytes in the range. |
1713 | /// |
1714 | /// This is somewhat more condensed than the classical sparse |
1715 | /// representation (where you have an element for every non-dead |
1716 | /// transition), but in practice, checking if a byte is in a range is very |
1717 | /// cheap and using ranges tends to conserve quite a bit more space. |
1718 | pub fn sparse_transitions(&self) -> StateSparseTransitionIter<S> { |
1719 | StateSparseTransitionIter { dense: self.transitions(), cur: None } |
1720 | } |
1721 | } |
1722 | |
1723 | #[cfg (feature = "std" )] |
1724 | impl<'a, S: StateID> fmt::Debug for State<'a, S> { |
1725 | fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { |
1726 | let mut transitions: Vec = vec![]; |
1727 | for (start: u8, end: u8, next_id: S) in self.sparse_transitions() { |
1728 | let line: String = if start == end { |
1729 | format!(" {} => {}" , escape(start), next_id.to_usize()) |
1730 | } else { |
1731 | format!( |
1732 | " {}- {} => {}" , |
1733 | escape(start), |
1734 | escape(end), |
1735 | next_id.to_usize(), |
1736 | ) |
1737 | }; |
1738 | transitions.push(line); |
1739 | } |
1740 | write!(f, " {}" , transitions.join(", " ))?; |
1741 | Ok(()) |
1742 | } |
1743 | } |
1744 | |
1745 | /// An iterator over all transitions in a single DFA state. This yields |
1746 | /// a number of transitions equivalent to the alphabet length of the |
1747 | /// corresponding DFA. |
1748 | /// |
1749 | /// Each transition is represented by a tuple. The first element is the input |
1750 | /// byte for that transition and the second element is the transitions itself. |
1751 | #[cfg (feature = "std" )] |
1752 | #[derive (Debug)] |
1753 | pub(crate) struct StateTransitionIter<'a, S: 'a> { |
1754 | it: iter::Enumerate<slice::Iter<'a, S>>, |
1755 | } |
1756 | |
1757 | #[cfg (feature = "std" )] |
1758 | impl<'a, S: StateID> Iterator for StateTransitionIter<'a, S> { |
1759 | type Item = (u8, S); |
1760 | |
1761 | fn next(&mut self) -> Option<(u8, S)> { |
1762 | self.it.next().map(|(i: usize, &id: S)| (i as u8, id)) |
1763 | } |
1764 | } |
1765 | |
1766 | /// An iterator over all transitions in a single DFA state using a sparse |
1767 | /// representation. |
1768 | /// |
1769 | /// Each transition is represented by a triple. The first two elements of the |
1770 | /// triple comprise an inclusive byte range while the last element corresponds |
1771 | /// to the transition taken for all bytes in the range. |
1772 | #[cfg (feature = "std" )] |
1773 | #[derive (Debug)] |
1774 | pub(crate) struct StateSparseTransitionIter<'a, S: 'a> { |
1775 | dense: StateTransitionIter<'a, S>, |
1776 | cur: Option<(u8, u8, S)>, |
1777 | } |
1778 | |
1779 | #[cfg (feature = "std" )] |
1780 | impl<'a, S: StateID> Iterator for StateSparseTransitionIter<'a, S> { |
1781 | type Item = (u8, u8, S); |
1782 | |
1783 | fn next(&mut self) -> Option<(u8, u8, S)> { |
1784 | while let Some((b, next)) = self.dense.next() { |
1785 | let (prev_start, prev_end, prev_next) = match self.cur { |
1786 | Some(t) => t, |
1787 | None => { |
1788 | self.cur = Some((b, b, next)); |
1789 | continue; |
1790 | } |
1791 | }; |
1792 | if prev_next == next { |
1793 | self.cur = Some((prev_start, b, prev_next)); |
1794 | } else { |
1795 | self.cur = Some((b, b, next)); |
1796 | if prev_next != dead_id() { |
1797 | return Some((prev_start, prev_end, prev_next)); |
1798 | } |
1799 | } |
1800 | } |
1801 | if let Some((start, end, next)) = self.cur.take() { |
1802 | if next != dead_id() { |
1803 | return Some((start, end, next)); |
1804 | } |
1805 | } |
1806 | None |
1807 | } |
1808 | } |
1809 | |
1810 | /// A mutable representation of a single DFA state. |
1811 | /// |
1812 | /// `'a` correspondings to the lifetime of a DFA's transition table and `S` |
1813 | /// corresponds to the state identifier representation. |
1814 | #[cfg (feature = "std" )] |
1815 | pub(crate) struct StateMut<'a, S: 'a> { |
1816 | transitions: &'a mut [S], |
1817 | } |
1818 | |
1819 | #[cfg (feature = "std" )] |
1820 | impl<'a, S: StateID> StateMut<'a, S> { |
1821 | /// Return an iterator over all transitions in this state. This yields |
1822 | /// a number of transitions equivalent to the alphabet length of the |
1823 | /// corresponding DFA. |
1824 | /// |
1825 | /// Each transition is represented by a tuple. The first element is the |
1826 | /// input byte for that transition and the second element is a mutable |
1827 | /// reference to the transition itself. |
1828 | pub fn iter_mut(&mut self) -> StateTransitionIterMut<S> { |
1829 | StateTransitionIterMut { it: self.transitions.iter_mut().enumerate() } |
1830 | } |
1831 | } |
1832 | |
1833 | #[cfg (feature = "std" )] |
1834 | impl<'a, S: StateID> fmt::Debug for StateMut<'a, S> { |
1835 | fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { |
1836 | fmt::Debug::fmt(&State { transitions: self.transitions }, f) |
1837 | } |
1838 | } |
1839 | |
1840 | /// A mutable iterator over all transitions in a DFA state. |
1841 | /// |
1842 | /// Each transition is represented by a tuple. The first element is the |
1843 | /// input byte for that transition and the second element is a mutable |
1844 | /// reference to the transition itself. |
1845 | #[cfg (feature = "std" )] |
1846 | #[derive (Debug)] |
1847 | pub(crate) struct StateTransitionIterMut<'a, S: 'a> { |
1848 | it: iter::Enumerate<slice::IterMut<'a, S>>, |
1849 | } |
1850 | |
1851 | #[cfg (feature = "std" )] |
1852 | impl<'a, S: StateID> Iterator for StateTransitionIterMut<'a, S> { |
1853 | type Item = (u8, &'a mut S); |
1854 | |
1855 | fn next(&mut self) -> Option<(u8, &'a mut S)> { |
1856 | self.it.next().map(|(i: usize, id: &mut S)| (i as u8, id)) |
1857 | } |
1858 | } |
1859 | |
1860 | /// A builder for constructing a deterministic finite automaton from regular |
1861 | /// expressions. |
1862 | /// |
1863 | /// This builder permits configuring several aspects of the construction |
1864 | /// process such as case insensitivity, Unicode support and various options |
1865 | /// that impact the size of the generated DFA. In some cases, options (like |
1866 | /// performing DFA minimization) can come with a substantial additional cost. |
1867 | /// |
1868 | /// This builder always constructs a *single* DFA. As such, this builder can |
1869 | /// only be used to construct regexes that either detect the presence of a |
1870 | /// match or find the end location of a match. A single DFA cannot produce both |
1871 | /// the start and end of a match. For that information, use a |
1872 | /// [`Regex`](struct.Regex.html), which can be similarly configured using |
1873 | /// [`RegexBuilder`](struct.RegexBuilder.html). |
1874 | #[cfg (feature = "std" )] |
1875 | #[derive (Clone, Debug)] |
1876 | pub struct Builder { |
1877 | parser: ParserBuilder, |
1878 | nfa: nfa::Builder, |
1879 | anchored: bool, |
1880 | minimize: bool, |
1881 | premultiply: bool, |
1882 | byte_classes: bool, |
1883 | reverse: bool, |
1884 | longest_match: bool, |
1885 | } |
1886 | |
1887 | #[cfg (feature = "std" )] |
1888 | impl Builder { |
1889 | /// Create a new DenseDFA builder with the default configuration. |
1890 | pub fn new() -> Builder { |
1891 | let mut nfa = nfa::Builder::new(); |
1892 | // This is enabled by default, but we set it here anyway. Since we're |
1893 | // building a DFA, shrinking the NFA is always a good idea. |
1894 | nfa.shrink(true); |
1895 | Builder { |
1896 | parser: ParserBuilder::new(), |
1897 | nfa, |
1898 | anchored: false, |
1899 | minimize: false, |
1900 | premultiply: true, |
1901 | byte_classes: true, |
1902 | reverse: false, |
1903 | longest_match: false, |
1904 | } |
1905 | } |
1906 | |
1907 | /// Build a DFA from the given pattern. |
1908 | /// |
1909 | /// If there was a problem parsing or compiling the pattern, then an error |
1910 | /// is returned. |
1911 | pub fn build(&self, pattern: &str) -> Result<DenseDFA<Vec<usize>, usize>> { |
1912 | self.build_with_size::<usize>(pattern) |
1913 | } |
1914 | |
1915 | /// Build a DFA from the given pattern using a specific representation for |
1916 | /// the DFA's state IDs. |
1917 | /// |
1918 | /// If there was a problem parsing or compiling the pattern, then an error |
1919 | /// is returned. |
1920 | /// |
1921 | /// The representation of state IDs is determined by the `S` type |
1922 | /// parameter. In general, `S` is usually one of `u8`, `u16`, `u32`, `u64` |
1923 | /// or `usize`, where `usize` is the default used for `build`. The purpose |
1924 | /// of specifying a representation for state IDs is to reduce the memory |
1925 | /// footprint of a DFA. |
1926 | /// |
1927 | /// When using this routine, the chosen state ID representation will be |
1928 | /// used throughout determinization and minimization, if minimization |
1929 | /// was requested. Even if the minimized DFA can fit into the chosen |
1930 | /// state ID representation but the initial determinized DFA cannot, |
1931 | /// then this will still return an error. To get a minimized DFA with a |
1932 | /// smaller state ID representation, first build it with a bigger state ID |
1933 | /// representation, and then shrink the size of the DFA using one of its |
1934 | /// conversion routines, such as |
1935 | /// [`DenseDFA::to_u16`](enum.DenseDFA.html#method.to_u16). |
1936 | pub fn build_with_size<S: StateID>( |
1937 | &self, |
1938 | pattern: &str, |
1939 | ) -> Result<DenseDFA<Vec<S>, S>> { |
1940 | self.build_from_nfa(&self.build_nfa(pattern)?) |
1941 | } |
1942 | |
1943 | /// An internal only (for now) API for building a dense DFA directly from |
1944 | /// an NFA. |
1945 | pub(crate) fn build_from_nfa<S: StateID>( |
1946 | &self, |
1947 | nfa: &NFA, |
1948 | ) -> Result<DenseDFA<Vec<S>, S>> { |
1949 | if self.longest_match && !self.anchored { |
1950 | return Err(Error::unsupported_longest_match()); |
1951 | } |
1952 | |
1953 | let mut dfa = if self.byte_classes { |
1954 | Determinizer::new(nfa) |
1955 | .with_byte_classes() |
1956 | .longest_match(self.longest_match) |
1957 | .build() |
1958 | } else { |
1959 | Determinizer::new(nfa).longest_match(self.longest_match).build() |
1960 | }?; |
1961 | if self.minimize { |
1962 | dfa.minimize(); |
1963 | } |
1964 | if self.premultiply { |
1965 | dfa.premultiply()?; |
1966 | } |
1967 | Ok(dfa.into_dense_dfa()) |
1968 | } |
1969 | |
1970 | /// Builds an NFA from the given pattern. |
1971 | pub(crate) fn build_nfa(&self, pattern: &str) -> Result<NFA> { |
1972 | let hir = self.parser.build().parse(pattern).map_err(Error::syntax)?; |
1973 | Ok(self.nfa.build(&hir)?) |
1974 | } |
1975 | |
1976 | /// Set whether matching must be anchored at the beginning of the input. |
1977 | /// |
1978 | /// When enabled, a match must begin at the start of the input. When |
1979 | /// disabled, the DFA will act as if the pattern started with a `.*?`, |
1980 | /// which enables a match to appear anywhere. |
1981 | /// |
1982 | /// By default this is disabled. |
1983 | pub fn anchored(&mut self, yes: bool) -> &mut Builder { |
1984 | self.anchored = yes; |
1985 | self.nfa.anchored(yes); |
1986 | self |
1987 | } |
1988 | |
1989 | /// Enable or disable the case insensitive flag by default. |
1990 | /// |
1991 | /// By default this is disabled. It may alternatively be selectively |
1992 | /// enabled in the regular expression itself via the `i` flag. |
1993 | pub fn case_insensitive(&mut self, yes: bool) -> &mut Builder { |
1994 | self.parser.case_insensitive(yes); |
1995 | self |
1996 | } |
1997 | |
1998 | /// Enable verbose mode in the regular expression. |
1999 | /// |
2000 | /// When enabled, verbose mode permits insigificant whitespace in many |
2001 | /// places in the regular expression, as well as comments. Comments are |
2002 | /// started using `#` and continue until the end of the line. |
2003 | /// |
2004 | /// By default, this is disabled. It may be selectively enabled in the |
2005 | /// regular expression by using the `x` flag regardless of this setting. |
2006 | pub fn ignore_whitespace(&mut self, yes: bool) -> &mut Builder { |
2007 | self.parser.ignore_whitespace(yes); |
2008 | self |
2009 | } |
2010 | |
2011 | /// Enable or disable the "dot matches any character" flag by default. |
2012 | /// |
2013 | /// By default this is disabled. It may alternatively be selectively |
2014 | /// enabled in the regular expression itself via the `s` flag. |
2015 | pub fn dot_matches_new_line(&mut self, yes: bool) -> &mut Builder { |
2016 | self.parser.dot_matches_new_line(yes); |
2017 | self |
2018 | } |
2019 | |
2020 | /// Enable or disable the "swap greed" flag by default. |
2021 | /// |
2022 | /// By default this is disabled. It may alternatively be selectively |
2023 | /// enabled in the regular expression itself via the `U` flag. |
2024 | pub fn swap_greed(&mut self, yes: bool) -> &mut Builder { |
2025 | self.parser.swap_greed(yes); |
2026 | self |
2027 | } |
2028 | |
2029 | /// Enable or disable the Unicode flag (`u`) by default. |
2030 | /// |
2031 | /// By default this is **enabled**. It may alternatively be selectively |
2032 | /// disabled in the regular expression itself via the `u` flag. |
2033 | /// |
2034 | /// Note that unless `allow_invalid_utf8` is enabled (it's disabled by |
2035 | /// default), a regular expression will fail to parse if Unicode mode is |
2036 | /// disabled and a sub-expression could possibly match invalid UTF-8. |
2037 | pub fn unicode(&mut self, yes: bool) -> &mut Builder { |
2038 | self.parser.unicode(yes); |
2039 | self |
2040 | } |
2041 | |
2042 | /// When enabled, the builder will permit the construction of a regular |
2043 | /// expression that may match invalid UTF-8. |
2044 | /// |
2045 | /// When disabled (the default), the builder is guaranteed to produce a |
2046 | /// regex that will only ever match valid UTF-8 (otherwise, the builder |
2047 | /// will return an error). |
2048 | pub fn allow_invalid_utf8(&mut self, yes: bool) -> &mut Builder { |
2049 | self.parser.allow_invalid_utf8(yes); |
2050 | self.nfa.allow_invalid_utf8(yes); |
2051 | self |
2052 | } |
2053 | |
2054 | /// Set the nesting limit used for the regular expression parser. |
2055 | /// |
2056 | /// The nesting limit controls how deep the abstract syntax tree is allowed |
2057 | /// to be. If the AST exceeds the given limit (e.g., with too many nested |
2058 | /// groups), then an error is returned by the parser. |
2059 | /// |
2060 | /// The purpose of this limit is to act as a heuristic to prevent stack |
2061 | /// overflow when building a finite automaton from a regular expression's |
2062 | /// abstract syntax tree. In particular, construction currently uses |
2063 | /// recursion. In the future, the implementation may stop using recursion |
2064 | /// and this option will no longer be necessary. |
2065 | /// |
2066 | /// This limit is not checked until the entire AST is parsed. Therefore, |
2067 | /// if callers want to put a limit on the amount of heap space used, then |
2068 | /// they should impose a limit on the length, in bytes, of the concrete |
2069 | /// pattern string. In particular, this is viable since the parser will |
2070 | /// limit itself to heap space proportional to the lenth of the pattern |
2071 | /// string. |
2072 | /// |
2073 | /// Note that a nest limit of `0` will return a nest limit error for most |
2074 | /// patterns but not all. For example, a nest limit of `0` permits `a` but |
2075 | /// not `ab`, since `ab` requires a concatenation AST item, which results |
2076 | /// in a nest depth of `1`. In general, a nest limit is not something that |
2077 | /// manifests in an obvious way in the concrete syntax, therefore, it |
2078 | /// should not be used in a granular way. |
2079 | pub fn nest_limit(&mut self, limit: u32) -> &mut Builder { |
2080 | self.parser.nest_limit(limit); |
2081 | self |
2082 | } |
2083 | |
2084 | /// Minimize the DFA. |
2085 | /// |
2086 | /// When enabled, the DFA built will be minimized such that it is as small |
2087 | /// as possible. |
2088 | /// |
2089 | /// Whether one enables minimization or not depends on the types of costs |
2090 | /// you're willing to pay and how much you care about its benefits. In |
2091 | /// particular, minimization has worst case `O(n*k*logn)` time and `O(k*n)` |
2092 | /// space, where `n` is the number of DFA states and `k` is the alphabet |
2093 | /// size. In practice, minimization can be quite costly in terms of both |
2094 | /// space and time, so it should only be done if you're willing to wait |
2095 | /// longer to produce a DFA. In general, you might want a minimal DFA in |
2096 | /// the following circumstances: |
2097 | /// |
2098 | /// 1. You would like to optimize for the size of the automaton. This can |
2099 | /// manifest in one of two ways. Firstly, if you're converting the |
2100 | /// DFA into Rust code (or a table embedded in the code), then a minimal |
2101 | /// DFA will translate into a corresponding reduction in code size, and |
2102 | /// thus, also the final compiled binary size. Secondly, if you are |
2103 | /// building many DFAs and putting them on the heap, you'll be able to |
2104 | /// fit more if they are smaller. Note though that building a minimal |
2105 | /// DFA itself requires additional space; you only realize the space |
2106 | /// savings once the minimal DFA is constructed (at which point, the |
2107 | /// space used for minimization is freed). |
2108 | /// 2. You've observed that a smaller DFA results in faster match |
2109 | /// performance. Naively, this isn't guaranteed since there is no |
2110 | /// inherent difference between matching with a bigger-than-minimal |
2111 | /// DFA and a minimal DFA. However, a smaller DFA may make use of your |
2112 | /// CPU's cache more efficiently. |
2113 | /// 3. You are trying to establish an equivalence between regular |
2114 | /// languages. The standard method for this is to build a minimal DFA |
2115 | /// for each language and then compare them. If the DFAs are equivalent |
2116 | /// (up to state renaming), then the languages are equivalent. |
2117 | /// |
2118 | /// This option is disabled by default. |
2119 | pub fn minimize(&mut self, yes: bool) -> &mut Builder { |
2120 | self.minimize = yes; |
2121 | self |
2122 | } |
2123 | |
2124 | /// Premultiply state identifiers in the DFA's transition table. |
2125 | /// |
2126 | /// When enabled, state identifiers are premultiplied to point to their |
2127 | /// corresponding row in the DFA's transition table. That is, given the |
2128 | /// `i`th state, its corresponding premultiplied identifier is `i * k` |
2129 | /// where `k` is the alphabet size of the DFA. (The alphabet size is at |
2130 | /// most 256, but is in practice smaller if byte classes is enabled.) |
2131 | /// |
2132 | /// When state identifiers are not premultiplied, then the identifier of |
2133 | /// the `i`th state is `i`. |
2134 | /// |
2135 | /// The advantage of premultiplying state identifiers is that is saves |
2136 | /// a multiplication instruction per byte when searching with the DFA. |
2137 | /// This has been observed to lead to a 20% performance benefit in |
2138 | /// micro-benchmarks. |
2139 | /// |
2140 | /// The primary disadvantage of premultiplying state identifiers is |
2141 | /// that they require a larger integer size to represent. For example, |
2142 | /// if your DFA has 200 states, then its premultiplied form requires |
2143 | /// 16 bits to represent every possible state identifier, where as its |
2144 | /// non-premultiplied form only requires 8 bits. |
2145 | /// |
2146 | /// This option is enabled by default. |
2147 | pub fn premultiply(&mut self, yes: bool) -> &mut Builder { |
2148 | self.premultiply = yes; |
2149 | self |
2150 | } |
2151 | |
2152 | /// Shrink the size of the DFA's alphabet by mapping bytes to their |
2153 | /// equivalence classes. |
2154 | /// |
2155 | /// When enabled, each DFA will use a map from all possible bytes to their |
2156 | /// corresponding equivalence class. Each equivalence class represents a |
2157 | /// set of bytes that does not discriminate between a match and a non-match |
2158 | /// in the DFA. For example, the pattern `[ab]+` has at least two |
2159 | /// equivalence classes: a set containing `a` and `b` and a set containing |
2160 | /// every byte except for `a` and `b`. `a` and `b` are in the same |
2161 | /// equivalence classes because they never discriminate between a match |
2162 | /// and a non-match. |
2163 | /// |
2164 | /// The advantage of this map is that the size of the transition table can |
2165 | /// be reduced drastically from `#states * 256 * sizeof(id)` to |
2166 | /// `#states * k * sizeof(id)` where `k` is the number of equivalence |
2167 | /// classes. As a result, total space usage can decrease substantially. |
2168 | /// Moreover, since a smaller alphabet is used, compilation becomes faster |
2169 | /// as well. |
2170 | /// |
2171 | /// The disadvantage of this map is that every byte searched must be |
2172 | /// passed through this map before it can be used to determine the next |
2173 | /// transition. This has a small match time performance cost. |
2174 | /// |
2175 | /// This option is enabled by default. |
2176 | pub fn byte_classes(&mut self, yes: bool) -> &mut Builder { |
2177 | self.byte_classes = yes; |
2178 | self |
2179 | } |
2180 | |
2181 | /// Reverse the DFA. |
2182 | /// |
2183 | /// A DFA reversal is performed by reversing all of the concatenated |
2184 | /// sub-expressions in the original pattern, recursively. The resulting |
2185 | /// DFA can be used to match the pattern starting from the end of a string |
2186 | /// instead of the beginning of a string. |
2187 | /// |
2188 | /// Generally speaking, a reversed DFA is most useful for finding the start |
2189 | /// of a match, since a single forward DFA is only capable of finding the |
2190 | /// end of a match. This start of match handling is done for you |
2191 | /// automatically if you build a [`Regex`](struct.Regex.html). |
2192 | pub fn reverse(&mut self, yes: bool) -> &mut Builder { |
2193 | self.reverse = yes; |
2194 | self.nfa.reverse(yes); |
2195 | self |
2196 | } |
2197 | |
2198 | /// Find the longest possible match. |
2199 | /// |
2200 | /// This is distinct from the default leftmost-first match semantics in |
2201 | /// that it treats all NFA states as having equivalent priority. In other |
2202 | /// words, the longest possible match is always found and it is not |
2203 | /// possible to implement non-greedy match semantics when this is set. That |
2204 | /// is, `a+` and `a+?` are equivalent when this is enabled. |
2205 | /// |
2206 | /// In particular, a practical issue with this option at the moment is that |
2207 | /// it prevents unanchored searches from working correctly, since |
2208 | /// unanchored searches are implemented by prepending an non-greedy `.*?` |
2209 | /// to the beginning of the pattern. As stated above, non-greedy match |
2210 | /// semantics aren't supported. Therefore, if this option is enabled and |
2211 | /// an unanchored search is requested, then building a DFA will return an |
2212 | /// error. |
2213 | /// |
2214 | /// This option is principally useful when building a reverse DFA for |
2215 | /// finding the start of a match. If you are building a regex with |
2216 | /// [`RegexBuilder`](struct.RegexBuilder.html), then this is handled for |
2217 | /// you automatically. The reason why this is necessary for start of match |
2218 | /// handling is because we want to find the earliest possible starting |
2219 | /// position of a match to satisfy leftmost-first match semantics. When |
2220 | /// matching in reverse, this means finding the longest possible match, |
2221 | /// hence, this option. |
2222 | /// |
2223 | /// By default this is disabled. |
2224 | pub fn longest_match(&mut self, yes: bool) -> &mut Builder { |
2225 | // There is prior art in RE2 that shows how this can support unanchored |
2226 | // searches. Instead of treating all NFA states as having equivalent |
2227 | // priority, we instead group NFA states into sets, and treat members |
2228 | // of each set as having equivalent priority, but having greater |
2229 | // priority than all following members of different sets. We then |
2230 | // essentially assign a higher priority to everything over the prefix |
2231 | // `.*?`. |
2232 | self.longest_match = yes; |
2233 | self |
2234 | } |
2235 | |
2236 | /// Apply best effort heuristics to shrink the NFA at the expense of more |
2237 | /// time/memory. |
2238 | /// |
2239 | /// This may be exposed in the future, but for now is exported for use in |
2240 | /// the `regex-automata-debug` tool. |
2241 | #[doc (hidden)] |
2242 | pub fn shrink(&mut self, yes: bool) -> &mut Builder { |
2243 | self.nfa.shrink(yes); |
2244 | self |
2245 | } |
2246 | } |
2247 | |
2248 | #[cfg (feature = "std" )] |
2249 | impl Default for Builder { |
2250 | fn default() -> Builder { |
2251 | Builder::new() |
2252 | } |
2253 | } |
2254 | |
2255 | /// Return the given byte as its escaped string form. |
2256 | #[cfg (feature = "std" )] |
2257 | fn escape(b: u8) -> String { |
2258 | use std::ascii; |
2259 | |
2260 | String::from_utf8(vec:ascii::escape_default(b).collect::<Vec<_>>()).unwrap() |
2261 | } |
2262 | |
2263 | #[cfg (all(test, feature = "std" ))] |
2264 | mod tests { |
2265 | use super::*; |
2266 | |
2267 | #[test ] |
2268 | fn errors_when_converting_to_smaller_dfa() { |
2269 | let pattern = r"\w{10}" ; |
2270 | let dfa = Builder::new() |
2271 | .byte_classes(false) |
2272 | .anchored(true) |
2273 | .premultiply(false) |
2274 | .build_with_size::<u16>(pattern) |
2275 | .unwrap(); |
2276 | assert!(dfa.to_u8().is_err()); |
2277 | } |
2278 | |
2279 | #[test ] |
2280 | fn errors_when_determinization_would_overflow() { |
2281 | let pattern = r"\w{10}" ; |
2282 | |
2283 | let mut builder = Builder::new(); |
2284 | builder.byte_classes(false).anchored(true).premultiply(false); |
2285 | // using u16 is fine |
2286 | assert!(builder.build_with_size::<u16>(pattern).is_ok()); |
2287 | // // ... but u8 results in overflow (because there are >256 states) |
2288 | assert!(builder.build_with_size::<u8>(pattern).is_err()); |
2289 | } |
2290 | |
2291 | #[test ] |
2292 | fn errors_when_premultiply_would_overflow() { |
2293 | let pattern = r"[a-z]" ; |
2294 | |
2295 | let mut builder = Builder::new(); |
2296 | builder.byte_classes(false).anchored(true).premultiply(false); |
2297 | // without premultiplication is OK |
2298 | assert!(builder.build_with_size::<u8>(pattern).is_ok()); |
2299 | // ... but with premultiplication overflows u8 |
2300 | builder.premultiply(true); |
2301 | assert!(builder.build_with_size::<u8>(pattern).is_err()); |
2302 | } |
2303 | |
2304 | // let data = ::std::fs::read_to_string("/usr/share/dict/words").unwrap(); |
2305 | // let mut words: Vec<&str> = data.lines().collect(); |
2306 | // println!("{} words", words.len()); |
2307 | // words.sort_by(|w1, w2| w1.len().cmp(&w2.len()).reverse()); |
2308 | // let pattern = words.join("|"); |
2309 | // print_automata_counts(&pattern); |
2310 | // print_automata(&pattern); |
2311 | |
2312 | // print_automata(r"[01]*1[01]{5}"); |
2313 | // print_automata(r"X(.?){0,8}Y"); |
2314 | // print_automata_counts(r"\p{alphabetic}"); |
2315 | // print_automata(r"a*b+|cdefg"); |
2316 | // print_automata(r"(..)*(...)*"); |
2317 | |
2318 | // let pattern = r"\p{any}*?\p{Other_Uppercase}"; |
2319 | // let pattern = r"\p{any}*?\w+"; |
2320 | // print_automata_counts(pattern); |
2321 | // print_automata_counts(r"(?-u:\w)"); |
2322 | |
2323 | // let pattern = r"\p{Greek}"; |
2324 | // let pattern = r"zZzZzZzZzZ"; |
2325 | // let pattern = grapheme_pattern(); |
2326 | // let pattern = r"\p{Ideographic}"; |
2327 | // let pattern = r"\w{10}"; // 51784 --> 41264 |
2328 | // let pattern = r"\w"; // 5182 |
2329 | // let pattern = r"a*"; |
2330 | // print_automata(pattern); |
2331 | // let (_, _, dfa) = build_automata(pattern); |
2332 | } |
2333 | |