| 1 |
use crate::{ |
| 2 |
hybrid::{ |
| 3 |
dfa::{Cache, OverlappingState, DFA}, |
| 4 |
id::LazyStateID, |
| 5 |
}, |
| 6 |
util::{ |
| 7 |
prefilter::Prefilter, |
| 8 |
search::{HalfMatch, Input, MatchError, Span}, |
| 9 |
}, |
| 10 |
}; |
| 11 |
|
| 12 |
#[inline(never)] |
| 13 |
pub(crate) fn find_fwd( |
| 14 |
dfa: &DFA, |
| 15 |
cache: &mut Cache, |
| 16 |
input: &Input<'_>, |
| 17 |
) -> Result<Option<HalfMatch>, MatchError> { |
| 18 |
if input.is_done() { |
| 19 |
return Ok(None); |
| 20 |
} |
| 21 |
let pre = if input.get_anchored().is_anchored() { |
| 22 |
None |
| 23 |
} else { |
| 24 |
dfa.get_config().get_prefilter() |
| 25 |
}; |
| 26 |
// So what we do here is specialize four different versions of 'find_fwd': |
| 27 |
// one for each of the combinations for 'has prefilter' and 'is earliest |
| 28 |
// search'. The reason for doing this is that both of these things require |
| 29 |
// branches and special handling in some code that can be very hot, |
| 30 |
// and shaving off as much as we can when we don't need it tends to be |
| 31 |
// beneficial in ad hoc benchmarks. To see these differences, you often |
| 32 |
// need a query with a high match count. In other words, specializing these |
| 33 |
// four routines *tends* to help latency more than throughput. |
| 34 |
if pre.is_some() { |
| 35 |
if input.get_earliest() { |
| 36 |
find_fwd_imp(dfa, cache, input, pre, true) |
| 37 |
} else { |
| 38 |
find_fwd_imp(dfa, cache, input, pre, false) |
| 39 |
} |
| 40 |
} else { |
| 41 |
if input.get_earliest() { |
| 42 |
find_fwd_imp(dfa, cache, input, None, true) |
| 43 |
} else { |
| 44 |
find_fwd_imp(dfa, cache, input, None, false) |
| 45 |
} |
| 46 |
} |
| 47 |
} |
| 48 |
|
| 49 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 50 |
fn find_fwd_imp( |
| 51 |
dfa: &DFA, |
| 52 |
cache: &mut Cache, |
| 53 |
input: &Input<'_>, |
| 54 |
pre: Option<&'_ Prefilter>, |
| 55 |
earliest: bool, |
| 56 |
) -> Result<Option<HalfMatch>, MatchError> { |
| 57 |
// See 'prefilter_restart' docs for explanation. |
| 58 |
let universal_start = dfa.get_nfa().look_set_prefix_any().is_empty(); |
| 59 |
let mut mat = None; |
| 60 |
let mut sid = init_fwd(dfa, cache, input)?; |
| 61 |
let mut at = input.start(); |
| 62 |
// This could just be a closure, but then I think it would be unsound |
| 63 |
// because it would need to be safe to invoke. This way, the lack of safety |
| 64 |
// is clearer in the code below. |
| 65 |
macro_rules! next_unchecked { |
| 66 |
($sid:expr, $at:expr) => {{ |
| 67 |
let byte = *input.haystack().get_unchecked($at); |
| 68 |
dfa.next_state_untagged_unchecked(cache, $sid, byte) |
| 69 |
}}; |
| 70 |
} |
| 71 |
|
| 72 |
if let Some(ref pre) = pre { |
| 73 |
let span = Span::from(at..input.end()); |
| 74 |
match pre.find(input.haystack(), span) { |
| 75 |
None => return Ok(mat), |
| 76 |
Some(ref span) => { |
| 77 |
at = span.start; |
| 78 |
if !universal_start { |
| 79 |
sid = prefilter_restart(dfa, cache, &input, at)?; |
| 80 |
} |
| 81 |
} |
| 82 |
} |
| 83 |
} |
| 84 |
cache.search_start(at); |
| 85 |
while at < input.end() { |
| 86 |
if sid.is_tagged() { |
| 87 |
cache.search_update(at); |
| 88 |
sid = dfa |
| 89 |
.next_state(cache, sid, input.haystack()[at]) |
| 90 |
.map_err(|_| gave_up(at))?; |
| 91 |
} else { |
| 92 |
// SAFETY: There are two safety invariants we need to uphold |
| 93 |
// here in the loops below: that 'sid' and 'prev_sid' are valid |
| 94 |
// state IDs for this DFA, and that 'at' is a valid index into |
| 95 |
// 'haystack'. For the former, we rely on the invariant that |
| 96 |
// next_state* and start_state_forward always returns a valid state |
| 97 |
// ID (given a valid state ID in the former case), and that we are |
| 98 |
// only at this place in the code if 'sid' is untagged. Moreover, |
| 99 |
// every call to next_state_untagged_unchecked below is guarded by |
| 100 |
// a check that sid is untagged. For the latter safety invariant, |
| 101 |
// we always guard unchecked access with a check that 'at' is less |
| 102 |
// than 'end', where 'end <= haystack.len()'. In the unrolled loop |
| 103 |
// below, we ensure that 'at' is always in bounds. |
| 104 |
// |
| 105 |
// PERF: For justification of omitting bounds checks, it gives us a |
| 106 |
// ~10% bump in search time. This was used for a benchmark: |
| 107 |
// |
| 108 |
// regex-cli find hybrid dfa @bigfile '(?m)^.+$' -UBb |
| 109 |
// |
| 110 |
// PERF: For justification for the loop unrolling, we use a few |
| 111 |
// different tests: |
| 112 |
// |
| 113 |
// regex-cli find hybrid dfa @$bigfile '\w{50}' -UBb |
| 114 |
// regex-cli find hybrid dfa @$bigfile '(?m)^.+$' -UBb |
| 115 |
// regex-cli find hybrid dfa @$bigfile 'ZQZQZQZQ' -UBb |
| 116 |
// |
| 117 |
// And there are three different configurations: |
| 118 |
// |
| 119 |
// nounroll: this entire 'else' block vanishes and we just |
| 120 |
// always use 'dfa.next_state(..)'. |
| 121 |
// unroll1: just the outer loop below |
| 122 |
// unroll2: just the inner loop below |
| 123 |
// unroll3: both the outer and inner loops below |
| 124 |
// |
| 125 |
// This results in a matrix of timings for each of the above |
| 126 |
// regexes with each of the above unrolling configurations: |
| 127 |
// |
| 128 |
// '\w{50}' '(?m)^.+$' 'ZQZQZQZQ' |
| 129 |
// nounroll 1.51s 2.34s 1.51s |
| 130 |
// unroll1 1.53s 2.32s 1.56s |
| 131 |
// unroll2 2.22s 1.50s 0.61s |
| 132 |
// unroll3 1.67s 1.45s 0.61s |
| 133 |
// |
| 134 |
// Ideally we'd be able to find a configuration that yields the |
| 135 |
// best time for all regexes, but alas we settle for unroll3 that |
| 136 |
// gives us *almost* the best for '\w{50}' and the best for the |
| 137 |
// other two regexes. |
| 138 |
// |
| 139 |
// So what exactly is going on here? The first unrolling (grouping |
| 140 |
// together runs of untagged transitions) specifically targets |
| 141 |
// our choice of representation. The second unrolling (grouping |
| 142 |
// together runs of self-transitions) specifically targets a common |
| 143 |
// DFA topology. Let's dig in a little bit by looking at our |
| 144 |
// regexes: |
| 145 |
// |
| 146 |
// '\w{50}': This regex spends a lot of time outside of the DFA's |
| 147 |
// start state matching some part of the '\w' repetition. This |
| 148 |
// means that it's a bit of a worst case for loop unrolling that |
| 149 |
// targets self-transitions since the self-transitions in '\w{50}' |
| 150 |
// are not particularly active for this haystack. However, the |
| 151 |
// first unrolling (grouping together untagged transitions) |
| 152 |
// does apply quite well here since very few transitions hit |
| 153 |
// match/dead/quit/unknown states. It is however worth mentioning |
| 154 |
// that if start states are configured to be tagged (which you |
| 155 |
// typically want to do if you have a prefilter), then this regex |
| 156 |
// actually slows way down because it is constantly ping-ponging |
| 157 |
// out of the unrolled loop and into the handling of a tagged start |
| 158 |
// state below. But when start states aren't tagged, the unrolled |
| 159 |
// loop stays hot. (This is why it's imperative that start state |
| 160 |
// tagging be disabled when there isn't a prefilter!) |
| 161 |
// |
| 162 |
// '(?m)^.+$': There are two important aspects of this regex: 1) |
| 163 |
// on this haystack, its match count is very high, much higher |
| 164 |
// than the other two regex and 2) it spends the vast majority |
| 165 |
// of its time matching '.+'. Since Unicode mode is disabled, |
| 166 |
// this corresponds to repeatedly following self transitions for |
| 167 |
// the vast majority of the input. This does benefit from the |
| 168 |
// untagged unrolling since most of the transitions will be to |
| 169 |
// untagged states, but the untagged unrolling does more work than |
| 170 |
// what is actually required. Namely, it has to keep track of the |
| 171 |
// previous and next state IDs, which I guess requires a bit more |
| 172 |
// shuffling. This is supported by the fact that nounroll+unroll1 |
| 173 |
// are both slower than unroll2+unroll3, where the latter has a |
| 174 |
// loop unrolling that specifically targets self-transitions. |
| 175 |
// |
| 176 |
// 'ZQZQZQZQ': This one is very similar to '(?m)^.+$' because it |
| 177 |
// spends the vast majority of its time in self-transitions for |
| 178 |
// the (implicit) unanchored prefix. The main difference with |
| 179 |
// '(?m)^.+$' is that it has a much lower match count. So there |
| 180 |
// isn't much time spent in the overhead of reporting matches. This |
| 181 |
// is the primary explainer in the perf difference here. We include |
| 182 |
// this regex and the former to make sure we have comparison points |
| 183 |
// with high and low match counts. |
| 184 |
// |
| 185 |
// NOTE: I used 'OpenSubtitles2018.raw.sample.en' for 'bigfile'. |
| 186 |
// |
| 187 |
// NOTE: In a follow-up, it turns out that the "inner" loop |
| 188 |
// mentioned above was a pretty big pessimization in some other |
| 189 |
// cases. Namely, it resulted in too much ping-ponging into and out |
| 190 |
// of the loop, which resulted in nearly ~2x regressions in search |
| 191 |
// time when compared to the originally lazy DFA in the regex crate. |
| 192 |
// So I've removed the second loop unrolling that targets the |
| 193 |
// self-transition case. |
| 194 |
let mut prev_sid = sid; |
| 195 |
while at < input.end() { |
| 196 |
prev_sid = unsafe { next_unchecked!(sid, at) }; |
| 197 |
if prev_sid.is_tagged() || at + 3 >= input.end() { |
| 198 |
core::mem::swap(&mut prev_sid, &mut sid); |
| 199 |
break; |
| 200 |
} |
| 201 |
at += 1; |
| 202 |
|
| 203 |
sid = unsafe { next_unchecked!(prev_sid, at) }; |
| 204 |
if sid.is_tagged() { |
| 205 |
break; |
| 206 |
} |
| 207 |
at += 1; |
| 208 |
|
| 209 |
prev_sid = unsafe { next_unchecked!(sid, at) }; |
| 210 |
if prev_sid.is_tagged() { |
| 211 |
core::mem::swap(&mut prev_sid, &mut sid); |
| 212 |
break; |
| 213 |
} |
| 214 |
at += 1; |
| 215 |
|
| 216 |
sid = unsafe { next_unchecked!(prev_sid, at) }; |
| 217 |
if sid.is_tagged() { |
| 218 |
break; |
| 219 |
} |
| 220 |
at += 1; |
| 221 |
} |
| 222 |
// If we quit out of the code above with an unknown state ID at |
| 223 |
// any point, then we need to re-compute that transition using |
| 224 |
// 'next_state', which will do NFA powerset construction for us. |
| 225 |
if sid.is_unknown() { |
| 226 |
cache.search_update(at); |
| 227 |
sid = dfa |
| 228 |
.next_state(cache, prev_sid, input.haystack()[at]) |
| 229 |
.map_err(|_| gave_up(at))?; |
| 230 |
} |
| 231 |
} |
| 232 |
if sid.is_tagged() { |
| 233 |
if sid.is_start() { |
| 234 |
if let Some(ref pre) = pre { |
| 235 |
let span = Span::from(at..input.end()); |
| 236 |
match pre.find(input.haystack(), span) { |
| 237 |
None => { |
| 238 |
cache.search_finish(span.end); |
| 239 |
return Ok(mat); |
| 240 |
} |
| 241 |
Some(ref span) => { |
| 242 |
// We want to skip any update to 'at' below |
| 243 |
// at the end of this iteration and just |
| 244 |
// jump immediately back to the next state |
| 245 |
// transition at the leading position of the |
| 246 |
// candidate match. |
| 247 |
// |
| 248 |
// ... but only if we actually made progress |
| 249 |
// with our prefilter, otherwise if the start |
| 250 |
// state has a self-loop, we can get stuck. |
| 251 |
if span.start > at { |
| 252 |
at = span.start; |
| 253 |
if !universal_start { |
| 254 |
sid = prefilter_restart( |
| 255 |
dfa, cache, &input, at, |
| 256 |
)?; |
| 257 |
} |
| 258 |
continue; |
| 259 |
} |
| 260 |
} |
| 261 |
} |
| 262 |
} |
| 263 |
} else if sid.is_match() { |
| 264 |
let pattern = dfa.match_pattern(cache, sid, 0); |
| 265 |
// Since slice ranges are inclusive at the beginning and |
| 266 |
// exclusive at the end, and since forward searches report |
| 267 |
// the end, we can return 'at' as-is. This only works because |
| 268 |
// matches are delayed by 1 byte. So by the time we observe a |
| 269 |
// match, 'at' has already been set to 1 byte past the actual |
| 270 |
// match location, which is precisely the exclusive ending |
| 271 |
// bound of the match. |
| 272 |
mat = Some(HalfMatch::new(pattern, at)); |
| 273 |
if earliest { |
| 274 |
cache.search_finish(at); |
| 275 |
return Ok(mat); |
| 276 |
} |
| 277 |
} else if sid.is_dead() { |
| 278 |
cache.search_finish(at); |
| 279 |
return Ok(mat); |
| 280 |
} else if sid.is_quit() { |
| 281 |
cache.search_finish(at); |
| 282 |
return Err(MatchError::quit(input.haystack()[at], at)); |
| 283 |
} else { |
| 284 |
debug_assert!(sid.is_unknown()); |
| 285 |
unreachable!("sid being unknown is a bug"); |
| 286 |
} |
| 287 |
} |
| 288 |
at += 1; |
| 289 |
} |
| 290 |
eoi_fwd(dfa, cache, input, &mut sid, &mut mat)?; |
| 291 |
cache.search_finish(input.end()); |
| 292 |
Ok(mat) |
| 293 |
} |
| 294 |
|
| 295 |
#[inline(never)] |
| 296 |
pub(crate) fn find_rev( |
| 297 |
dfa: &DFA, |
| 298 |
cache: &mut Cache, |
| 299 |
input: &Input<'_>, |
| 300 |
) -> Result<Option<HalfMatch>, MatchError> { |
| 301 |
if input.is_done() { |
| 302 |
return Ok(None); |
| 303 |
} |
| 304 |
if input.get_earliest() { |
| 305 |
find_rev_imp(dfa, cache, input, true) |
| 306 |
} else { |
| 307 |
find_rev_imp(dfa, cache, input, false) |
| 308 |
} |
| 309 |
} |
| 310 |
|
| 311 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 312 |
fn find_rev_imp( |
| 313 |
dfa: &DFA, |
| 314 |
cache: &mut Cache, |
| 315 |
input: &Input<'_>, |
| 316 |
earliest: bool, |
| 317 |
) -> Result<Option<HalfMatch>, MatchError> { |
| 318 |
let mut mat = None; |
| 319 |
let mut sid = init_rev(dfa, cache, input)?; |
| 320 |
// In reverse search, the loop below can't handle the case of searching an |
| 321 |
// empty slice. Ideally we could write something congruent to the forward |
| 322 |
// search, i.e., 'while at >= start', but 'start' might be 0. Since we use |
| 323 |
// an unsigned offset, 'at >= 0' is trivially always true. We could avoid |
| 324 |
// this extra case handling by using a signed offset, but Rust makes it |
| 325 |
// annoying to do. So... We just handle the empty case separately. |
| 326 |
if input.start() == input.end() { |
| 327 |
eoi_rev(dfa, cache, input, &mut sid, &mut mat)?; |
| 328 |
return Ok(mat); |
| 329 |
} |
| 330 |
|
| 331 |
let mut at = input.end() - 1; |
| 332 |
macro_rules! next_unchecked { |
| 333 |
($sid:expr, $at:expr) => {{ |
| 334 |
let byte = *input.haystack().get_unchecked($at); |
| 335 |
dfa.next_state_untagged_unchecked(cache, $sid, byte) |
| 336 |
}}; |
| 337 |
} |
| 338 |
cache.search_start(at); |
| 339 |
loop { |
| 340 |
if sid.is_tagged() { |
| 341 |
cache.search_update(at); |
| 342 |
sid = dfa |
| 343 |
.next_state(cache, sid, input.haystack()[at]) |
| 344 |
.map_err(|_| gave_up(at))?; |
| 345 |
} else { |
| 346 |
// SAFETY: See comments in 'find_fwd' for a safety argument. |
| 347 |
// |
| 348 |
// PERF: The comments in 'find_fwd' also provide a justification |
| 349 |
// from a performance perspective as to 1) why we elide bounds |
| 350 |
// checks and 2) why we do a specialized version of unrolling |
| 351 |
// below. The reverse search does have a slightly different |
| 352 |
// consideration in that most reverse searches tend to be |
| 353 |
// anchored and on shorter haystacks. However, this still makes a |
| 354 |
// difference. Take this command for example: |
| 355 |
// |
| 356 |
// regex-cli find hybrid regex @$bigfile '(?m)^.+$' -UBb |
| 357 |
// |
| 358 |
// (Notice that we use 'find hybrid regex', not 'find hybrid dfa' |
| 359 |
// like in the justification for the forward direction. The 'regex' |
| 360 |
// sub-command will find start-of-match and thus run the reverse |
| 361 |
// direction.) |
| 362 |
// |
| 363 |
// Without unrolling below, the above command takes around 3.76s. |
| 364 |
// But with the unrolling below, we get down to 2.55s. If we keep |
| 365 |
// the unrolling but add in bounds checks, then we get 2.86s. |
| 366 |
// |
| 367 |
// NOTE: I used 'OpenSubtitles2018.raw.sample.en' for 'bigfile'. |
| 368 |
let mut prev_sid = sid; |
| 369 |
while at >= input.start() { |
| 370 |
prev_sid = unsafe { next_unchecked!(sid, at) }; |
| 371 |
if prev_sid.is_tagged() |
| 372 |
|| at <= input.start().saturating_add(3) |
| 373 |
{ |
| 374 |
core::mem::swap(&mut prev_sid, &mut sid); |
| 375 |
break; |
| 376 |
} |
| 377 |
at -= 1; |
| 378 |
|
| 379 |
sid = unsafe { next_unchecked!(prev_sid, at) }; |
| 380 |
if sid.is_tagged() { |
| 381 |
break; |
| 382 |
} |
| 383 |
at -= 1; |
| 384 |
|
| 385 |
prev_sid = unsafe { next_unchecked!(sid, at) }; |
| 386 |
if prev_sid.is_tagged() { |
| 387 |
core::mem::swap(&mut prev_sid, &mut sid); |
| 388 |
break; |
| 389 |
} |
| 390 |
at -= 1; |
| 391 |
|
| 392 |
sid = unsafe { next_unchecked!(prev_sid, at) }; |
| 393 |
if sid.is_tagged() { |
| 394 |
break; |
| 395 |
} |
| 396 |
at -= 1; |
| 397 |
} |
| 398 |
// If we quit out of the code above with an unknown state ID at |
| 399 |
// any point, then we need to re-compute that transition using |
| 400 |
// 'next_state', which will do NFA powerset construction for us. |
| 401 |
if sid.is_unknown() { |
| 402 |
cache.search_update(at); |
| 403 |
sid = dfa |
| 404 |
.next_state(cache, prev_sid, input.haystack()[at]) |
| 405 |
.map_err(|_| gave_up(at))?; |
| 406 |
} |
| 407 |
} |
| 408 |
if sid.is_tagged() { |
| 409 |
if sid.is_start() { |
| 410 |
// do nothing |
| 411 |
} else if sid.is_match() { |
| 412 |
let pattern = dfa.match_pattern(cache, sid, 0); |
| 413 |
// Since reverse searches report the beginning of a match |
| 414 |
// and the beginning is inclusive (not exclusive like the |
| 415 |
// end of a match), we add 1 to make it inclusive. |
| 416 |
mat = Some(HalfMatch::new(pattern, at + 1)); |
| 417 |
if earliest { |
| 418 |
cache.search_finish(at); |
| 419 |
return Ok(mat); |
| 420 |
} |
| 421 |
} else if sid.is_dead() { |
| 422 |
cache.search_finish(at); |
| 423 |
return Ok(mat); |
| 424 |
} else if sid.is_quit() { |
| 425 |
cache.search_finish(at); |
| 426 |
return Err(MatchError::quit(input.haystack()[at], at)); |
| 427 |
} else { |
| 428 |
debug_assert!(sid.is_unknown()); |
| 429 |
unreachable!("sid being unknown is a bug"); |
| 430 |
} |
| 431 |
} |
| 432 |
if at == input.start() { |
| 433 |
break; |
| 434 |
} |
| 435 |
at -= 1; |
| 436 |
} |
| 437 |
cache.search_finish(input.start()); |
| 438 |
eoi_rev(dfa, cache, input, &mut sid, &mut mat)?; |
| 439 |
Ok(mat) |
| 440 |
} |
| 441 |
|
| 442 |
#[inline(never)] |
| 443 |
pub(crate) fn find_overlapping_fwd( |
| 444 |
dfa: &DFA, |
| 445 |
cache: &mut Cache, |
| 446 |
input: &Input<'_>, |
| 447 |
state: &mut OverlappingState, |
| 448 |
) -> Result<(), MatchError> { |
| 449 |
state.mat = None; |
| 450 |
if input.is_done() { |
| 451 |
return Ok(()); |
| 452 |
} |
| 453 |
let pre = if input.get_anchored().is_anchored() { |
| 454 |
None |
| 455 |
} else { |
| 456 |
dfa.get_config().get_prefilter() |
| 457 |
}; |
| 458 |
if pre.is_some() { |
| 459 |
find_overlapping_fwd_imp(dfa, cache, input, pre, state) |
| 460 |
} else { |
| 461 |
find_overlapping_fwd_imp(dfa, cache, input, None, state) |
| 462 |
} |
| 463 |
} |
| 464 |
|
| 465 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 466 |
fn find_overlapping_fwd_imp( |
| 467 |
dfa: &DFA, |
| 468 |
cache: &mut Cache, |
| 469 |
input: &Input<'_>, |
| 470 |
pre: Option<&'_ Prefilter>, |
| 471 |
state: &mut OverlappingState, |
| 472 |
) -> Result<(), MatchError> { |
| 473 |
// See 'prefilter_restart' docs for explanation. |
| 474 |
let universal_start = dfa.get_nfa().look_set_prefix_any().is_empty(); |
| 475 |
let mut sid = match state.id { |
| 476 |
None => { |
| 477 |
state.at = input.start(); |
| 478 |
init_fwd(dfa, cache, input)? |
| 479 |
} |
| 480 |
Some(sid) => { |
| 481 |
if let Some(match_index) = state.next_match_index { |
| 482 |
let match_len = dfa.match_len(cache, sid); |
| 483 |
if match_index < match_len { |
| 484 |
state.next_match_index = Some(match_index + 1); |
| 485 |
let pattern = dfa.match_pattern(cache, sid, match_index); |
| 486 |
state.mat = Some(HalfMatch::new(pattern, state.at)); |
| 487 |
return Ok(()); |
| 488 |
} |
| 489 |
} |
| 490 |
// Once we've reported all matches at a given position, we need to |
| 491 |
// advance the search to the next position. |
| 492 |
state.at += 1; |
| 493 |
if state.at > input.end() { |
| 494 |
return Ok(()); |
| 495 |
} |
| 496 |
sid |
| 497 |
} |
| 498 |
}; |
| 499 |
|
| 500 |
// NOTE: We don't optimize the crap out of this routine primarily because |
| 501 |
// it seems like most overlapping searches will have higher match counts, |
| 502 |
// and thus, throughput is perhaps not as important. But if you have a use |
| 503 |
// case for something faster, feel free to file an issue. |
| 504 |
cache.search_start(state.at); |
| 505 |
while state.at < input.end() { |
| 506 |
sid = dfa |
| 507 |
.next_state(cache, sid, input.haystack()[state.at]) |
| 508 |
.map_err(|_| gave_up(state.at))?; |
| 509 |
if sid.is_tagged() { |
| 510 |
state.id = Some(sid); |
| 511 |
if sid.is_start() { |
| 512 |
if let Some(ref pre) = pre { |
| 513 |
let span = Span::from(state.at..input.end()); |
| 514 |
match pre.find(input.haystack(), span) { |
| 515 |
None => return Ok(()), |
| 516 |
Some(ref span) => { |
| 517 |
if span.start > state.at { |
| 518 |
state.at = span.start; |
| 519 |
if !universal_start { |
| 520 |
sid = prefilter_restart( |
| 521 |
dfa, cache, &input, state.at, |
| 522 |
)?; |
| 523 |
} |
| 524 |
continue; |
| 525 |
} |
| 526 |
} |
| 527 |
} |
| 528 |
} |
| 529 |
} else if sid.is_match() { |
| 530 |
state.next_match_index = Some(1); |
| 531 |
let pattern = dfa.match_pattern(cache, sid, 0); |
| 532 |
state.mat = Some(HalfMatch::new(pattern, state.at)); |
| 533 |
cache.search_finish(state.at); |
| 534 |
return Ok(()); |
| 535 |
} else if sid.is_dead() { |
| 536 |
cache.search_finish(state.at); |
| 537 |
return Ok(()); |
| 538 |
} else if sid.is_quit() { |
| 539 |
cache.search_finish(state.at); |
| 540 |
return Err(MatchError::quit( |
| 541 |
input.haystack()[state.at], |
| 542 |
state.at, |
| 543 |
)); |
| 544 |
} else { |
| 545 |
debug_assert!(sid.is_unknown()); |
| 546 |
unreachable!("sid being unknown is a bug"); |
| 547 |
} |
| 548 |
} |
| 549 |
state.at += 1; |
| 550 |
cache.search_update(state.at); |
| 551 |
} |
| 552 |
|
| 553 |
let result = eoi_fwd(dfa, cache, input, &mut sid, &mut state.mat); |
| 554 |
state.id = Some(sid); |
| 555 |
if state.mat.is_some() { |
| 556 |
// '1' is always correct here since if we get to this point, this |
| 557 |
// always corresponds to the first (index '0') match discovered at |
| 558 |
// this position. So the next match to report at this position (if |
| 559 |
// it exists) is at index '1'. |
| 560 |
state.next_match_index = Some(1); |
| 561 |
} |
| 562 |
cache.search_finish(input.end()); |
| 563 |
result |
| 564 |
} |
| 565 |
|
| 566 |
#[inline(never)] |
| 567 |
pub(crate) fn find_overlapping_rev( |
| 568 |
dfa: &DFA, |
| 569 |
cache: &mut Cache, |
| 570 |
input: &Input<'_>, |
| 571 |
state: &mut OverlappingState, |
| 572 |
) -> Result<(), MatchError> { |
| 573 |
state.mat = None; |
| 574 |
if input.is_done() { |
| 575 |
return Ok(()); |
| 576 |
} |
| 577 |
let mut sid = match state.id { |
| 578 |
None => { |
| 579 |
let sid = init_rev(dfa, cache, input)?; |
| 580 |
state.id = Some(sid); |
| 581 |
if input.start() == input.end() { |
| 582 |
state.rev_eoi = true; |
| 583 |
} else { |
| 584 |
state.at = input.end() - 1; |
| 585 |
} |
| 586 |
sid |
| 587 |
} |
| 588 |
Some(sid) => { |
| 589 |
if let Some(match_index) = state.next_match_index { |
| 590 |
let match_len = dfa.match_len(cache, sid); |
| 591 |
if match_index < match_len { |
| 592 |
state.next_match_index = Some(match_index + 1); |
| 593 |
let pattern = dfa.match_pattern(cache, sid, match_index); |
| 594 |
state.mat = Some(HalfMatch::new(pattern, state.at)); |
| 595 |
return Ok(()); |
| 596 |
} |
| 597 |
} |
| 598 |
// Once we've reported all matches at a given position, we need |
| 599 |
// to advance the search to the next position. However, if we've |
| 600 |
// already followed the EOI transition, then we know we're done |
| 601 |
// with the search and there cannot be any more matches to report. |
| 602 |
if state.rev_eoi { |
| 603 |
return Ok(()); |
| 604 |
} else if state.at == input.start() { |
| 605 |
// At this point, we should follow the EOI transition. This |
| 606 |
// will cause us the skip the main loop below and fall through |
| 607 |
// to the final 'eoi_rev' transition. |
| 608 |
state.rev_eoi = true; |
| 609 |
} else { |
| 610 |
// We haven't hit the end of the search yet, so move on. |
| 611 |
state.at -= 1; |
| 612 |
} |
| 613 |
sid |
| 614 |
} |
| 615 |
}; |
| 616 |
cache.search_start(state.at); |
| 617 |
while !state.rev_eoi { |
| 618 |
sid = dfa |
| 619 |
.next_state(cache, sid, input.haystack()[state.at]) |
| 620 |
.map_err(|_| gave_up(state.at))?; |
| 621 |
if sid.is_tagged() { |
| 622 |
state.id = Some(sid); |
| 623 |
if sid.is_start() { |
| 624 |
// do nothing |
| 625 |
} else if sid.is_match() { |
| 626 |
state.next_match_index = Some(1); |
| 627 |
let pattern = dfa.match_pattern(cache, sid, 0); |
| 628 |
state.mat = Some(HalfMatch::new(pattern, state.at + 1)); |
| 629 |
cache.search_finish(state.at); |
| 630 |
return Ok(()); |
| 631 |
} else if sid.is_dead() { |
| 632 |
cache.search_finish(state.at); |
| 633 |
return Ok(()); |
| 634 |
} else if sid.is_quit() { |
| 635 |
cache.search_finish(state.at); |
| 636 |
return Err(MatchError::quit( |
| 637 |
input.haystack()[state.at], |
| 638 |
state.at, |
| 639 |
)); |
| 640 |
} else { |
| 641 |
debug_assert!(sid.is_unknown()); |
| 642 |
unreachable!("sid being unknown is a bug"); |
| 643 |
} |
| 644 |
} |
| 645 |
if state.at == input.start() { |
| 646 |
break; |
| 647 |
} |
| 648 |
state.at -= 1; |
| 649 |
cache.search_update(state.at); |
| 650 |
} |
| 651 |
|
| 652 |
let result = eoi_rev(dfa, cache, input, &mut sid, &mut state.mat); |
| 653 |
state.rev_eoi = true; |
| 654 |
state.id = Some(sid); |
| 655 |
if state.mat.is_some() { |
| 656 |
// '1' is always correct here since if we get to this point, this |
| 657 |
// always corresponds to the first (index '0') match discovered at |
| 658 |
// this position. So the next match to report at this position (if |
| 659 |
// it exists) is at index '1'. |
| 660 |
state.next_match_index = Some(1); |
| 661 |
} |
| 662 |
cache.search_finish(input.start()); |
| 663 |
result |
| 664 |
} |
| 665 |
|
| 666 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 667 |
fn init_fwd( |
| 668 |
dfa: &DFA, |
| 669 |
cache: &mut Cache, |
| 670 |
input: &Input<'_>, |
| 671 |
) -> Result<LazyStateID, MatchError> { |
| 672 |
let sid = dfa.start_state_forward(cache, input)?; |
| 673 |
// Start states can never be match states, since all matches are delayed |
| 674 |
// by 1 byte. |
| 675 |
debug_assert!(!sid.is_match()); |
| 676 |
Ok(sid) |
| 677 |
} |
| 678 |
|
| 679 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 680 |
fn init_rev( |
| 681 |
dfa: &DFA, |
| 682 |
cache: &mut Cache, |
| 683 |
input: &Input<'_>, |
| 684 |
) -> Result<LazyStateID, MatchError> { |
| 685 |
let sid = dfa.start_state_reverse(cache, input)?; |
| 686 |
// Start states can never be match states, since all matches are delayed |
| 687 |
// by 1 byte. |
| 688 |
debug_assert!(!sid.is_match()); |
| 689 |
Ok(sid) |
| 690 |
} |
| 691 |
|
| 692 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 693 |
fn eoi_fwd( |
| 694 |
dfa: &DFA, |
| 695 |
cache: &mut Cache, |
| 696 |
input: &Input<'_>, |
| 697 |
sid: &mut LazyStateID, |
| 698 |
mat: &mut Option<HalfMatch>, |
| 699 |
) -> Result<(), MatchError> { |
| 700 |
let sp = input.get_span(); |
| 701 |
match input.haystack().get(sp.end) { |
| 702 |
Some(&b) => { |
| 703 |
*sid = |
| 704 |
dfa.next_state(cache, *sid, b).map_err(|_| gave_up(sp.end))?; |
| 705 |
if sid.is_match() { |
| 706 |
let pattern = dfa.match_pattern(cache, *sid, 0); |
| 707 |
*mat = Some(HalfMatch::new(pattern, sp.end)); |
| 708 |
} else if sid.is_quit() { |
| 709 |
return Err(MatchError::quit(b, sp.end)); |
| 710 |
} |
| 711 |
} |
| 712 |
None => { |
| 713 |
*sid = dfa |
| 714 |
.next_eoi_state(cache, *sid) |
| 715 |
.map_err(|_| gave_up(input.haystack().len()))?; |
| 716 |
if sid.is_match() { |
| 717 |
let pattern = dfa.match_pattern(cache, *sid, 0); |
| 718 |
*mat = Some(HalfMatch::new(pattern, input.haystack().len())); |
| 719 |
} |
| 720 |
// N.B. We don't have to check 'is_quit' here because the EOI |
| 721 |
// transition can never lead to a quit state. |
| 722 |
debug_assert!(!sid.is_quit()); |
| 723 |
} |
| 724 |
} |
| 725 |
Ok(()) |
| 726 |
} |
| 727 |
|
| 728 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 729 |
fn eoi_rev( |
| 730 |
dfa: &DFA, |
| 731 |
cache: &mut Cache, |
| 732 |
input: &Input<'_>, |
| 733 |
sid: &mut LazyStateID, |
| 734 |
mat: &mut Option<HalfMatch>, |
| 735 |
) -> Result<(), MatchError> { |
| 736 |
let sp = input.get_span(); |
| 737 |
if sp.start > 0 { |
| 738 |
let byte = input.haystack()[sp.start - 1]; |
| 739 |
*sid = dfa |
| 740 |
.next_state(cache, *sid, byte) |
| 741 |
.map_err(|_| gave_up(sp.start))?; |
| 742 |
if sid.is_match() { |
| 743 |
let pattern = dfa.match_pattern(cache, *sid, 0); |
| 744 |
*mat = Some(HalfMatch::new(pattern, sp.start)); |
| 745 |
} else if sid.is_quit() { |
| 746 |
return Err(MatchError::quit(byte, sp.start - 1)); |
| 747 |
} |
| 748 |
} else { |
| 749 |
*sid = |
| 750 |
dfa.next_eoi_state(cache, *sid).map_err(|_| gave_up(sp.start))?; |
| 751 |
if sid.is_match() { |
| 752 |
let pattern = dfa.match_pattern(cache, *sid, 0); |
| 753 |
*mat = Some(HalfMatch::new(pattern, 0)); |
| 754 |
} |
| 755 |
// N.B. We don't have to check 'is_quit' here because the EOI |
| 756 |
// transition can never lead to a quit state. |
| 757 |
debug_assert!(!sid.is_quit()); |
| 758 |
} |
| 759 |
Ok(()) |
| 760 |
} |
| 761 |
|
| 762 |
/// Re-compute the starting state that a DFA should be in after finding a |
| 763 |
/// prefilter candidate match at the position `at`. |
| 764 |
/// |
| 765 |
/// It is always correct to call this, but not always necessary. Namely, |
| 766 |
/// whenever the DFA has a universal start state, the DFA can remain in the |
| 767 |
/// start state that it was in when it ran the prefilter. Why? Because in that |
| 768 |
/// case, there is only one start state. |
| 769 |
/// |
| 770 |
/// When does a DFA have a universal start state? In precisely cases where |
| 771 |
/// it has no look-around assertions in its prefix. So for example, `\bfoo` |
| 772 |
/// does not have a universal start state because the start state depends on |
| 773 |
/// whether the byte immediately before the start position is a word byte or |
| 774 |
/// not. However, `foo\b` does have a universal start state because the word |
| 775 |
/// boundary does not appear in the pattern's prefix. |
| 776 |
/// |
| 777 |
/// So... most cases don't need this, but when a pattern doesn't have a |
| 778 |
/// universal start state, then after a prefilter candidate has been found, the |
| 779 |
/// current state *must* be re-litigated as if computing the start state at the |
| 780 |
/// beginning of the search because it might change. That is, not all start |
| 781 |
/// states are created equal. |
| 782 |
/// |
| 783 |
/// Why avoid it? Because while it's not super expensive, it isn't a trivial |
| 784 |
/// operation to compute the start state. It is much better to avoid it and |
| 785 |
/// just state in the current state if you know it to be correct. |
| 786 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 787 |
fn prefilter_restart( |
| 788 |
dfa: &DFA, |
| 789 |
cache: &mut Cache, |
| 790 |
input: &Input<'_>, |
| 791 |
at: usize, |
| 792 |
) -> Result<LazyStateID, MatchError> { |
| 793 |
let mut input = input.clone(); |
| 794 |
input.set_start(at); |
| 795 |
init_fwd(dfa, cache, &input) |
| 796 |
} |
| 797 |
|
| 798 |
/// A convenience routine for constructing a "gave up" match error. |
| 799 |
#[cfg_attr(feature = "perf-inline", inline(always))] |
| 800 |
fn gave_up(offset: usize) -> MatchError { |
| 801 |
MatchError::gave_up(offset) |
| 802 |
} |