regex_automata/dfa/
minimize.rs

1use core::{cell::RefCell, fmt, mem};
2
3use alloc::{collections::BTreeMap, rc::Rc, vec, vec::Vec};
4
5use crate::{
6    dfa::{automaton::Automaton, dense, DEAD},
7    util::{
8        alphabet,
9        primitives::{PatternID, StateID},
10    },
11};
12
13/// An implementation of Hopcroft's algorithm for minimizing DFAs.
14///
15/// The algorithm implemented here is mostly taken from Wikipedia:
16/// https://en.wikipedia.org/wiki/DFA_minimization#Hopcroft's_algorithm
17///
18/// This code has had some light optimization attention paid to it,
19/// particularly in the form of reducing allocation as much as possible.
20/// However, it is still generally slow. Future optimization work should
21/// probably focus on the bigger picture rather than micro-optimizations. For
22/// example:
23///
24/// 1. Figure out how to more intelligently create initial partitions. That is,
25///    Hopcroft's algorithm starts by creating two partitions of DFA states
26///    that are known to NOT be equivalent: match states and non-match states.
27///    The algorithm proceeds by progressively refining these partitions into
28///    smaller partitions. If we could start with more partitions, then we
29///    could reduce the amount of work that Hopcroft's algorithm needs to do.
30/// 2. For every partition that we visit, we find all incoming transitions to
31///    every state in the partition for *every* element in the alphabet. (This
32///    is why using byte classes can significantly decrease minimization times,
33///    since byte classes shrink the alphabet.) This is quite costly and there
34///    is perhaps some redundant work being performed depending on the specific
35///    states in the set. For example, we might be able to only visit some
36///    elements of the alphabet based on the transitions.
37/// 3. Move parts of minimization into determinization. If minimization has
38///    fewer states to deal with, then it should run faster. A prime example
39///    of this might be large Unicode classes, which are generated in way that
40///    can create a lot of redundant states. (Some work has been done on this
41///    point during NFA compilation via the algorithm described in the
42///    "Incremental Construction of MinimalAcyclic Finite-State Automata"
43///    paper.)
44pub(crate) struct Minimizer<'a> {
45    dfa: &'a mut dense::OwnedDFA,
46    in_transitions: Vec<Vec<Vec<StateID>>>,
47    partitions: Vec<StateSet>,
48    waiting: Vec<StateSet>,
49}
50
51impl<'a> fmt::Debug for Minimizer<'a> {
52    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
53        f.debug_struct("Minimizer")
54            .field("dfa", &self.dfa)
55            .field("in_transitions", &self.in_transitions)
56            .field("partitions", &self.partitions)
57            .field("waiting", &self.waiting)
58            .finish()
59    }
60}
61
62/// A set of states. A state set makes up a single partition in Hopcroft's
63/// algorithm.
64///
65/// It is represented by an ordered set of state identifiers. We use shared
66/// ownership so that a single state set can be in both the set of partitions
67/// and in the set of waiting sets simultaneously without an additional
68/// allocation. Generally, once a state set is built, it becomes immutable.
69///
70/// We use this representation because it avoids the overhead of more
71/// traditional set data structures (HashSet/BTreeSet), and also because
72/// computing intersection/subtraction on this representation is especially
73/// fast.
74#[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord)]
75struct StateSet {
76    ids: Rc<RefCell<Vec<StateID>>>,
77}
78
79impl<'a> Minimizer<'a> {
80    pub fn new(dfa: &'a mut dense::OwnedDFA) -> Minimizer<'a> {
81        let in_transitions = Minimizer::incoming_transitions(dfa);
82        let partitions = Minimizer::initial_partitions(dfa);
83        let waiting = partitions.clone();
84        Minimizer { dfa, in_transitions, partitions, waiting }
85    }
86
87    pub fn run(mut self) {
88        let stride2 = self.dfa.stride2();
89        let as_state_id = |index: usize| -> StateID {
90            StateID::new(index << stride2).unwrap()
91        };
92        let as_index = |id: StateID| -> usize { id.as_usize() >> stride2 };
93
94        let mut incoming = StateSet::empty();
95        let mut scratch1 = StateSet::empty();
96        let mut scratch2 = StateSet::empty();
97        let mut newparts = vec![];
98
99        // This loop is basically Hopcroft's algorithm. Everything else is just
100        // shuffling data around to fit our representation.
101        while let Some(set) = self.waiting.pop() {
102            for b in self.dfa.byte_classes().iter() {
103                self.find_incoming_to(b, &set, &mut incoming);
104                // If incoming is empty, then the intersection with any other
105                // set must also be empty. So 'newparts' just ends up being
106                // 'self.partitions'. So there's no need to go through the loop
107                // below.
108                //
109                // This actually turns out to be rather large optimization. On
110                // the order of making minimization 4-5x faster. It's likely
111                // that the vast majority of all states have very few incoming
112                // transitions.
113                if incoming.is_empty() {
114                    continue;
115                }
116
117                for p in 0..self.partitions.len() {
118                    self.partitions[p].intersection(&incoming, &mut scratch1);
119                    if scratch1.is_empty() {
120                        newparts.push(self.partitions[p].clone());
121                        continue;
122                    }
123
124                    self.partitions[p].subtract(&incoming, &mut scratch2);
125                    if scratch2.is_empty() {
126                        newparts.push(self.partitions[p].clone());
127                        continue;
128                    }
129
130                    let (x, y) =
131                        (scratch1.deep_clone(), scratch2.deep_clone());
132                    newparts.push(x.clone());
133                    newparts.push(y.clone());
134                    match self.find_waiting(&self.partitions[p]) {
135                        Some(i) => {
136                            self.waiting[i] = x;
137                            self.waiting.push(y);
138                        }
139                        None => {
140                            if x.len() <= y.len() {
141                                self.waiting.push(x);
142                            } else {
143                                self.waiting.push(y);
144                            }
145                        }
146                    }
147                }
148                newparts = mem::replace(&mut self.partitions, newparts);
149                newparts.clear();
150            }
151        }
152
153        // At this point, we now have a minimal partitioning of states, where
154        // each partition is an equivalence class of DFA states. Now we need to
155        // use this partitioning to update the DFA to only contain one state for
156        // each partition.
157
158        // Create a map from DFA state ID to the representative ID of the
159        // equivalence class to which it belongs. The representative ID of an
160        // equivalence class of states is the minimum ID in that class.
161        let mut state_to_part = vec![DEAD; self.dfa.state_len()];
162        for p in &self.partitions {
163            p.iter(|id| state_to_part[as_index(id)] = p.min());
164        }
165
166        // Generate a new contiguous sequence of IDs for minimal states, and
167        // create a map from equivalence IDs to the new IDs. Thus, the new
168        // minimal ID of *any* state in the unminimized DFA can be obtained
169        // with minimals_ids[state_to_part[old_id]].
170        let mut minimal_ids = vec![DEAD; self.dfa.state_len()];
171        let mut new_index = 0;
172        for state in self.dfa.states() {
173            if state_to_part[as_index(state.id())] == state.id() {
174                minimal_ids[as_index(state.id())] = as_state_id(new_index);
175                new_index += 1;
176            }
177        }
178        // The total number of states in the minimal DFA.
179        let minimal_count = new_index;
180        // Convenience function for remapping state IDs. This takes an old ID,
181        // looks up its Hopcroft partition and then maps that to the new ID
182        // range.
183        let remap = |old| minimal_ids[as_index(state_to_part[as_index(old)])];
184
185        // Re-map this DFA in place such that the only states remaining
186        // correspond to the representative states of every equivalence class.
187        for id in (0..self.dfa.state_len()).map(as_state_id) {
188            // If this state isn't a representative for an equivalence class,
189            // then we skip it since it won't appear in the minimal DFA.
190            if state_to_part[as_index(id)] != id {
191                continue;
192            }
193            self.dfa.remap_state(id, remap);
194            self.dfa.swap_states(id, minimal_ids[as_index(id)]);
195        }
196        // Trim off all unused states from the pre-minimized DFA. This
197        // represents all states that were merged into a non-singleton
198        // equivalence class of states, and appeared after the first state
199        // in each such class. (Because the state with the smallest ID in each
200        // equivalence class is its representative ID.)
201        self.dfa.truncate_states(minimal_count);
202
203        // Update the new start states, which is now just the minimal ID of
204        // whatever state the old start state was collapsed into. Also, we
205        // collect everything before-hand to work around the borrow checker.
206        // We're already allocating so much that this is probably fine. If this
207        // turns out to be costly, then I guess add a `starts_mut` iterator.
208        let starts: Vec<_> = self.dfa.starts().collect();
209        for (old_start_id, anchored, start_type) in starts {
210            self.dfa.set_start_state(
211                anchored,
212                start_type,
213                remap(old_start_id),
214            );
215        }
216
217        // Update the match state pattern ID list for multi-regexes. All we
218        // need to do is remap the match state IDs. The pattern ID lists are
219        // always the same as they were since match states with distinct
220        // pattern ID lists are always considered distinct states.
221        let mut pmap = BTreeMap::new();
222        for (match_id, pattern_ids) in self.dfa.pattern_map() {
223            let new_id = remap(match_id);
224            pmap.insert(new_id, pattern_ids);
225        }
226        // This unwrap is OK because minimization never increases the number of
227        // match states or patterns in those match states. Since minimization
228        // runs after the pattern map has already been set at least once, we
229        // know that our match states cannot error.
230        self.dfa.set_pattern_map(&pmap).unwrap();
231
232        // In order to update the ID of the maximum match state, we need to
233        // find the maximum ID among all of the match states in the minimized
234        // DFA. This is not necessarily the new ID of the unminimized maximum
235        // match state, since that could have been collapsed with a much
236        // earlier match state. Therefore, to find the new max match state,
237        // we iterate over all previous match states, find their corresponding
238        // new minimal ID, and take the maximum of those.
239        let old = self.dfa.special().clone();
240        let new = self.dfa.special_mut();
241        // ... but only remap if we had match states.
242        if old.matches() {
243            new.min_match = StateID::MAX;
244            new.max_match = StateID::ZERO;
245            for i in as_index(old.min_match)..=as_index(old.max_match) {
246                let new_id = remap(as_state_id(i));
247                if new_id < new.min_match {
248                    new.min_match = new_id;
249                }
250                if new_id > new.max_match {
251                    new.max_match = new_id;
252                }
253            }
254        }
255        // ... same, but for start states.
256        if old.starts() {
257            new.min_start = StateID::MAX;
258            new.max_start = StateID::ZERO;
259            for i in as_index(old.min_start)..=as_index(old.max_start) {
260                let new_id = remap(as_state_id(i));
261                if new_id == DEAD {
262                    continue;
263                }
264                if new_id < new.min_start {
265                    new.min_start = new_id;
266                }
267                if new_id > new.max_start {
268                    new.max_start = new_id;
269                }
270            }
271            if new.max_start == DEAD {
272                new.min_start = DEAD;
273            }
274        }
275        new.quit_id = remap(new.quit_id);
276        new.set_max();
277    }
278
279    fn find_waiting(&self, set: &StateSet) -> Option<usize> {
280        self.waiting.iter().position(|s| s == set)
281    }
282
283    fn find_incoming_to(
284        &self,
285        b: alphabet::Unit,
286        set: &StateSet,
287        incoming: &mut StateSet,
288    ) {
289        incoming.clear();
290        set.iter(|id| {
291            for &inid in
292                &self.in_transitions[self.dfa.to_index(id)][b.as_usize()]
293            {
294                incoming.add(inid);
295            }
296        });
297        incoming.canonicalize();
298    }
299
300    fn initial_partitions(dfa: &dense::OwnedDFA) -> Vec<StateSet> {
301        // For match states, we know that two match states with different
302        // pattern ID lists will *always* be distinct, so we can partition them
303        // initially based on that.
304        let mut matching: BTreeMap<Vec<PatternID>, StateSet> = BTreeMap::new();
305        let mut is_quit = StateSet::empty();
306        let mut no_match = StateSet::empty();
307        for state in dfa.states() {
308            if dfa.is_match_state(state.id()) {
309                let mut pids = vec![];
310                for i in 0..dfa.match_len(state.id()) {
311                    pids.push(dfa.match_pattern(state.id(), i));
312                }
313                matching
314                    .entry(pids)
315                    .or_insert(StateSet::empty())
316                    .add(state.id());
317            } else if dfa.is_quit_state(state.id()) {
318                is_quit.add(state.id());
319            } else {
320                no_match.add(state.id());
321            }
322        }
323
324        let mut sets: Vec<StateSet> =
325            matching.into_iter().map(|(_, set)| set).collect();
326        sets.push(no_match);
327        sets.push(is_quit);
328        sets
329    }
330
331    fn incoming_transitions(dfa: &dense::OwnedDFA) -> Vec<Vec<Vec<StateID>>> {
332        let mut incoming = vec![];
333        for _ in dfa.states() {
334            incoming.push(vec![vec![]; dfa.alphabet_len()]);
335        }
336        for state in dfa.states() {
337            for (b, next) in state.transitions() {
338                incoming[dfa.to_index(next)][b.as_usize()].push(state.id());
339            }
340        }
341        incoming
342    }
343}
344
345impl StateSet {
346    fn empty() -> StateSet {
347        StateSet { ids: Rc::new(RefCell::new(vec![])) }
348    }
349
350    fn add(&mut self, id: StateID) {
351        self.ids.borrow_mut().push(id);
352    }
353
354    fn min(&self) -> StateID {
355        self.ids.borrow()[0]
356    }
357
358    fn canonicalize(&mut self) {
359        self.ids.borrow_mut().sort();
360        self.ids.borrow_mut().dedup();
361    }
362
363    fn clear(&mut self) {
364        self.ids.borrow_mut().clear();
365    }
366
367    fn len(&self) -> usize {
368        self.ids.borrow().len()
369    }
370
371    fn is_empty(&self) -> bool {
372        self.len() == 0
373    }
374
375    fn deep_clone(&self) -> StateSet {
376        let ids = self.ids.borrow().iter().cloned().collect();
377        StateSet { ids: Rc::new(RefCell::new(ids)) }
378    }
379
380    fn iter<F: FnMut(StateID)>(&self, mut f: F) {
381        for &id in self.ids.borrow().iter() {
382            f(id);
383        }
384    }
385
386    fn intersection(&self, other: &StateSet, dest: &mut StateSet) {
387        dest.clear();
388        if self.is_empty() || other.is_empty() {
389            return;
390        }
391
392        let (seta, setb) = (self.ids.borrow(), other.ids.borrow());
393        let (mut ita, mut itb) = (seta.iter().cloned(), setb.iter().cloned());
394        let (mut a, mut b) = (ita.next().unwrap(), itb.next().unwrap());
395        loop {
396            if a == b {
397                dest.add(a);
398                a = match ita.next() {
399                    None => break,
400                    Some(a) => a,
401                };
402                b = match itb.next() {
403                    None => break,
404                    Some(b) => b,
405                };
406            } else if a < b {
407                a = match ita.next() {
408                    None => break,
409                    Some(a) => a,
410                };
411            } else {
412                b = match itb.next() {
413                    None => break,
414                    Some(b) => b,
415                };
416            }
417        }
418    }
419
420    fn subtract(&self, other: &StateSet, dest: &mut StateSet) {
421        dest.clear();
422        if self.is_empty() || other.is_empty() {
423            self.iter(|s| dest.add(s));
424            return;
425        }
426
427        let (seta, setb) = (self.ids.borrow(), other.ids.borrow());
428        let (mut ita, mut itb) = (seta.iter().cloned(), setb.iter().cloned());
429        let (mut a, mut b) = (ita.next().unwrap(), itb.next().unwrap());
430        loop {
431            if a == b {
432                a = match ita.next() {
433                    None => break,
434                    Some(a) => a,
435                };
436                b = match itb.next() {
437                    None => {
438                        dest.add(a);
439                        break;
440                    }
441                    Some(b) => b,
442                };
443            } else if a < b {
444                dest.add(a);
445                a = match ita.next() {
446                    None => break,
447                    Some(a) => a,
448                };
449            } else {
450                b = match itb.next() {
451                    None => {
452                        dest.add(a);
453                        break;
454                    }
455                    Some(b) => b,
456                };
457            }
458        }
459        for a in ita {
460            dest.add(a);
461        }
462    }
463}