• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*!
2 The DFA matching engine.
3 
4 A DFA provides faster matching because the engine is in exactly one state at
5 any point in time. In the NFA, there may be multiple active states, and
6 considerable CPU cycles are spent shuffling them around. In finite automata
7 speak, the DFA follows epsilon transitions in the regex far less than the NFA.
8 
9 A DFA is a classic trade off between time and space. The NFA is slower, but
10 its memory requirements are typically small and predictable. The DFA is faster,
11 but given the right regex and the right input, the number of states in the
12 DFA can grow exponentially. To mitigate this space problem, we do two things:
13 
14 1. We implement an *online* DFA. That is, the DFA is constructed from the NFA
15    during a search. When a new state is computed, it is stored in a cache so
16    that it may be reused. An important consequence of this implementation
17    is that states that are never reached for a particular input are never
18    computed. (This is impossible in an "offline" DFA which needs to compute
19    all possible states up front.)
20 2. If the cache gets too big, we wipe it and continue matching.
21 
22 In pathological cases, a new state can be created for every byte of input.
23 (e.g., The regex `(a|b)*a(a|b){20}` on a long sequence of a's and b's.)
24 In this case, performance regresses to slightly slower than the full NFA
25 simulation, in large part because the cache becomes useless. If the cache
26 is wiped too frequently, the DFA quits and control falls back to one of the
27 NFA simulations.
28 
29 Because of the "lazy" nature of this DFA, the inner matching loop is
30 considerably more complex than one might expect out of a DFA. A number of
31 tricks are employed to make it fast. Tread carefully.
32 
33 N.B. While this implementation is heavily commented, Russ Cox's series of
34 articles on regexes is strongly recommended: https://swtch.com/~rsc/regexp/
35 (As is the DFA implementation in RE2, which heavily influenced this
36 implementation.)
37 */
38 
39 use std::collections::HashMap;
40 use std::fmt;
41 use std::iter::repeat;
42 use std::mem;
43 use std::sync::Arc;
44 
45 use exec::ProgramCache;
46 use prog::{Inst, Program};
47 use sparse::SparseSet;
48 
49 /// Return true if and only if the given program can be executed by a DFA.
50 ///
51 /// Generally, a DFA is always possible. A pathological case where it is not
52 /// possible is if the number of NFA states exceeds `u32::MAX`, in which case,
53 /// this function will return false.
54 ///
55 /// This function will also return false if the given program has any Unicode
56 /// instructions (Char or Ranges) since the DFA operates on bytes only.
can_exec(insts: &Program) -> bool57 pub fn can_exec(insts: &Program) -> bool {
58     use prog::Inst::*;
59     // If for some reason we manage to allocate a regex program with more
60     // than i32::MAX instructions, then we can't execute the DFA because we
61     // use 32 bit instruction pointer deltas for memory savings.
62     // If i32::MAX is the largest positive delta,
63     // then -i32::MAX == i32::MIN + 1 is the largest negative delta,
64     // and we are OK to use 32 bits.
65     if insts.dfa_size_limit == 0 || insts.len() > ::std::i32::MAX as usize {
66         return false;
67     }
68     for inst in insts {
69         match *inst {
70             Char(_) | Ranges(_) => return false,
71             EmptyLook(_) | Match(_) | Save(_) | Split(_) | Bytes(_) => {}
72         }
73     }
74     true
75 }
76 
77 /// A reusable cache of DFA states.
78 ///
79 /// This cache is reused between multiple invocations of the same regex
80 /// program. (It is not shared simultaneously between threads. If there is
81 /// contention, then new caches are created.)
82 #[derive(Debug)]
83 pub struct Cache {
84     /// Group persistent DFA related cache state together. The sparse sets
85     /// listed below are used as scratch space while computing uncached states.
86     inner: CacheInner,
87     /// qcur and qnext are ordered sets with constant time
88     /// addition/membership/clearing-whole-set and linear time iteration. They
89     /// are used to manage the sets of NFA states in DFA states when computing
90     /// cached DFA states. In particular, the order of the NFA states matters
91     /// for leftmost-first style matching. Namely, when computing a cached
92     /// state, the set of NFA states stops growing as soon as the first Match
93     /// instruction is observed.
94     qcur: SparseSet,
95     qnext: SparseSet,
96 }
97 
98 /// `CacheInner` is logically just a part of Cache, but groups together fields
99 /// that aren't passed as function parameters throughout search. (This split
100 /// is mostly an artifact of the borrow checker. It is happily paid.)
101 #[derive(Debug)]
102 struct CacheInner {
103     /// A cache of pre-compiled DFA states, keyed by the set of NFA states
104     /// and the set of empty-width flags set at the byte in the input when the
105     /// state was observed.
106     ///
107     /// A StatePtr is effectively a `*State`, but to avoid various inconvenient
108     /// things, we just pass indexes around manually. The performance impact of
109     /// this is probably an instruction or two in the inner loop. However, on
110     /// 64 bit, each StatePtr is half the size of a *State.
111     compiled: StateMap,
112     /// The transition table.
113     ///
114     /// The transition table is laid out in row-major order, where states are
115     /// rows and the transitions for each state are columns. At a high level,
116     /// given state `s` and byte `b`, the next state can be found at index
117     /// `s * 256 + b`.
118     ///
119     /// This is, of course, a lie. A StatePtr is actually a pointer to the
120     /// *start* of a row in this table. When indexing in the DFA's inner loop,
121     /// this removes the need to multiply the StatePtr by the stride. Yes, it
122     /// matters. This reduces the number of states we can store, but: the
123     /// stride is rarely 256 since we define transitions in terms of
124     /// *equivalence classes* of bytes. Each class corresponds to a set of
125     /// bytes that never discriminate a distinct path through the DFA from each
126     /// other.
127     trans: Transitions,
128     /// A set of cached start states, which are limited to the number of
129     /// permutations of flags set just before the initial byte of input. (The
130     /// index into this vec is a `EmptyFlags`.)
131     ///
132     /// N.B. A start state can be "dead" (i.e., no possible match), so we
133     /// represent it with a StatePtr.
134     start_states: Vec<StatePtr>,
135     /// Stack scratch space used to follow epsilon transitions in the NFA.
136     /// (This permits us to avoid recursion.)
137     ///
138     /// The maximum stack size is the number of NFA states.
139     stack: Vec<InstPtr>,
140     /// The total number of times this cache has been flushed by the DFA
141     /// because of space constraints.
142     flush_count: u64,
143     /// The total heap size of the DFA's cache. We use this to determine when
144     /// we should flush the cache.
145     size: usize,
146     /// Scratch space used when building instruction pointer lists for new
147     /// states. This helps amortize allocation.
148     insts_scratch_space: Vec<u8>,
149 }
150 
151 /// The transition table.
152 ///
153 /// It is laid out in row-major order, with states as rows and byte class
154 /// transitions as columns.
155 ///
156 /// The transition table is responsible for producing valid `StatePtrs`. A
157 /// `StatePtr` points to the start of a particular row in this table. When
158 /// indexing to find the next state this allows us to avoid a multiplication
159 /// when computing an index into the table.
160 #[derive(Clone)]
161 struct Transitions {
162     /// The table.
163     table: Vec<StatePtr>,
164     /// The stride.
165     num_byte_classes: usize,
166 }
167 
168 /// Fsm encapsulates the actual execution of the DFA.
169 #[derive(Debug)]
170 pub struct Fsm<'a> {
171     /// prog contains the NFA instruction opcodes. DFA execution uses either
172     /// the `dfa` instructions or the `dfa_reverse` instructions from
173     /// `exec::ExecReadOnly`. (It never uses `ExecReadOnly.nfa`, which may have
174     /// Unicode opcodes that cannot be executed by the DFA.)
175     prog: &'a Program,
176     /// The start state. We record it here because the pointer may change
177     /// when the cache is wiped.
178     start: StatePtr,
179     /// The current position in the input.
180     at: usize,
181     /// Should we quit after seeing the first match? e.g., When the caller
182     /// uses `is_match` or `shortest_match`.
183     quit_after_match: bool,
184     /// The last state that matched.
185     ///
186     /// When no match has occurred, this is set to STATE_UNKNOWN.
187     ///
188     /// This is only useful when matching regex sets. The last match state
189     /// is useful because it contains all of the match instructions seen,
190     /// thereby allowing us to enumerate which regexes in the set matched.
191     last_match_si: StatePtr,
192     /// The input position of the last cache flush. We use this to determine
193     /// if we're thrashing in the cache too often. If so, the DFA quits so
194     /// that we can fall back to the NFA algorithm.
195     last_cache_flush: usize,
196     /// All cached DFA information that is persisted between searches.
197     cache: &'a mut CacheInner,
198 }
199 
200 /// The result of running the DFA.
201 ///
202 /// Generally, the result is either a match or not a match, but sometimes the
203 /// DFA runs too slowly because the cache size is too small. In that case, it
204 /// gives up with the intent of falling back to the NFA algorithm.
205 ///
206 /// The DFA can also give up if it runs out of room to create new states, or if
207 /// it sees non-ASCII bytes in the presence of a Unicode word boundary.
208 #[derive(Clone, Debug)]
209 pub enum Result<T> {
210     Match(T),
211     NoMatch(usize),
212     Quit,
213 }
214 
215 impl<T> Result<T> {
216     /// Returns true if this result corresponds to a match.
is_match(&self) -> bool217     pub fn is_match(&self) -> bool {
218         match *self {
219             Result::Match(_) => true,
220             Result::NoMatch(_) | Result::Quit => false,
221         }
222     }
223 
224     /// Maps the given function onto T and returns the result.
225     ///
226     /// If this isn't a match, then this is a no-op.
227     #[cfg(feature = "perf-literal")]
map<U, F: FnMut(T) -> U>(self, mut f: F) -> Result<U>228     pub fn map<U, F: FnMut(T) -> U>(self, mut f: F) -> Result<U> {
229         match self {
230             Result::Match(t) => Result::Match(f(t)),
231             Result::NoMatch(x) => Result::NoMatch(x),
232             Result::Quit => Result::Quit,
233         }
234     }
235 
236     /// Sets the non-match position.
237     ///
238     /// If this isn't a non-match, then this is a no-op.
set_non_match(self, at: usize) -> Result<T>239     fn set_non_match(self, at: usize) -> Result<T> {
240         match self {
241             Result::NoMatch(_) => Result::NoMatch(at),
242             r => r,
243         }
244     }
245 }
246 
247 /// `State` is a DFA state. It contains an ordered set of NFA states (not
248 /// necessarily complete) and a smattering of flags.
249 ///
250 /// The flags are packed into the first byte of data.
251 ///
252 /// States don't carry their transitions. Instead, transitions are stored in
253 /// a single row-major table.
254 ///
255 /// Delta encoding is used to store the instruction pointers.
256 /// The first instruction pointer is stored directly starting
257 /// at data[1], and each following pointer is stored as an offset
258 /// to the previous one. If a delta is in the range -127..127,
259 /// it is packed into a single byte; Otherwise the byte 128 (-128 as an i8)
260 /// is coded as a flag, followed by 4 bytes encoding the delta.
261 #[derive(Clone, Eq, Hash, PartialEq)]
262 struct State {
263     data: Arc<[u8]>,
264 }
265 
266 /// `InstPtr` is a 32 bit pointer into a sequence of opcodes (i.e., it indexes
267 /// an NFA state).
268 ///
269 /// Throughout this library, this is usually set to `usize`, but we force a
270 /// `u32` here for the DFA to save on space.
271 type InstPtr = u32;
272 
273 /// Adds ip to data using delta encoding with respect to prev.
274 ///
275 /// After completion, `data` will contain `ip` and `prev` will be set to `ip`.
push_inst_ptr(data: &mut Vec<u8>, prev: &mut InstPtr, ip: InstPtr)276 fn push_inst_ptr(data: &mut Vec<u8>, prev: &mut InstPtr, ip: InstPtr) {
277     let delta = (ip as i32) - (*prev as i32);
278     write_vari32(data, delta);
279     *prev = ip;
280 }
281 
282 struct InstPtrs<'a> {
283     base: usize,
284     data: &'a [u8],
285 }
286 
287 impl<'a> Iterator for InstPtrs<'a> {
288     type Item = usize;
289 
next(&mut self) -> Option<usize>290     fn next(&mut self) -> Option<usize> {
291         if self.data.is_empty() {
292             return None;
293         }
294         let (delta, nread) = read_vari32(self.data);
295         let base = self.base as i32 + delta;
296         debug_assert!(base >= 0);
297         debug_assert!(nread > 0);
298         self.data = &self.data[nread..];
299         self.base = base as usize;
300         Some(self.base)
301     }
302 }
303 
304 impl State {
flags(&self) -> StateFlags305     fn flags(&self) -> StateFlags {
306         StateFlags(self.data[0])
307     }
308 
inst_ptrs(&self) -> InstPtrs309     fn inst_ptrs(&self) -> InstPtrs {
310         InstPtrs { base: 0, data: &self.data[1..] }
311     }
312 }
313 
314 /// `StatePtr` is a 32 bit pointer to the start of a row in the transition
315 /// table.
316 ///
317 /// It has many special values. There are two types of special values:
318 /// sentinels and flags.
319 ///
320 /// Sentinels corresponds to special states that carry some kind of
321 /// significance. There are three such states: unknown, dead and quit states.
322 ///
323 /// Unknown states are states that haven't been computed yet. They indicate
324 /// that a transition should be filled in that points to either an existing
325 /// cached state or a new state altogether. In general, an unknown state means
326 /// "follow the NFA's epsilon transitions."
327 ///
328 /// Dead states are states that can never lead to a match, no matter what
329 /// subsequent input is observed. This means that the DFA should quit
330 /// immediately and return the longest match it has found thus far.
331 ///
332 /// Quit states are states that imply the DFA is not capable of matching the
333 /// regex correctly. Currently, this is only used when a Unicode word boundary
334 /// exists in the regex *and* a non-ASCII byte is observed.
335 ///
336 /// The other type of state pointer is a state pointer with special flag bits.
337 /// There are two flags: a start flag and a match flag. The lower bits of both
338 /// kinds always contain a "valid" `StatePtr` (indicated by the `STATE_MAX`
339 /// mask).
340 ///
341 /// The start flag means that the state is a start state, and therefore may be
342 /// subject to special prefix scanning optimizations.
343 ///
344 /// The match flag means that the state is a match state, and therefore the
345 /// current position in the input (while searching) should be recorded.
346 ///
347 /// The above exists mostly in the service of making the inner loop fast.
348 /// In particular, the inner *inner* loop looks something like this:
349 ///
350 /// ```ignore
351 /// while state <= STATE_MAX and i < len(text):
352 ///     state = state.next[i]
353 /// ```
354 ///
355 /// This is nice because it lets us execute a lazy DFA as if it were an
356 /// entirely offline DFA (i.e., with very few instructions). The loop will
357 /// quit only when we need to examine a case that needs special attention.
358 type StatePtr = u32;
359 
360 /// An unknown state means that the state has not been computed yet, and that
361 /// the only way to progress is to compute it.
362 const STATE_UNKNOWN: StatePtr = 1 << 31;
363 
364 /// A dead state means that the state has been computed and it is known that
365 /// once it is entered, no future match can ever occur.
366 const STATE_DEAD: StatePtr = STATE_UNKNOWN + 1;
367 
368 /// A quit state means that the DFA came across some input that it doesn't
369 /// know how to process correctly. The DFA should quit and another matching
370 /// engine should be run in its place.
371 const STATE_QUIT: StatePtr = STATE_DEAD + 1;
372 
373 /// A start state is a state that the DFA can start in.
374 ///
375 /// Note that start states have their lower bits set to a state pointer.
376 const STATE_START: StatePtr = 1 << 30;
377 
378 /// A match state means that the regex has successfully matched.
379 ///
380 /// Note that match states have their lower bits set to a state pointer.
381 const STATE_MATCH: StatePtr = 1 << 29;
382 
383 /// The maximum state pointer. This is useful to mask out the "valid" state
384 /// pointer from a state with the "start" or "match" bits set.
385 ///
386 /// It doesn't make sense to use this with unknown, dead or quit state
387 /// pointers, since those pointers are sentinels and never have their lower
388 /// bits set to anything meaningful.
389 const STATE_MAX: StatePtr = STATE_MATCH - 1;
390 
391 /// Byte is a u8 in spirit, but a u16 in practice so that we can represent the
392 /// special EOF sentinel value.
393 #[derive(Copy, Clone, Debug)]
394 struct Byte(u16);
395 
396 /// A set of flags for zero-width assertions.
397 #[derive(Clone, Copy, Eq, Debug, Default, Hash, PartialEq)]
398 struct EmptyFlags {
399     start: bool,
400     end: bool,
401     start_line: bool,
402     end_line: bool,
403     word_boundary: bool,
404     not_word_boundary: bool,
405 }
406 
407 /// A set of flags describing various configurations of a DFA state. This is
408 /// represented by a `u8` so that it is compact.
409 #[derive(Clone, Copy, Eq, Default, Hash, PartialEq)]
410 struct StateFlags(u8);
411 
412 impl Cache {
413     /// Create new empty cache for the DFA engine.
new(prog: &Program) -> Self414     pub fn new(prog: &Program) -> Self {
415         // We add 1 to account for the special EOF byte.
416         let num_byte_classes = (prog.byte_classes[255] as usize + 1) + 1;
417         let starts = vec![STATE_UNKNOWN; 256];
418         let mut cache = Cache {
419             inner: CacheInner {
420                 compiled: StateMap::new(num_byte_classes),
421                 trans: Transitions::new(num_byte_classes),
422                 start_states: starts,
423                 stack: vec![],
424                 flush_count: 0,
425                 size: 0,
426                 insts_scratch_space: vec![],
427             },
428             qcur: SparseSet::new(prog.insts.len()),
429             qnext: SparseSet::new(prog.insts.len()),
430         };
431         cache.inner.reset_size();
432         cache
433     }
434 }
435 
436 impl CacheInner {
437     /// Resets the cache size to account for fixed costs, such as the program
438     /// and stack sizes.
reset_size(&mut self)439     fn reset_size(&mut self) {
440         self.size = (self.start_states.len() * mem::size_of::<StatePtr>())
441             + (self.stack.len() * mem::size_of::<InstPtr>());
442     }
443 }
444 
445 impl<'a> Fsm<'a> {
446     #[cfg_attr(feature = "perf-inline", inline(always))]
forward( prog: &'a Program, cache: &ProgramCache, quit_after_match: bool, text: &[u8], at: usize, ) -> Result<usize>447     pub fn forward(
448         prog: &'a Program,
449         cache: &ProgramCache,
450         quit_after_match: bool,
451         text: &[u8],
452         at: usize,
453     ) -> Result<usize> {
454         let mut cache = cache.borrow_mut();
455         let cache = &mut cache.dfa;
456         let mut dfa = Fsm {
457             prog: prog,
458             start: 0, // filled in below
459             at: at,
460             quit_after_match: quit_after_match,
461             last_match_si: STATE_UNKNOWN,
462             last_cache_flush: at,
463             cache: &mut cache.inner,
464         };
465         let (empty_flags, state_flags) = dfa.start_flags(text, at);
466         dfa.start =
467             match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
468                 None => return Result::Quit,
469                 Some(STATE_DEAD) => return Result::NoMatch(at),
470                 Some(si) => si,
471             };
472         debug_assert!(dfa.start != STATE_UNKNOWN);
473         dfa.exec_at(&mut cache.qcur, &mut cache.qnext, text)
474     }
475 
476     #[cfg_attr(feature = "perf-inline", inline(always))]
reverse( prog: &'a Program, cache: &ProgramCache, quit_after_match: bool, text: &[u8], at: usize, ) -> Result<usize>477     pub fn reverse(
478         prog: &'a Program,
479         cache: &ProgramCache,
480         quit_after_match: bool,
481         text: &[u8],
482         at: usize,
483     ) -> Result<usize> {
484         let mut cache = cache.borrow_mut();
485         let cache = &mut cache.dfa_reverse;
486         let mut dfa = Fsm {
487             prog: prog,
488             start: 0, // filled in below
489             at: at,
490             quit_after_match: quit_after_match,
491             last_match_si: STATE_UNKNOWN,
492             last_cache_flush: at,
493             cache: &mut cache.inner,
494         };
495         let (empty_flags, state_flags) = dfa.start_flags_reverse(text, at);
496         dfa.start =
497             match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
498                 None => return Result::Quit,
499                 Some(STATE_DEAD) => return Result::NoMatch(at),
500                 Some(si) => si,
501             };
502         debug_assert!(dfa.start != STATE_UNKNOWN);
503         dfa.exec_at_reverse(&mut cache.qcur, &mut cache.qnext, text)
504     }
505 
506     #[cfg_attr(feature = "perf-inline", inline(always))]
forward_many( prog: &'a Program, cache: &ProgramCache, matches: &mut [bool], text: &[u8], at: usize, ) -> Result<usize>507     pub fn forward_many(
508         prog: &'a Program,
509         cache: &ProgramCache,
510         matches: &mut [bool],
511         text: &[u8],
512         at: usize,
513     ) -> Result<usize> {
514         debug_assert!(matches.len() == prog.matches.len());
515         let mut cache = cache.borrow_mut();
516         let cache = &mut cache.dfa;
517         let mut dfa = Fsm {
518             prog: prog,
519             start: 0, // filled in below
520             at: at,
521             quit_after_match: false,
522             last_match_si: STATE_UNKNOWN,
523             last_cache_flush: at,
524             cache: &mut cache.inner,
525         };
526         let (empty_flags, state_flags) = dfa.start_flags(text, at);
527         dfa.start =
528             match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
529                 None => return Result::Quit,
530                 Some(STATE_DEAD) => return Result::NoMatch(at),
531                 Some(si) => si,
532             };
533         debug_assert!(dfa.start != STATE_UNKNOWN);
534         let result = dfa.exec_at(&mut cache.qcur, &mut cache.qnext, text);
535         if result.is_match() {
536             if matches.len() == 1 {
537                 matches[0] = true;
538             } else {
539                 debug_assert!(dfa.last_match_si != STATE_UNKNOWN);
540                 debug_assert!(dfa.last_match_si != STATE_DEAD);
541                 for ip in dfa.state(dfa.last_match_si).inst_ptrs() {
542                     if let Inst::Match(slot) = dfa.prog[ip] {
543                         matches[slot] = true;
544                     }
545                 }
546             }
547         }
548         result
549     }
550 
551     /// Executes the DFA on a forward NFA.
552     ///
553     /// {qcur,qnext} are scratch ordered sets which may be non-empty.
554     #[cfg_attr(feature = "perf-inline", inline(always))]
exec_at( &mut self, qcur: &mut SparseSet, qnext: &mut SparseSet, text: &[u8], ) -> Result<usize>555     fn exec_at(
556         &mut self,
557         qcur: &mut SparseSet,
558         qnext: &mut SparseSet,
559         text: &[u8],
560     ) -> Result<usize> {
561         // For the most part, the DFA is basically:
562         //
563         //   last_match = null
564         //   while current_byte != EOF:
565         //     si = current_state.next[current_byte]
566         //     if si is match
567         //       last_match = si
568         //   return last_match
569         //
570         // However, we need to deal with a few things:
571         //
572         //   1. This is an *online* DFA, so the current state's next list
573         //      may not point to anywhere yet, so we must go out and compute
574         //      them. (They are then cached into the current state's next list
575         //      to avoid re-computation.)
576         //   2. If we come across a state that is known to be dead (i.e., never
577         //      leads to a match), then we can quit early.
578         //   3. If the caller just wants to know if a match occurs, then we
579         //      can quit as soon as we know we have a match. (Full leftmost
580         //      first semantics require continuing on.)
581         //   4. If we're in the start state, then we can use a pre-computed set
582         //      of prefix literals to skip quickly along the input.
583         //   5. After the input is exhausted, we run the DFA on one symbol
584         //      that stands for EOF. This is useful for handling empty width
585         //      assertions.
586         //   6. We can't actually do state.next[byte]. Instead, we have to do
587         //      state.next[byte_classes[byte]], which permits us to keep the
588         //      'next' list very small.
589         //
590         // Since there's a bunch of extra stuff we need to consider, we do some
591         // pretty hairy tricks to get the inner loop to run as fast as
592         // possible.
593         debug_assert!(!self.prog.is_reverse);
594 
595         // The last match is the currently known ending match position. It is
596         // reported as an index to the most recent byte that resulted in a
597         // transition to a match state and is always stored in capture slot `1`
598         // when searching forwards. Its maximum value is `text.len()`.
599         let mut result = Result::NoMatch(self.at);
600         let (mut prev_si, mut next_si) = (self.start, self.start);
601         let mut at = self.at;
602         while at < text.len() {
603             // This is the real inner loop. We take advantage of special bits
604             // set in the state pointer to determine whether a state is in the
605             // "common" case or not. Specifically, the common case is a
606             // non-match non-start non-dead state that has already been
607             // computed. So long as we remain in the common case, this inner
608             // loop will chew through the input.
609             //
610             // We also unroll the loop 4 times to amortize the cost of checking
611             // whether we've consumed the entire input. We are also careful
612             // to make sure that `prev_si` always represents the previous state
613             // and `next_si` always represents the next state after the loop
614             // exits, even if it isn't always true inside the loop.
615             while next_si <= STATE_MAX && at < text.len() {
616                 // Argument for safety is in the definition of next_si.
617                 prev_si = unsafe { self.next_si(next_si, text, at) };
618                 at += 1;
619                 if prev_si > STATE_MAX || at + 2 >= text.len() {
620                     mem::swap(&mut prev_si, &mut next_si);
621                     break;
622                 }
623                 next_si = unsafe { self.next_si(prev_si, text, at) };
624                 at += 1;
625                 if next_si > STATE_MAX {
626                     break;
627                 }
628                 prev_si = unsafe { self.next_si(next_si, text, at) };
629                 at += 1;
630                 if prev_si > STATE_MAX {
631                     mem::swap(&mut prev_si, &mut next_si);
632                     break;
633                 }
634                 next_si = unsafe { self.next_si(prev_si, text, at) };
635                 at += 1;
636             }
637             if next_si & STATE_MATCH > 0 {
638                 // A match state is outside of the common case because it needs
639                 // special case analysis. In particular, we need to record the
640                 // last position as having matched and possibly quit the DFA if
641                 // we don't need to keep matching.
642                 next_si &= !STATE_MATCH;
643                 result = Result::Match(at - 1);
644                 if self.quit_after_match {
645                     return result;
646                 }
647                 self.last_match_si = next_si;
648                 prev_si = next_si;
649 
650                 // This permits short-circuiting when matching a regex set.
651                 // In particular, if this DFA state contains only match states,
652                 // then it's impossible to extend the set of matches since
653                 // match states are final. Therefore, we can quit.
654                 if self.prog.matches.len() > 1 {
655                     let state = self.state(next_si);
656                     let just_matches =
657                         state.inst_ptrs().all(|ip| self.prog[ip].is_match());
658                     if just_matches {
659                         return result;
660                     }
661                 }
662 
663                 // Another inner loop! If the DFA stays in this particular
664                 // match state, then we can rip through all of the input
665                 // very quickly, and only recording the match location once
666                 // we've left this particular state.
667                 let cur = at;
668                 while (next_si & !STATE_MATCH) == prev_si
669                     && at + 2 < text.len()
670                 {
671                     // Argument for safety is in the definition of next_si.
672                     next_si = unsafe {
673                         self.next_si(next_si & !STATE_MATCH, text, at)
674                     };
675                     at += 1;
676                 }
677                 if at > cur {
678                     result = Result::Match(at - 2);
679                 }
680             } else if next_si & STATE_START > 0 {
681                 // A start state isn't in the common case because we may
682                 // want to do quick prefix scanning. If the program doesn't
683                 // have a detected prefix, then start states are actually
684                 // considered common and this case is never reached.
685                 debug_assert!(self.has_prefix());
686                 next_si &= !STATE_START;
687                 prev_si = next_si;
688                 at = match self.prefix_at(text, at) {
689                     None => return Result::NoMatch(text.len()),
690                     Some(i) => i,
691                 };
692             } else if next_si >= STATE_UNKNOWN {
693                 if next_si == STATE_QUIT {
694                     return Result::Quit;
695                 }
696                 // Finally, this corresponds to the case where the transition
697                 // entered a state that can never lead to a match or a state
698                 // that hasn't been computed yet. The latter being the "slow"
699                 // path.
700                 let byte = Byte::byte(text[at - 1]);
701                 // We no longer care about the special bits in the state
702                 // pointer.
703                 prev_si &= STATE_MAX;
704                 // Record where we are. This is used to track progress for
705                 // determining whether we should quit if we've flushed the
706                 // cache too much.
707                 self.at = at;
708                 next_si = match self.next_state(qcur, qnext, prev_si, byte) {
709                     None => return Result::Quit,
710                     Some(STATE_DEAD) => return result.set_non_match(at),
711                     Some(si) => si,
712                 };
713                 debug_assert!(next_si != STATE_UNKNOWN);
714                 if next_si & STATE_MATCH > 0 {
715                     next_si &= !STATE_MATCH;
716                     result = Result::Match(at - 1);
717                     if self.quit_after_match {
718                         return result;
719                     }
720                     self.last_match_si = next_si;
721                 }
722                 prev_si = next_si;
723             } else {
724                 prev_si = next_si;
725             }
726         }
727 
728         // Run the DFA once more on the special EOF sentinel value.
729         // We don't care about the special bits in the state pointer any more,
730         // so get rid of them.
731         prev_si &= STATE_MAX;
732         prev_si = match self.next_state(qcur, qnext, prev_si, Byte::eof()) {
733             None => return Result::Quit,
734             Some(STATE_DEAD) => return result.set_non_match(text.len()),
735             Some(si) => si & !STATE_START,
736         };
737         debug_assert!(prev_si != STATE_UNKNOWN);
738         if prev_si & STATE_MATCH > 0 {
739             prev_si &= !STATE_MATCH;
740             self.last_match_si = prev_si;
741             result = Result::Match(text.len());
742         }
743         result
744     }
745 
746     /// Executes the DFA on a reverse NFA.
747     #[cfg_attr(feature = "perf-inline", inline(always))]
exec_at_reverse( &mut self, qcur: &mut SparseSet, qnext: &mut SparseSet, text: &[u8], ) -> Result<usize>748     fn exec_at_reverse(
749         &mut self,
750         qcur: &mut SparseSet,
751         qnext: &mut SparseSet,
752         text: &[u8],
753     ) -> Result<usize> {
754         // The comments in `exec_at` above mostly apply here too. The main
755         // difference is that we move backwards over the input and we look for
756         // the longest possible match instead of the leftmost-first match.
757         //
758         // N.B. The code duplication here is regrettable. Efforts to improve
759         // it without sacrificing performance are welcome. ---AG
760         debug_assert!(self.prog.is_reverse);
761         let mut result = Result::NoMatch(self.at);
762         let (mut prev_si, mut next_si) = (self.start, self.start);
763         let mut at = self.at;
764         while at > 0 {
765             while next_si <= STATE_MAX && at > 0 {
766                 // Argument for safety is in the definition of next_si.
767                 at -= 1;
768                 prev_si = unsafe { self.next_si(next_si, text, at) };
769                 if prev_si > STATE_MAX || at <= 4 {
770                     mem::swap(&mut prev_si, &mut next_si);
771                     break;
772                 }
773                 at -= 1;
774                 next_si = unsafe { self.next_si(prev_si, text, at) };
775                 if next_si > STATE_MAX {
776                     break;
777                 }
778                 at -= 1;
779                 prev_si = unsafe { self.next_si(next_si, text, at) };
780                 if prev_si > STATE_MAX {
781                     mem::swap(&mut prev_si, &mut next_si);
782                     break;
783                 }
784                 at -= 1;
785                 next_si = unsafe { self.next_si(prev_si, text, at) };
786             }
787             if next_si & STATE_MATCH > 0 {
788                 next_si &= !STATE_MATCH;
789                 result = Result::Match(at + 1);
790                 if self.quit_after_match {
791                     return result;
792                 }
793                 self.last_match_si = next_si;
794                 prev_si = next_si;
795                 let cur = at;
796                 while (next_si & !STATE_MATCH) == prev_si && at >= 2 {
797                     // Argument for safety is in the definition of next_si.
798                     at -= 1;
799                     next_si = unsafe {
800                         self.next_si(next_si & !STATE_MATCH, text, at)
801                     };
802                 }
803                 if at < cur {
804                     result = Result::Match(at + 2);
805                 }
806             } else if next_si >= STATE_UNKNOWN {
807                 if next_si == STATE_QUIT {
808                     return Result::Quit;
809                 }
810                 let byte = Byte::byte(text[at]);
811                 prev_si &= STATE_MAX;
812                 self.at = at;
813                 next_si = match self.next_state(qcur, qnext, prev_si, byte) {
814                     None => return Result::Quit,
815                     Some(STATE_DEAD) => return result.set_non_match(at),
816                     Some(si) => si,
817                 };
818                 debug_assert!(next_si != STATE_UNKNOWN);
819                 if next_si & STATE_MATCH > 0 {
820                     next_si &= !STATE_MATCH;
821                     result = Result::Match(at + 1);
822                     if self.quit_after_match {
823                         return result;
824                     }
825                     self.last_match_si = next_si;
826                 }
827                 prev_si = next_si;
828             } else {
829                 prev_si = next_si;
830             }
831         }
832 
833         // Run the DFA once more on the special EOF sentinel value.
834         prev_si = match self.next_state(qcur, qnext, prev_si, Byte::eof()) {
835             None => return Result::Quit,
836             Some(STATE_DEAD) => return result.set_non_match(0),
837             Some(si) => si,
838         };
839         debug_assert!(prev_si != STATE_UNKNOWN);
840         if prev_si & STATE_MATCH > 0 {
841             prev_si &= !STATE_MATCH;
842             self.last_match_si = prev_si;
843             result = Result::Match(0);
844         }
845         result
846     }
847 
848     /// next_si transitions to the next state, where the transition input
849     /// corresponds to text[i].
850     ///
851     /// This elides bounds checks, and is therefore not safe.
852     #[cfg_attr(feature = "perf-inline", inline(always))]
next_si(&self, si: StatePtr, text: &[u8], i: usize) -> StatePtr853     unsafe fn next_si(&self, si: StatePtr, text: &[u8], i: usize) -> StatePtr {
854         // What is the argument for safety here?
855         // We have three unchecked accesses that could possibly violate safety:
856         //
857         //   1. The given byte of input (`text[i]`).
858         //   2. The class of the byte of input (`classes[text[i]]`).
859         //   3. The transition for the class (`trans[si + cls]`).
860         //
861         // (1) is only safe when calling next_si is guarded by
862         // `i < text.len()`.
863         //
864         // (2) is the easiest case to guarantee since `text[i]` is always a
865         // `u8` and `self.prog.byte_classes` always has length `u8::MAX`.
866         // (See `ByteClassSet.byte_classes` in `compile.rs`.)
867         //
868         // (3) is only safe if (1)+(2) are safe. Namely, the transitions
869         // of every state are defined to have length equal to the number of
870         // byte classes in the program. Therefore, a valid class leads to a
871         // valid transition. (All possible transitions are valid lookups, even
872         // if it points to a state that hasn't been computed yet.) (3) also
873         // relies on `si` being correct, but StatePtrs should only ever be
874         // retrieved from the transition table, which ensures they are correct.
875         debug_assert!(i < text.len());
876         let b = *text.get_unchecked(i);
877         debug_assert!((b as usize) < self.prog.byte_classes.len());
878         let cls = *self.prog.byte_classes.get_unchecked(b as usize);
879         self.cache.trans.next_unchecked(si, cls as usize)
880     }
881 
882     /// Computes the next state given the current state and the current input
883     /// byte (which may be EOF).
884     ///
885     /// If STATE_DEAD is returned, then there is no valid state transition.
886     /// This implies that no permutation of future input can lead to a match
887     /// state.
888     ///
889     /// STATE_UNKNOWN can never be returned.
exec_byte( &mut self, qcur: &mut SparseSet, qnext: &mut SparseSet, mut si: StatePtr, b: Byte, ) -> Option<StatePtr>890     fn exec_byte(
891         &mut self,
892         qcur: &mut SparseSet,
893         qnext: &mut SparseSet,
894         mut si: StatePtr,
895         b: Byte,
896     ) -> Option<StatePtr> {
897         use prog::Inst::*;
898 
899         // Initialize a queue with the current DFA state's NFA states.
900         qcur.clear();
901         for ip in self.state(si).inst_ptrs() {
902             qcur.insert(ip);
903         }
904 
905         // Before inspecting the current byte, we may need to also inspect
906         // whether the position immediately preceding the current byte
907         // satisfies the empty assertions found in the current state.
908         //
909         // We only need to do this step if there are any empty assertions in
910         // the current state.
911         let is_word_last = self.state(si).flags().is_word();
912         let is_word = b.is_ascii_word();
913         if self.state(si).flags().has_empty() {
914             // Compute the flags immediately preceding the current byte.
915             // This means we only care about the "end" or "end line" flags.
916             // (The "start" flags are computed immediately following the
917             // current byte and are handled below.)
918             let mut flags = EmptyFlags::default();
919             if b.is_eof() {
920                 flags.end = true;
921                 flags.end_line = true;
922             } else if b.as_byte().map_or(false, |b| b == b'\n') {
923                 flags.end_line = true;
924             }
925             if is_word_last == is_word {
926                 flags.not_word_boundary = true;
927             } else {
928                 flags.word_boundary = true;
929             }
930             // Now follow epsilon transitions from every NFA state, but make
931             // sure we only follow transitions that satisfy our flags.
932             qnext.clear();
933             for &ip in &*qcur {
934                 self.follow_epsilons(usize_to_u32(ip), qnext, flags);
935             }
936             mem::swap(qcur, qnext);
937         }
938 
939         // Now we set flags for immediately after the current byte. Since start
940         // states are processed separately, and are the only states that can
941         // have the StartText flag set, we therefore only need to worry about
942         // the StartLine flag here.
943         //
944         // We do also keep track of whether this DFA state contains a NFA state
945         // that is a matching state. This is precisely how we delay the DFA
946         // matching by one byte in order to process the special EOF sentinel
947         // byte. Namely, if this DFA state containing a matching NFA state,
948         // then it is the *next* DFA state that is marked as a match.
949         let mut empty_flags = EmptyFlags::default();
950         let mut state_flags = StateFlags::default();
951         empty_flags.start_line = b.as_byte().map_or(false, |b| b == b'\n');
952         if b.is_ascii_word() {
953             state_flags.set_word();
954         }
955         // Now follow all epsilon transitions again, but only after consuming
956         // the current byte.
957         qnext.clear();
958         for &ip in &*qcur {
959             match self.prog[ip as usize] {
960                 // These states never happen in a byte-based program.
961                 Char(_) | Ranges(_) => unreachable!(),
962                 // These states are handled when following epsilon transitions.
963                 Save(_) | Split(_) | EmptyLook(_) => {}
964                 Match(_) => {
965                     state_flags.set_match();
966                     if !self.continue_past_first_match() {
967                         break;
968                     } else if self.prog.matches.len() > 1
969                         && !qnext.contains(ip as usize)
970                     {
971                         // If we are continuing on to find other matches,
972                         // then keep a record of the match states we've seen.
973                         qnext.insert(ip);
974                     }
975                 }
976                 Bytes(ref inst) => {
977                     if b.as_byte().map_or(false, |b| inst.matches(b)) {
978                         self.follow_epsilons(
979                             inst.goto as InstPtr,
980                             qnext,
981                             empty_flags,
982                         );
983                     }
984                 }
985             }
986         }
987 
988         let cache = if b.is_eof() && self.prog.matches.len() > 1 {
989             // If we're processing the last byte of the input and we're
990             // matching a regex set, then make the next state contain the
991             // previous states transitions. We do this so that the main
992             // matching loop can extract all of the match instructions.
993             mem::swap(qcur, qnext);
994             // And don't cache this state because it's totally bunk.
995             false
996         } else {
997             true
998         };
999 
1000         // We've now built up the set of NFA states that ought to comprise the
1001         // next DFA state, so try to find it in the cache, and if it doesn't
1002         // exist, cache it.
1003         //
1004         // N.B. We pass `&mut si` here because the cache may clear itself if
1005         // it has gotten too full. When that happens, the location of the
1006         // current state may change.
1007         let mut next =
1008             match self.cached_state(qnext, state_flags, Some(&mut si)) {
1009                 None => return None,
1010                 Some(next) => next,
1011             };
1012         if (self.start & !STATE_START) == next {
1013             // Start states can never be match states since all matches are
1014             // delayed by one byte.
1015             debug_assert!(!self.state(next).flags().is_match());
1016             next = self.start_ptr(next);
1017         }
1018         if next <= STATE_MAX && self.state(next).flags().is_match() {
1019             next |= STATE_MATCH;
1020         }
1021         debug_assert!(next != STATE_UNKNOWN);
1022         // And now store our state in the current state's next list.
1023         if cache {
1024             let cls = self.byte_class(b);
1025             self.cache.trans.set_next(si, cls, next);
1026         }
1027         Some(next)
1028     }
1029 
1030     /// Follows the epsilon transitions starting at (and including) `ip`. The
1031     /// resulting states are inserted into the ordered set `q`.
1032     ///
1033     /// Conditional epsilon transitions (i.e., empty width assertions) are only
1034     /// followed if they are satisfied by the given flags, which should
1035     /// represent the flags set at the current location in the input.
1036     ///
1037     /// If the current location corresponds to the empty string, then only the
1038     /// end line and/or end text flags may be set. If the current location
1039     /// corresponds to a real byte in the input, then only the start line
1040     /// and/or start text flags may be set.
1041     ///
1042     /// As an exception to the above, when finding the initial state, any of
1043     /// the above flags may be set:
1044     ///
1045     /// If matching starts at the beginning of the input, then start text and
1046     /// start line should be set. If the input is empty, then end text and end
1047     /// line should also be set.
1048     ///
1049     /// If matching starts after the beginning of the input, then only start
1050     /// line should be set if the preceding byte is `\n`. End line should never
1051     /// be set in this case. (Even if the following byte is a `\n`, it will
1052     /// be handled in a subsequent DFA state.)
follow_epsilons( &mut self, ip: InstPtr, q: &mut SparseSet, flags: EmptyFlags, )1053     fn follow_epsilons(
1054         &mut self,
1055         ip: InstPtr,
1056         q: &mut SparseSet,
1057         flags: EmptyFlags,
1058     ) {
1059         use prog::EmptyLook::*;
1060         use prog::Inst::*;
1061 
1062         // We need to traverse the NFA to follow epsilon transitions, so avoid
1063         // recursion with an explicit stack.
1064         self.cache.stack.push(ip);
1065         while let Some(mut ip) = self.cache.stack.pop() {
1066             // Try to munch through as many states as possible without
1067             // pushes/pops to the stack.
1068             loop {
1069                 // Don't visit states we've already added.
1070                 if q.contains(ip as usize) {
1071                     break;
1072                 }
1073                 q.insert(ip as usize);
1074                 match self.prog[ip as usize] {
1075                     Char(_) | Ranges(_) => unreachable!(),
1076                     Match(_) | Bytes(_) => {
1077                         break;
1078                     }
1079                     EmptyLook(ref inst) => {
1080                         // Only follow empty assertion states if our flags
1081                         // satisfy the assertion.
1082                         match inst.look {
1083                             StartLine if flags.start_line => {
1084                                 ip = inst.goto as InstPtr;
1085                             }
1086                             EndLine if flags.end_line => {
1087                                 ip = inst.goto as InstPtr;
1088                             }
1089                             StartText if flags.start => {
1090                                 ip = inst.goto as InstPtr;
1091                             }
1092                             EndText if flags.end => {
1093                                 ip = inst.goto as InstPtr;
1094                             }
1095                             WordBoundaryAscii if flags.word_boundary => {
1096                                 ip = inst.goto as InstPtr;
1097                             }
1098                             NotWordBoundaryAscii
1099                                 if flags.not_word_boundary =>
1100                             {
1101                                 ip = inst.goto as InstPtr;
1102                             }
1103                             WordBoundary if flags.word_boundary => {
1104                                 ip = inst.goto as InstPtr;
1105                             }
1106                             NotWordBoundary if flags.not_word_boundary => {
1107                                 ip = inst.goto as InstPtr;
1108                             }
1109                             StartLine | EndLine | StartText | EndText
1110                             | WordBoundaryAscii | NotWordBoundaryAscii
1111                             | WordBoundary | NotWordBoundary => {
1112                                 break;
1113                             }
1114                         }
1115                     }
1116                     Save(ref inst) => {
1117                         ip = inst.goto as InstPtr;
1118                     }
1119                     Split(ref inst) => {
1120                         self.cache.stack.push(inst.goto2 as InstPtr);
1121                         ip = inst.goto1 as InstPtr;
1122                     }
1123                 }
1124             }
1125         }
1126     }
1127 
1128     /// Find a previously computed state matching the given set of instructions
1129     /// and is_match bool.
1130     ///
1131     /// The given set of instructions should represent a single state in the
1132     /// NFA along with all states reachable without consuming any input.
1133     ///
1134     /// The is_match bool should be true if and only if the preceding DFA state
1135     /// contains an NFA matching state. The cached state produced here will
1136     /// then signify a match. (This enables us to delay a match by one byte,
1137     /// in order to account for the EOF sentinel byte.)
1138     ///
1139     /// If the cache is full, then it is wiped before caching a new state.
1140     ///
1141     /// The current state should be specified if it exists, since it will need
1142     /// to be preserved if the cache clears itself. (Start states are
1143     /// always saved, so they should not be passed here.) It takes a mutable
1144     /// pointer to the index because if the cache is cleared, the state's
1145     /// location may change.
cached_state( &mut self, q: &SparseSet, mut state_flags: StateFlags, current_state: Option<&mut StatePtr>, ) -> Option<StatePtr>1146     fn cached_state(
1147         &mut self,
1148         q: &SparseSet,
1149         mut state_flags: StateFlags,
1150         current_state: Option<&mut StatePtr>,
1151     ) -> Option<StatePtr> {
1152         // If we couldn't come up with a non-empty key to represent this state,
1153         // then it is dead and can never lead to a match.
1154         //
1155         // Note that inst_flags represent the set of empty width assertions
1156         // in q. We use this as an optimization in exec_byte to determine when
1157         // we should follow epsilon transitions at the empty string preceding
1158         // the current byte.
1159         let key = match self.cached_state_key(q, &mut state_flags) {
1160             None => return Some(STATE_DEAD),
1161             Some(v) => v,
1162         };
1163         // In the cache? Cool. Done.
1164         if let Some(si) = self.cache.compiled.get_ptr(&key) {
1165             return Some(si);
1166         }
1167         // If the cache has gotten too big, wipe it.
1168         if self.approximate_size() > self.prog.dfa_size_limit
1169             && !self.clear_cache_and_save(current_state)
1170         {
1171             // Ooops. DFA is giving up.
1172             return None;
1173         }
1174         // Allocate room for our state and add it.
1175         self.add_state(key)
1176     }
1177 
1178     /// Produces a key suitable for describing a state in the DFA cache.
1179     ///
1180     /// The key invariant here is that equivalent keys are produced for any two
1181     /// sets of ordered NFA states (and toggling of whether the previous NFA
1182     /// states contain a match state) that do not discriminate a match for any
1183     /// input.
1184     ///
1185     /// Specifically, q should be an ordered set of NFA states and is_match
1186     /// should be true if and only if the previous NFA states contained a match
1187     /// state.
cached_state_key( &mut self, q: &SparseSet, state_flags: &mut StateFlags, ) -> Option<State>1188     fn cached_state_key(
1189         &mut self,
1190         q: &SparseSet,
1191         state_flags: &mut StateFlags,
1192     ) -> Option<State> {
1193         use prog::Inst::*;
1194 
1195         // We need to build up enough information to recognize pre-built states
1196         // in the DFA. Generally speaking, this includes every instruction
1197         // except for those which are purely epsilon transitions, e.g., the
1198         // Save and Split instructions.
1199         //
1200         // Empty width assertions are also epsilon transitions, but since they
1201         // are conditional, we need to make them part of a state's key in the
1202         // cache.
1203 
1204         let mut insts =
1205             mem::replace(&mut self.cache.insts_scratch_space, vec![]);
1206         insts.clear();
1207         // Reserve 1 byte for flags.
1208         insts.push(0);
1209 
1210         let mut prev = 0;
1211         for &ip in q {
1212             let ip = usize_to_u32(ip);
1213             match self.prog[ip as usize] {
1214                 Char(_) | Ranges(_) => unreachable!(),
1215                 Save(_) | Split(_) => {}
1216                 Bytes(_) => push_inst_ptr(&mut insts, &mut prev, ip),
1217                 EmptyLook(_) => {
1218                     state_flags.set_empty();
1219                     push_inst_ptr(&mut insts, &mut prev, ip)
1220                 }
1221                 Match(_) => {
1222                     push_inst_ptr(&mut insts, &mut prev, ip);
1223                     if !self.continue_past_first_match() {
1224                         break;
1225                     }
1226                 }
1227             }
1228         }
1229         // If we couldn't transition to any other instructions and we didn't
1230         // see a match when expanding NFA states previously, then this is a
1231         // dead state and no amount of additional input can transition out
1232         // of this state.
1233         let opt_state = if insts.len() == 1 && !state_flags.is_match() {
1234             None
1235         } else {
1236             let StateFlags(f) = *state_flags;
1237             insts[0] = f;
1238             Some(State { data: Arc::from(&*insts) })
1239         };
1240         self.cache.insts_scratch_space = insts;
1241         opt_state
1242     }
1243 
1244     /// Clears the cache, but saves and restores current_state if it is not
1245     /// none.
1246     ///
1247     /// The current state must be provided here in case its location in the
1248     /// cache changes.
1249     ///
1250     /// This returns false if the cache is not cleared and the DFA should
1251     /// give up.
clear_cache_and_save( &mut self, current_state: Option<&mut StatePtr>, ) -> bool1252     fn clear_cache_and_save(
1253         &mut self,
1254         current_state: Option<&mut StatePtr>,
1255     ) -> bool {
1256         if self.cache.compiled.is_empty() {
1257             // Nothing to clear...
1258             return true;
1259         }
1260         match current_state {
1261             None => self.clear_cache(),
1262             Some(si) => {
1263                 let cur = self.state(*si).clone();
1264                 if !self.clear_cache() {
1265                     return false;
1266                 }
1267                 // The unwrap is OK because we just cleared the cache and
1268                 // therefore know that the next state pointer won't exceed
1269                 // STATE_MAX.
1270                 *si = self.restore_state(cur).unwrap();
1271                 true
1272             }
1273         }
1274     }
1275 
1276     /// Wipes the state cache, but saves and restores the current start state.
1277     ///
1278     /// This returns false if the cache is not cleared and the DFA should
1279     /// give up.
clear_cache(&mut self) -> bool1280     fn clear_cache(&mut self) -> bool {
1281         // Bail out of the DFA if we're moving too "slowly."
1282         // A heuristic from RE2: assume the DFA is too slow if it is processing
1283         // 10 or fewer bytes per state.
1284         // Additionally, we permit the cache to be flushed a few times before
1285         // caling it quits.
1286         let nstates = self.cache.compiled.len();
1287         if self.cache.flush_count >= 3
1288             && self.at >= self.last_cache_flush
1289             && (self.at - self.last_cache_flush) <= 10 * nstates
1290         {
1291             return false;
1292         }
1293         // Update statistics tracking cache flushes.
1294         self.last_cache_flush = self.at;
1295         self.cache.flush_count += 1;
1296 
1297         // OK, actually flush the cache.
1298         let start = self.state(self.start & !STATE_START).clone();
1299         let last_match = if self.last_match_si <= STATE_MAX {
1300             Some(self.state(self.last_match_si).clone())
1301         } else {
1302             None
1303         };
1304         self.cache.reset_size();
1305         self.cache.trans.clear();
1306         self.cache.compiled.clear();
1307         for s in &mut self.cache.start_states {
1308             *s = STATE_UNKNOWN;
1309         }
1310         // The unwraps are OK because we just cleared the cache and therefore
1311         // know that the next state pointer won't exceed STATE_MAX.
1312         let start_ptr = self.restore_state(start).unwrap();
1313         self.start = self.start_ptr(start_ptr);
1314         if let Some(last_match) = last_match {
1315             self.last_match_si = self.restore_state(last_match).unwrap();
1316         }
1317         true
1318     }
1319 
1320     /// Restores the given state back into the cache, and returns a pointer
1321     /// to it.
restore_state(&mut self, state: State) -> Option<StatePtr>1322     fn restore_state(&mut self, state: State) -> Option<StatePtr> {
1323         // If we've already stored this state, just return a pointer to it.
1324         // None will be the wiser.
1325         if let Some(si) = self.cache.compiled.get_ptr(&state) {
1326             return Some(si);
1327         }
1328         self.add_state(state)
1329     }
1330 
1331     /// Returns the next state given the current state si and current byte
1332     /// b. {qcur,qnext} are used as scratch space for storing ordered NFA
1333     /// states.
1334     ///
1335     /// This tries to fetch the next state from the cache, but if that fails,
1336     /// it computes the next state, caches it and returns a pointer to it.
1337     ///
1338     /// The pointer can be to a real state, or it can be STATE_DEAD.
1339     /// STATE_UNKNOWN cannot be returned.
1340     ///
1341     /// None is returned if a new state could not be allocated (i.e., the DFA
1342     /// ran out of space and thinks it's running too slowly).
next_state( &mut self, qcur: &mut SparseSet, qnext: &mut SparseSet, si: StatePtr, b: Byte, ) -> Option<StatePtr>1343     fn next_state(
1344         &mut self,
1345         qcur: &mut SparseSet,
1346         qnext: &mut SparseSet,
1347         si: StatePtr,
1348         b: Byte,
1349     ) -> Option<StatePtr> {
1350         if si == STATE_DEAD {
1351             return Some(STATE_DEAD);
1352         }
1353         match self.cache.trans.next(si, self.byte_class(b)) {
1354             STATE_UNKNOWN => self.exec_byte(qcur, qnext, si, b),
1355             STATE_QUIT => None,
1356             STATE_DEAD => Some(STATE_DEAD),
1357             nsi => Some(nsi),
1358         }
1359     }
1360 
1361     /// Computes and returns the start state, where searching begins at
1362     /// position `at` in `text`. If the state has already been computed,
1363     /// then it is pulled from the cache. If the state hasn't been cached,
1364     /// then it is computed, cached and a pointer to it is returned.
1365     ///
1366     /// This may return STATE_DEAD but never STATE_UNKNOWN.
1367     #[cfg_attr(feature = "perf-inline", inline(always))]
start_state( &mut self, q: &mut SparseSet, empty_flags: EmptyFlags, state_flags: StateFlags, ) -> Option<StatePtr>1368     fn start_state(
1369         &mut self,
1370         q: &mut SparseSet,
1371         empty_flags: EmptyFlags,
1372         state_flags: StateFlags,
1373     ) -> Option<StatePtr> {
1374         // Compute an index into our cache of start states based on the set
1375         // of empty/state flags set at the current position in the input. We
1376         // don't use every flag since not all flags matter. For example, since
1377         // matches are delayed by one byte, start states can never be match
1378         // states.
1379         let flagi = {
1380             (((empty_flags.start as u8) << 0)
1381                 | ((empty_flags.end as u8) << 1)
1382                 | ((empty_flags.start_line as u8) << 2)
1383                 | ((empty_flags.end_line as u8) << 3)
1384                 | ((empty_flags.word_boundary as u8) << 4)
1385                 | ((empty_flags.not_word_boundary as u8) << 5)
1386                 | ((state_flags.is_word() as u8) << 6)) as usize
1387         };
1388         match self.cache.start_states[flagi] {
1389             STATE_UNKNOWN => {}
1390             STATE_DEAD => return Some(STATE_DEAD),
1391             si => return Some(si),
1392         }
1393         q.clear();
1394         let start = usize_to_u32(self.prog.start);
1395         self.follow_epsilons(start, q, empty_flags);
1396         // Start states can never be match states because we delay every match
1397         // by one byte. Given an empty string and an empty match, the match
1398         // won't actually occur until the DFA processes the special EOF
1399         // sentinel byte.
1400         let sp = match self.cached_state(q, state_flags, None) {
1401             None => return None,
1402             Some(sp) => self.start_ptr(sp),
1403         };
1404         self.cache.start_states[flagi] = sp;
1405         Some(sp)
1406     }
1407 
1408     /// Computes the set of starting flags for the given position in text.
1409     ///
1410     /// This should only be used when executing the DFA forwards over the
1411     /// input.
start_flags(&self, text: &[u8], at: usize) -> (EmptyFlags, StateFlags)1412     fn start_flags(&self, text: &[u8], at: usize) -> (EmptyFlags, StateFlags) {
1413         let mut empty_flags = EmptyFlags::default();
1414         let mut state_flags = StateFlags::default();
1415         empty_flags.start = at == 0;
1416         empty_flags.end = text.is_empty();
1417         empty_flags.start_line = at == 0 || text[at - 1] == b'\n';
1418         empty_flags.end_line = text.is_empty();
1419 
1420         let is_word_last = at > 0 && Byte::byte(text[at - 1]).is_ascii_word();
1421         let is_word = at < text.len() && Byte::byte(text[at]).is_ascii_word();
1422         if is_word_last {
1423             state_flags.set_word();
1424         }
1425         if is_word == is_word_last {
1426             empty_flags.not_word_boundary = true;
1427         } else {
1428             empty_flags.word_boundary = true;
1429         }
1430         (empty_flags, state_flags)
1431     }
1432 
1433     /// Computes the set of starting flags for the given position in text.
1434     ///
1435     /// This should only be used when executing the DFA in reverse over the
1436     /// input.
start_flags_reverse( &self, text: &[u8], at: usize, ) -> (EmptyFlags, StateFlags)1437     fn start_flags_reverse(
1438         &self,
1439         text: &[u8],
1440         at: usize,
1441     ) -> (EmptyFlags, StateFlags) {
1442         let mut empty_flags = EmptyFlags::default();
1443         let mut state_flags = StateFlags::default();
1444         empty_flags.start = at == text.len();
1445         empty_flags.end = text.is_empty();
1446         empty_flags.start_line = at == text.len() || text[at] == b'\n';
1447         empty_flags.end_line = text.is_empty();
1448 
1449         let is_word_last =
1450             at < text.len() && Byte::byte(text[at]).is_ascii_word();
1451         let is_word = at > 0 && Byte::byte(text[at - 1]).is_ascii_word();
1452         if is_word_last {
1453             state_flags.set_word();
1454         }
1455         if is_word == is_word_last {
1456             empty_flags.not_word_boundary = true;
1457         } else {
1458             empty_flags.word_boundary = true;
1459         }
1460         (empty_flags, state_flags)
1461     }
1462 
1463     /// Returns a reference to a State given a pointer to it.
state(&self, si: StatePtr) -> &State1464     fn state(&self, si: StatePtr) -> &State {
1465         self.cache.compiled.get_state(si).unwrap()
1466     }
1467 
1468     /// Adds the given state to the DFA.
1469     ///
1470     /// This allocates room for transitions out of this state in
1471     /// self.cache.trans. The transitions can be set with the returned
1472     /// StatePtr.
1473     ///
1474     /// If None is returned, then the state limit was reached and the DFA
1475     /// should quit.
add_state(&mut self, state: State) -> Option<StatePtr>1476     fn add_state(&mut self, state: State) -> Option<StatePtr> {
1477         // This will fail if the next state pointer exceeds STATE_PTR. In
1478         // practice, the cache limit will prevent us from ever getting here,
1479         // but maybe callers will set the cache size to something ridiculous...
1480         let si = match self.cache.trans.add() {
1481             None => return None,
1482             Some(si) => si,
1483         };
1484         // If the program has a Unicode word boundary, then set any transitions
1485         // for non-ASCII bytes to STATE_QUIT. If the DFA stumbles over such a
1486         // transition, then it will quit and an alternative matching engine
1487         // will take over.
1488         if self.prog.has_unicode_word_boundary {
1489             for b in 128..256 {
1490                 let cls = self.byte_class(Byte::byte(b as u8));
1491                 self.cache.trans.set_next(si, cls, STATE_QUIT);
1492             }
1493         }
1494         // Finally, put our actual state on to our heap of states and index it
1495         // so we can find it later.
1496         self.cache.size += self.cache.trans.state_heap_size()
1497             + state.data.len()
1498             + (2 * mem::size_of::<State>())
1499             + mem::size_of::<StatePtr>();
1500         self.cache.compiled.insert(state, si);
1501         // Transition table and set of states and map should all be in sync.
1502         debug_assert!(
1503             self.cache.compiled.len() == self.cache.trans.num_states()
1504         );
1505         Some(si)
1506     }
1507 
1508     /// Quickly finds the next occurrence of any literal prefixes in the regex.
1509     /// If there are no literal prefixes, then the current position is
1510     /// returned. If there are literal prefixes and one could not be found,
1511     /// then None is returned.
1512     ///
1513     /// This should only be called when the DFA is in a start state.
prefix_at(&self, text: &[u8], at: usize) -> Option<usize>1514     fn prefix_at(&self, text: &[u8], at: usize) -> Option<usize> {
1515         self.prog.prefixes.find(&text[at..]).map(|(s, _)| at + s)
1516     }
1517 
1518     /// Returns the number of byte classes required to discriminate transitions
1519     /// in each state.
1520     ///
1521     /// invariant: num_byte_classes() == len(State.next)
num_byte_classes(&self) -> usize1522     fn num_byte_classes(&self) -> usize {
1523         // We add 1 to account for the special EOF byte.
1524         (self.prog.byte_classes[255] as usize + 1) + 1
1525     }
1526 
1527     /// Given an input byte or the special EOF sentinel, return its
1528     /// corresponding byte class.
1529     #[cfg_attr(feature = "perf-inline", inline(always))]
byte_class(&self, b: Byte) -> usize1530     fn byte_class(&self, b: Byte) -> usize {
1531         match b.as_byte() {
1532             None => self.num_byte_classes() - 1,
1533             Some(b) => self.u8_class(b),
1534         }
1535     }
1536 
1537     /// Like byte_class, but explicitly for u8s.
1538     #[cfg_attr(feature = "perf-inline", inline(always))]
u8_class(&self, b: u8) -> usize1539     fn u8_class(&self, b: u8) -> usize {
1540         self.prog.byte_classes[b as usize] as usize
1541     }
1542 
1543     /// Returns true if the DFA should continue searching past the first match.
1544     ///
1545     /// Leftmost first semantics in the DFA are preserved by not following NFA
1546     /// transitions after the first match is seen.
1547     ///
1548     /// On occasion, we want to avoid leftmost first semantics to find either
1549     /// the longest match (for reverse search) or all possible matches (for
1550     /// regex sets).
continue_past_first_match(&self) -> bool1551     fn continue_past_first_match(&self) -> bool {
1552         self.prog.is_reverse || self.prog.matches.len() > 1
1553     }
1554 
1555     /// Returns true if there is a prefix we can quickly search for.
has_prefix(&self) -> bool1556     fn has_prefix(&self) -> bool {
1557         !self.prog.is_reverse
1558             && !self.prog.prefixes.is_empty()
1559             && !self.prog.is_anchored_start
1560     }
1561 
1562     /// Sets the STATE_START bit in the given state pointer if and only if
1563     /// we have a prefix to scan for.
1564     ///
1565     /// If there's no prefix, then it's a waste to treat the start state
1566     /// specially.
start_ptr(&self, si: StatePtr) -> StatePtr1567     fn start_ptr(&self, si: StatePtr) -> StatePtr {
1568         if self.has_prefix() {
1569             si | STATE_START
1570         } else {
1571             si
1572         }
1573     }
1574 
1575     /// Approximate size returns the approximate heap space currently used by
1576     /// the DFA. It is used to determine whether the DFA's state cache needs to
1577     /// be wiped. Namely, it is possible that for certain regexes on certain
1578     /// inputs, a new state could be created for every byte of input. (This is
1579     /// bad for memory use, so we bound it with a cache.)
approximate_size(&self) -> usize1580     fn approximate_size(&self) -> usize {
1581         self.cache.size + self.prog.approximate_size()
1582     }
1583 }
1584 
1585 /// An abstraction for representing a map of states. The map supports two
1586 /// different ways of state lookup. One is fast constant time access via a
1587 /// state pointer. The other is a hashmap lookup based on the DFA's
1588 /// constituent NFA states.
1589 ///
1590 /// A DFA state internally uses an Arc such that we only need to store the
1591 /// set of NFA states on the heap once, even though we support looking up
1592 /// states by two different means. A more natural way to express this might
1593 /// use raw pointers, but an Arc is safe and effectively achieves the same
1594 /// thing.
1595 #[derive(Debug)]
1596 struct StateMap {
1597     /// The keys are not actually static but rely on always pointing to a
1598     /// buffer in `states` which will never be moved except when clearing
1599     /// the map or on drop, in which case the keys of this map will be
1600     /// removed before
1601     map: HashMap<State, StatePtr>,
1602     /// Our set of states. Note that `StatePtr / num_byte_classes` indexes
1603     /// this Vec rather than just a `StatePtr`.
1604     states: Vec<State>,
1605     /// The number of byte classes in the DFA. Used to index `states`.
1606     num_byte_classes: usize,
1607 }
1608 
1609 impl StateMap {
new(num_byte_classes: usize) -> StateMap1610     fn new(num_byte_classes: usize) -> StateMap {
1611         StateMap {
1612             map: HashMap::new(),
1613             states: vec![],
1614             num_byte_classes: num_byte_classes,
1615         }
1616     }
1617 
len(&self) -> usize1618     fn len(&self) -> usize {
1619         self.states.len()
1620     }
1621 
is_empty(&self) -> bool1622     fn is_empty(&self) -> bool {
1623         self.states.is_empty()
1624     }
1625 
get_ptr(&self, state: &State) -> Option<StatePtr>1626     fn get_ptr(&self, state: &State) -> Option<StatePtr> {
1627         self.map.get(state).cloned()
1628     }
1629 
get_state(&self, si: StatePtr) -> Option<&State>1630     fn get_state(&self, si: StatePtr) -> Option<&State> {
1631         self.states.get(si as usize / self.num_byte_classes)
1632     }
1633 
insert(&mut self, state: State, si: StatePtr)1634     fn insert(&mut self, state: State, si: StatePtr) {
1635         self.map.insert(state.clone(), si);
1636         self.states.push(state);
1637     }
1638 
clear(&mut self)1639     fn clear(&mut self) {
1640         self.map.clear();
1641         self.states.clear();
1642     }
1643 }
1644 
1645 impl Transitions {
1646     /// Create a new transition table.
1647     ///
1648     /// The number of byte classes corresponds to the stride. Every state will
1649     /// have `num_byte_classes` slots for transitions.
new(num_byte_classes: usize) -> Transitions1650     fn new(num_byte_classes: usize) -> Transitions {
1651         Transitions { table: vec![], num_byte_classes: num_byte_classes }
1652     }
1653 
1654     /// Returns the total number of states currently in this table.
num_states(&self) -> usize1655     fn num_states(&self) -> usize {
1656         self.table.len() / self.num_byte_classes
1657     }
1658 
1659     /// Allocates room for one additional state and returns a pointer to it.
1660     ///
1661     /// If there's no more room, None is returned.
add(&mut self) -> Option<StatePtr>1662     fn add(&mut self) -> Option<StatePtr> {
1663         let si = self.table.len();
1664         if si > STATE_MAX as usize {
1665             return None;
1666         }
1667         self.table.extend(repeat(STATE_UNKNOWN).take(self.num_byte_classes));
1668         Some(usize_to_u32(si))
1669     }
1670 
1671     /// Clears the table of all states.
clear(&mut self)1672     fn clear(&mut self) {
1673         self.table.clear();
1674     }
1675 
1676     /// Sets the transition from (si, cls) to next.
set_next(&mut self, si: StatePtr, cls: usize, next: StatePtr)1677     fn set_next(&mut self, si: StatePtr, cls: usize, next: StatePtr) {
1678         self.table[si as usize + cls] = next;
1679     }
1680 
1681     /// Returns the transition corresponding to (si, cls).
next(&self, si: StatePtr, cls: usize) -> StatePtr1682     fn next(&self, si: StatePtr, cls: usize) -> StatePtr {
1683         self.table[si as usize + cls]
1684     }
1685 
1686     /// The heap size, in bytes, of a single state in the transition table.
state_heap_size(&self) -> usize1687     fn state_heap_size(&self) -> usize {
1688         self.num_byte_classes * mem::size_of::<StatePtr>()
1689     }
1690 
1691     /// Like `next`, but uses unchecked access and is therefore not safe.
next_unchecked(&self, si: StatePtr, cls: usize) -> StatePtr1692     unsafe fn next_unchecked(&self, si: StatePtr, cls: usize) -> StatePtr {
1693         debug_assert!((si as usize) < self.table.len());
1694         debug_assert!(cls < self.num_byte_classes);
1695         *self.table.get_unchecked(si as usize + cls)
1696     }
1697 }
1698 
1699 impl StateFlags {
is_match(&self) -> bool1700     fn is_match(&self) -> bool {
1701         self.0 & 0b0000000_1 > 0
1702     }
1703 
set_match(&mut self)1704     fn set_match(&mut self) {
1705         self.0 |= 0b0000000_1;
1706     }
1707 
is_word(&self) -> bool1708     fn is_word(&self) -> bool {
1709         self.0 & 0b000000_1_0 > 0
1710     }
1711 
set_word(&mut self)1712     fn set_word(&mut self) {
1713         self.0 |= 0b000000_1_0;
1714     }
1715 
has_empty(&self) -> bool1716     fn has_empty(&self) -> bool {
1717         self.0 & 0b00000_1_00 > 0
1718     }
1719 
set_empty(&mut self)1720     fn set_empty(&mut self) {
1721         self.0 |= 0b00000_1_00;
1722     }
1723 }
1724 
1725 impl Byte {
byte(b: u8) -> Self1726     fn byte(b: u8) -> Self {
1727         Byte(b as u16)
1728     }
eof() -> Self1729     fn eof() -> Self {
1730         Byte(256)
1731     }
is_eof(&self) -> bool1732     fn is_eof(&self) -> bool {
1733         self.0 == 256
1734     }
1735 
is_ascii_word(&self) -> bool1736     fn is_ascii_word(&self) -> bool {
1737         let b = match self.as_byte() {
1738             None => return false,
1739             Some(b) => b,
1740         };
1741         match b {
1742             b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'_' => true,
1743             _ => false,
1744         }
1745     }
1746 
as_byte(&self) -> Option<u8>1747     fn as_byte(&self) -> Option<u8> {
1748         if self.is_eof() {
1749             None
1750         } else {
1751             Some(self.0 as u8)
1752         }
1753     }
1754 }
1755 
1756 impl fmt::Debug for State {
fmt(&self, f: &mut fmt::Formatter) -> fmt::Result1757     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
1758         let ips: Vec<usize> = self.inst_ptrs().collect();
1759         f.debug_struct("State")
1760             .field("flags", &self.flags())
1761             .field("insts", &ips)
1762             .finish()
1763     }
1764 }
1765 
1766 impl fmt::Debug for Transitions {
fmt(&self, f: &mut fmt::Formatter) -> fmt::Result1767     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
1768         let mut fmtd = f.debug_map();
1769         for si in 0..self.num_states() {
1770             let s = si * self.num_byte_classes;
1771             let e = s + self.num_byte_classes;
1772             fmtd.entry(&si.to_string(), &TransitionsRow(&self.table[s..e]));
1773         }
1774         fmtd.finish()
1775     }
1776 }
1777 
1778 struct TransitionsRow<'a>(&'a [StatePtr]);
1779 
1780 impl<'a> fmt::Debug for TransitionsRow<'a> {
fmt(&self, f: &mut fmt::Formatter) -> fmt::Result1781     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
1782         let mut fmtd = f.debug_map();
1783         for (b, si) in self.0.iter().enumerate() {
1784             match *si {
1785                 STATE_UNKNOWN => {}
1786                 STATE_DEAD => {
1787                     fmtd.entry(&vb(b as usize), &"DEAD");
1788                 }
1789                 si => {
1790                     fmtd.entry(&vb(b as usize), &si.to_string());
1791                 }
1792             }
1793         }
1794         fmtd.finish()
1795     }
1796 }
1797 
1798 impl fmt::Debug for StateFlags {
fmt(&self, f: &mut fmt::Formatter) -> fmt::Result1799     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
1800         f.debug_struct("StateFlags")
1801             .field("is_match", &self.is_match())
1802             .field("is_word", &self.is_word())
1803             .field("has_empty", &self.has_empty())
1804             .finish()
1805     }
1806 }
1807 
1808 /// Helper function for formatting a byte as a nice-to-read escaped string.
vb(b: usize) -> String1809 fn vb(b: usize) -> String {
1810     use std::ascii::escape_default;
1811 
1812     if b > ::std::u8::MAX as usize {
1813         "EOF".to_owned()
1814     } else {
1815         let escaped = escape_default(b as u8).collect::<Vec<u8>>();
1816         String::from_utf8_lossy(&escaped).into_owned()
1817     }
1818 }
1819 
usize_to_u32(n: usize) -> u321820 fn usize_to_u32(n: usize) -> u32 {
1821     if (n as u64) > (::std::u32::MAX as u64) {
1822         panic!("BUG: {} is too big to fit into u32", n)
1823     }
1824     n as u32
1825 }
1826 
1827 #[allow(dead_code)] // useful for debugging
show_state_ptr(si: StatePtr) -> String1828 fn show_state_ptr(si: StatePtr) -> String {
1829     let mut s = format!("{:?}", si & STATE_MAX);
1830     if si == STATE_UNKNOWN {
1831         s = format!("{} (unknown)", s);
1832     }
1833     if si == STATE_DEAD {
1834         s = format!("{} (dead)", s);
1835     }
1836     if si == STATE_QUIT {
1837         s = format!("{} (quit)", s);
1838     }
1839     if si & STATE_START > 0 {
1840         s = format!("{} (start)", s);
1841     }
1842     if si & STATE_MATCH > 0 {
1843         s = format!("{} (match)", s);
1844     }
1845     s
1846 }
1847 
1848 /// https://developers.google.com/protocol-buffers/docs/encoding#varints
write_vari32(data: &mut Vec<u8>, n: i32)1849 fn write_vari32(data: &mut Vec<u8>, n: i32) {
1850     let mut un = (n as u32) << 1;
1851     if n < 0 {
1852         un = !un;
1853     }
1854     write_varu32(data, un)
1855 }
1856 
1857 /// https://developers.google.com/protocol-buffers/docs/encoding#varints
read_vari32(data: &[u8]) -> (i32, usize)1858 fn read_vari32(data: &[u8]) -> (i32, usize) {
1859     let (un, i) = read_varu32(data);
1860     let mut n = (un >> 1) as i32;
1861     if un & 1 != 0 {
1862         n = !n;
1863     }
1864     (n, i)
1865 }
1866 
1867 /// https://developers.google.com/protocol-buffers/docs/encoding#varints
write_varu32(data: &mut Vec<u8>, mut n: u32)1868 fn write_varu32(data: &mut Vec<u8>, mut n: u32) {
1869     while n >= 0b1000_0000 {
1870         data.push((n as u8) | 0b1000_0000);
1871         n >>= 7;
1872     }
1873     data.push(n as u8);
1874 }
1875 
1876 /// https://developers.google.com/protocol-buffers/docs/encoding#varints
read_varu32(data: &[u8]) -> (u32, usize)1877 fn read_varu32(data: &[u8]) -> (u32, usize) {
1878     let mut n: u32 = 0;
1879     let mut shift: u32 = 0;
1880     for (i, &b) in data.iter().enumerate() {
1881         if b < 0b1000_0000 {
1882             return (n | ((b as u32) << shift), i + 1);
1883         }
1884         n |= ((b as u32) & 0b0111_1111) << shift;
1885         shift += 7;
1886     }
1887     (0, 0)
1888 }
1889 
1890 #[cfg(test)]
1891 mod tests {
1892     extern crate rand;
1893 
1894     use super::{
1895         push_inst_ptr, read_vari32, read_varu32, write_vari32, write_varu32,
1896         State, StateFlags,
1897     };
1898     use quickcheck::{quickcheck, Gen, QuickCheck};
1899     use std::sync::Arc;
1900 
1901     #[test]
prop_state_encode_decode()1902     fn prop_state_encode_decode() {
1903         fn p(mut ips: Vec<u32>, flags: u8) -> bool {
1904             // It looks like our encoding scheme can't handle instruction
1905             // pointers at or above 2**31. We should fix that, but it seems
1906             // unlikely to occur in real code due to the amount of memory
1907             // required for such a state machine. So for now, we just clamp
1908             // our test data.
1909             for ip in &mut ips {
1910                 if *ip >= 1 << 31 {
1911                     *ip = (1 << 31) - 1;
1912                 }
1913             }
1914             let mut data = vec![flags];
1915             let mut prev = 0;
1916             for &ip in ips.iter() {
1917                 push_inst_ptr(&mut data, &mut prev, ip);
1918             }
1919             let state = State { data: Arc::from(&data[..]) };
1920 
1921             let expected: Vec<usize> =
1922                 ips.into_iter().map(|ip| ip as usize).collect();
1923             let got: Vec<usize> = state.inst_ptrs().collect();
1924             expected == got && state.flags() == StateFlags(flags)
1925         }
1926         QuickCheck::new()
1927             .gen(Gen::new(10_000))
1928             .quickcheck(p as fn(Vec<u32>, u8) -> bool);
1929     }
1930 
1931     #[test]
prop_read_write_u32()1932     fn prop_read_write_u32() {
1933         fn p(n: u32) -> bool {
1934             let mut buf = vec![];
1935             write_varu32(&mut buf, n);
1936             let (got, nread) = read_varu32(&buf);
1937             nread == buf.len() && got == n
1938         }
1939         quickcheck(p as fn(u32) -> bool);
1940     }
1941 
1942     #[test]
prop_read_write_i32()1943     fn prop_read_write_i32() {
1944         fn p(n: i32) -> bool {
1945             let mut buf = vec![];
1946             write_vari32(&mut buf, n);
1947             let (got, nread) = read_vari32(&buf);
1948             nread == buf.len() && got == n
1949         }
1950         quickcheck(p as fn(i32) -> bool);
1951     }
1952 }
1953