1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
// This is the backtracking matching engine. It has the same exact capability
// as the full NFA simulation, except it is artificially restricted to small
// regexes on small inputs because of its memory requirements.
//
// In particular, this is a *bounded* backtracking engine. It retains worst
// case linear time by keeping track of the states that it has visited (using a
// bitmap). Namely, once a state is visited, it is never visited again. Since a
// state is keyed by `(instruction index, input index)`, we have that its time
// complexity is `O(mn)` (i.e., linear in the size of the search text).
//
// The backtracking engine can beat out the NFA simulation on small
// regexes/inputs because it doesn't have to keep track of multiple copies of
// the capture groups. In benchmarks, the backtracking engine is roughly twice
// as fast as the full NFA simulation. Note though that its performance doesn't
// scale, even if you're willing to live with the memory requirements. Namely,
// the bitset has to be zeroed on each execution, which becomes quite expensive
// on large bitsets.

use crate::exec::ProgramCache;
use crate::input::{Input, InputAt};
use crate::prog::{InstPtr, Program};
use crate::re_trait::Slot;

type Bits = u32;

const BIT_SIZE: usize = 32;
const MAX_SIZE_BYTES: usize = 256 * (1 << 10); // 256 KB

/// Returns true iff the given regex and input should be executed by this
/// engine with reasonable memory usage.
pub fn should_exec(num_insts: usize, text_len: usize) -> bool {
    // Total memory usage in bytes is determined by:
    //
    //   ((len(insts) * (len(input) + 1) + bits - 1) / bits) * (size_of(u32))
    //
    // The actual limit picked is pretty much a heuristic.
    // See: https://github.com/rust-lang/regex/issues/215
    let size = ((num_insts * (text_len + 1) + BIT_SIZE - 1) / BIT_SIZE) * 4;
    size <= MAX_SIZE_BYTES
}

/// A backtracking matching engine.
#[derive(Debug)]
pub struct Bounded<'a, 'm, 'r, 's, I> {
    prog: &'r Program,
    input: I,
    matches: &'m mut [bool],
    slots: &'s mut [Slot],
    m: &'a mut Cache,
}

/// Shared cached state between multiple invocations of a backtracking engine
/// in the same thread.
#[derive(Clone, Debug)]
pub struct Cache {
    jobs: Vec<Job>,
    visited: Vec<Bits>,
}

impl Cache {
    /// Create new empty cache for the backtracking engine.
    pub fn new(_prog: &Program) -> Self {
        Cache { jobs: vec![], visited: vec![] }
    }
}

/// A job is an explicit unit of stack space in the backtracking engine.
///
/// The "normal" representation is a single state transition, which corresponds
/// to an NFA state and a character in the input. However, the backtracking
/// engine must keep track of old capture group values. We use the explicit
/// stack to do it.
#[derive(Clone, Copy, Debug)]
enum Job {
    Inst { ip: InstPtr, at: InputAt },
    SaveRestore { slot: usize, old_pos: Option<usize> },
}

impl<'a, 'm, 'r, 's, I: Input> Bounded<'a, 'm, 'r, 's, I> {
    /// Execute the backtracking matching engine.
    ///
    /// If there's a match, `exec` returns `true` and populates the given
    /// captures accordingly.
    pub fn exec(
        prog: &'r Program,
        cache: &ProgramCache,
        matches: &'m mut [bool],
        slots: &'s mut [Slot],
        input: I,
        start: usize,
        end: usize,
    ) -> bool {
        let mut cache = cache.borrow_mut();
        let cache = &mut cache.backtrack;
        let start = input.at(start);
        let mut b = Bounded { prog, input, matches, slots, m: cache };
        b.exec_(start, end)
    }

    /// Clears the cache such that the backtracking engine can be executed
    /// on some input of fixed length.
    fn clear(&mut self) {
        // Reset the job memory so that we start fresh.
        self.m.jobs.clear();

        // Now we need to clear the bit state set.
        // We do this by figuring out how much space we need to keep track
        // of the states we've visited.
        // Then we reset all existing allocated space to 0.
        // Finally, we request more space if we need it.
        //
        // This is all a little circuitous, but doing this using unchecked
        // operations doesn't seem to have a measurable impact on performance.
        // (Probably because backtracking is limited to such small
        // inputs/regexes in the first place.)
        let visited_len =
            (self.prog.len() * (self.input.len() + 1) + BIT_SIZE - 1)
                / BIT_SIZE;
        self.m.visited.truncate(visited_len);
        for v in &mut self.m.visited {
            *v = 0;
        }
        if visited_len > self.m.visited.len() {
            let len = self.m.visited.len();
            self.m.visited.reserve_exact(visited_len - len);
            for _ in 0..(visited_len - len) {
                self.m.visited.push(0);
            }
        }
    }

    /// Start backtracking at the given position in the input, but also look
    /// for literal prefixes.
    fn exec_(&mut self, mut at: InputAt, end: usize) -> bool {
        self.clear();
        // If this is an anchored regex at the beginning of the input, then
        // we're either already done or we only need to try backtracking once.
        if self.prog.is_anchored_start {
            return if !at.is_start() { false } else { self.backtrack(at) };
        }
        let mut matched = false;
        loop {
            if !self.prog.prefixes.is_empty() {
                at = match self.input.prefix_at(&self.prog.prefixes, at) {
                    None => break,
                    Some(at) => at,
                };
            }
            matched = self.backtrack(at) || matched;
            if matched && self.prog.matches.len() == 1 {
                return true;
            }
            if at.pos() >= end {
                break;
            }
            at = self.input.at(at.next_pos());
        }
        matched
    }

    /// The main backtracking loop starting at the given input position.
    fn backtrack(&mut self, start: InputAt) -> bool {
        // N.B. We use an explicit stack to avoid recursion.
        // To avoid excessive pushing and popping, most transitions are handled
        // in the `step` helper function, which only pushes to the stack when
        // there's a capture or a branch.
        let mut matched = false;
        self.m.jobs.push(Job::Inst { ip: 0, at: start });
        while let Some(job) = self.m.jobs.pop() {
            match job {
                Job::Inst { ip, at } => {
                    if self.step(ip, at) {
                        // Only quit if we're matching one regex.
                        // If we're matching a regex set, then mush on and
                        // try to find other matches (if we want them).
                        if self.prog.matches.len() == 1 {
                            return true;
                        }
                        matched = true;
                    }
                }
                Job::SaveRestore { slot, old_pos } => {
                    if slot < self.slots.len() {
                        self.slots[slot] = old_pos;
                    }
                }
            }
        }
        matched
    }

    fn step(&mut self, mut ip: InstPtr, mut at: InputAt) -> bool {
        use crate::prog::Inst::*;
        loop {
            // This loop is an optimization to avoid constantly pushing/popping
            // from the stack. Namely, if we're pushing a job only to run it
            // next, avoid the push and just mutate `ip` (and possibly `at`)
            // in place.
            if self.has_visited(ip, at) {
                return false;
            }
            match self.prog[ip] {
                Match(slot) => {
                    if slot < self.matches.len() {
                        self.matches[slot] = true;
                    }
                    return true;
                }
                Save(ref inst) => {
                    if let Some(&old_pos) = self.slots.get(inst.slot) {
                        // If this path doesn't work out, then we save the old
                        // capture index (if one exists) in an alternate
                        // job. If the next path fails, then the alternate
                        // job is popped and the old capture index is restored.
                        self.m.jobs.push(Job::SaveRestore {
                            slot: inst.slot,
                            old_pos,
                        });
                        self.slots[inst.slot] = Some(at.pos());
                    }
                    ip = inst.goto;
                }
                Split(ref inst) => {
                    self.m.jobs.push(Job::Inst { ip: inst.goto2, at });
                    ip = inst.goto1;
                }
                EmptyLook(ref inst) => {
                    if self.input.is_empty_match(at, inst) {
                        ip = inst.goto;
                    } else {
                        return false;
                    }
                }
                Char(ref inst) => {
                    if inst.c == at.char() {
                        ip = inst.goto;
                        at = self.input.at(at.next_pos());
                    } else {
                        return false;
                    }
                }
                Ranges(ref inst) => {
                    if inst.matches(at.char()) {
                        ip = inst.goto;
                        at = self.input.at(at.next_pos());
                    } else {
                        return false;
                    }
                }
                Bytes(ref inst) => {
                    if let Some(b) = at.byte() {
                        if inst.matches(b) {
                            ip = inst.goto;
                            at = self.input.at(at.next_pos());
                            continue;
                        }
                    }
                    return false;
                }
            }
        }
    }

    fn has_visited(&mut self, ip: InstPtr, at: InputAt) -> bool {
        let k = ip * (self.input.len() + 1) + at.pos();
        let k1 = k / BIT_SIZE;
        let k2 = usize_to_u32(1 << (k & (BIT_SIZE - 1)));
        if self.m.visited[k1] & k2 == 0 {
            self.m.visited[k1] |= k2;
            false
        } else {
            true
        }
    }
}

fn usize_to_u32(n: usize) -> u32 {
    if (n as u64) > (::std::u32::MAX as u64) {
        panic!("BUG: {} is too big to fit into u32", n)
    }
    n as u32
}