| 1 |
use core::fmt::Debug; |
| 2 |
|
| 3 |
use alloc::{ |
| 4 |
boxed::Box, collections::BTreeMap, format, sync::Arc, vec, vec::Vec, |
| 5 |
}; |
| 6 |
|
| 7 |
use crate::{ |
| 8 |
packed::{ |
| 9 |
ext::Pointer, |
| 10 |
pattern::Patterns, |
| 11 |
vector::{FatVector, Vector}, |
| 12 |
}, |
| 13 |
util::int::U32, |
| 14 |
PatternID, |
| 15 |
}; |
| 16 |
|
| 17 |
/// A match type specialized to the Teddy implementations below. |
| 18 |
/// |
| 19 |
/// Essentially, instead of representing a match at byte offsets, we use |
| 20 |
/// raw pointers. This is because the implementations below operate on raw |
| 21 |
/// pointers, and so this is a more natural return type based on how the |
| 22 |
/// implementation works. |
| 23 |
/// |
| 24 |
/// Also, the `PatternID` used here is a `u16`. |
| 25 |
#[derive(Clone, Copy, Debug)] |
| 26 |
pub(crate) struct Match { |
| 27 |
pid: PatternID, |
| 28 |
start: *const u8, |
| 29 |
end: *const u8, |
| 30 |
} |
| 31 |
|
| 32 |
impl Match { |
| 33 |
/// Returns the ID of the pattern that matched. |
| 34 |
pub(crate) fn pattern(&self) -> PatternID { |
| 35 |
self.pid |
| 36 |
} |
| 37 |
|
| 38 |
/// Returns a pointer into the haystack at which the match starts. |
| 39 |
pub(crate) fn start(&self) -> *const u8 { |
| 40 |
self.start |
| 41 |
} |
| 42 |
|
| 43 |
/// Returns a pointer into the haystack at which the match ends. |
| 44 |
pub(crate) fn end(&self) -> *const u8 { |
| 45 |
self.end |
| 46 |
} |
| 47 |
} |
| 48 |
|
| 49 |
/// A "slim" Teddy implementation that is generic over both the vector type |
| 50 |
/// and the minimum length of the patterns being searched for. |
| 51 |
/// |
| 52 |
/// Only 1, 2, 3 and 4 bytes are supported as minimum lengths. |
| 53 |
#[derive(Clone, Debug)] |
| 54 |
pub(crate) struct Slim<V, const BYTES: usize> { |
| 55 |
/// A generic data structure for doing "slim" Teddy verification. |
| 56 |
teddy: Teddy<8>, |
| 57 |
/// The masks used as inputs to the shuffle operation to generate |
| 58 |
/// candidates (which are fed into the verification routines). |
| 59 |
masks: [Mask<V>; BYTES], |
| 60 |
} |
| 61 |
|
| 62 |
impl<V: Vector, const BYTES: usize> Slim<V, BYTES> { |
| 63 |
/// Create a new "slim" Teddy searcher for the given patterns. |
| 64 |
/// |
| 65 |
/// # Panics |
| 66 |
/// |
| 67 |
/// This panics when `BYTES` is any value other than 1, 2, 3 or 4. |
| 68 |
/// |
| 69 |
/// # Safety |
| 70 |
/// |
| 71 |
/// Callers must ensure that this is okay to call in the current target for |
| 72 |
/// the current CPU. |
| 73 |
#[inline(always)] |
| 74 |
pub(crate) unsafe fn new(patterns: Arc<Patterns>) -> Slim<V, BYTES> { |
| 75 |
assert!( |
| 76 |
1 <= BYTES && BYTES <= 4, |
| 77 |
"only 1, 2, 3 or 4 bytes are supported" |
| 78 |
); |
| 79 |
let teddy = Teddy::new(patterns); |
| 80 |
let masks = SlimMaskBuilder::from_teddy(&teddy); |
| 81 |
Slim { teddy, masks } |
| 82 |
} |
| 83 |
|
| 84 |
/// Returns the approximate total amount of heap used by this type, in |
| 85 |
/// units of bytes. |
| 86 |
#[inline(always)] |
| 87 |
pub(crate) fn memory_usage(&self) -> usize { |
| 88 |
self.teddy.memory_usage() |
| 89 |
} |
| 90 |
|
| 91 |
/// Returns the minimum length, in bytes, that a haystack must be in order |
| 92 |
/// to use it with this searcher. |
| 93 |
#[inline(always)] |
| 94 |
pub(crate) fn minimum_len(&self) -> usize { |
| 95 |
V::BYTES + (BYTES - 1) |
| 96 |
} |
| 97 |
} |
| 98 |
|
| 99 |
impl<V: Vector> Slim<V, 1> { |
| 100 |
/// Look for an occurrences of the patterns in this finder in the haystack |
| 101 |
/// given by the `start` and `end` pointers. |
| 102 |
/// |
| 103 |
/// If no match could be found, then `None` is returned. |
| 104 |
/// |
| 105 |
/// # Safety |
| 106 |
/// |
| 107 |
/// The given pointers representing the haystack must be valid to read |
| 108 |
/// from. They must also point to a region of memory that is at least the |
| 109 |
/// minimum length required by this searcher. |
| 110 |
/// |
| 111 |
/// Callers must ensure that this is okay to call in the current target for |
| 112 |
/// the current CPU. |
| 113 |
#[inline(always)] |
| 114 |
pub(crate) unsafe fn find( |
| 115 |
&self, |
| 116 |
start: *const u8, |
| 117 |
end: *const u8, |
| 118 |
) -> Option<Match> { |
| 119 |
let len = end.distance(start); |
| 120 |
debug_assert!(len >= self.minimum_len()); |
| 121 |
let mut cur = start; |
| 122 |
while cur <= end.sub(V::BYTES) { |
| 123 |
if let Some(m) = self.find_one(cur, end) { |
| 124 |
return Some(m); |
| 125 |
} |
| 126 |
cur = cur.add(V::BYTES); |
| 127 |
} |
| 128 |
if cur < end { |
| 129 |
cur = end.sub(V::BYTES); |
| 130 |
if let Some(m) = self.find_one(cur, end) { |
| 131 |
return Some(m); |
| 132 |
} |
| 133 |
} |
| 134 |
None |
| 135 |
} |
| 136 |
|
| 137 |
/// Look for a match starting at the `V::BYTES` at and after `cur`. If |
| 138 |
/// there isn't one, then `None` is returned. |
| 139 |
/// |
| 140 |
/// # Safety |
| 141 |
/// |
| 142 |
/// The given pointers representing the haystack must be valid to read |
| 143 |
/// from. They must also point to a region of memory that is at least the |
| 144 |
/// minimum length required by this searcher. |
| 145 |
/// |
| 146 |
/// Callers must ensure that this is okay to call in the current target for |
| 147 |
/// the current CPU. |
| 148 |
#[inline(always)] |
| 149 |
unsafe fn find_one( |
| 150 |
&self, |
| 151 |
cur: *const u8, |
| 152 |
end: *const u8, |
| 153 |
) -> Option<Match> { |
| 154 |
let c = self.candidate(cur); |
| 155 |
if !c.is_zero() { |
| 156 |
if let Some(m) = self.teddy.verify(cur, end, c) { |
| 157 |
return Some(m); |
| 158 |
} |
| 159 |
} |
| 160 |
None |
| 161 |
} |
| 162 |
|
| 163 |
/// Look for a candidate match (represented as a vector) starting at the |
| 164 |
/// `V::BYTES` at and after `cur`. If there isn't one, then a vector with |
| 165 |
/// all bits set to zero is returned. |
| 166 |
/// |
| 167 |
/// # Safety |
| 168 |
/// |
| 169 |
/// The given pointer representing the haystack must be valid to read |
| 170 |
/// from. |
| 171 |
/// |
| 172 |
/// Callers must ensure that this is okay to call in the current target for |
| 173 |
/// the current CPU. |
| 174 |
#[inline(always)] |
| 175 |
unsafe fn candidate(&self, cur: *const u8) -> V { |
| 176 |
let chunk = V::load_unaligned(cur); |
| 177 |
Mask::members1(chunk, self.masks) |
| 178 |
} |
| 179 |
} |
| 180 |
|
| 181 |
impl<V: Vector> Slim<V, 2> { |
| 182 |
/// See Slim<V, 1>::find. |
| 183 |
#[inline(always)] |
| 184 |
pub(crate) unsafe fn find( |
| 185 |
&self, |
| 186 |
start: *const u8, |
| 187 |
end: *const u8, |
| 188 |
) -> Option<Match> { |
| 189 |
let len = end.distance(start); |
| 190 |
debug_assert!(len >= self.minimum_len()); |
| 191 |
let mut cur = start.add(1); |
| 192 |
let mut prev0 = V::splat(0xFF); |
| 193 |
while cur <= end.sub(V::BYTES) { |
| 194 |
if let Some(m) = self.find_one(cur, end, &mut prev0) { |
| 195 |
return Some(m); |
| 196 |
} |
| 197 |
cur = cur.add(V::BYTES); |
| 198 |
} |
| 199 |
if cur < end { |
| 200 |
cur = end.sub(V::BYTES); |
| 201 |
prev0 = V::splat(0xFF); |
| 202 |
if let Some(m) = self.find_one(cur, end, &mut prev0) { |
| 203 |
return Some(m); |
| 204 |
} |
| 205 |
} |
| 206 |
None |
| 207 |
} |
| 208 |
|
| 209 |
/// See Slim<V, 1>::find_one. |
| 210 |
#[inline(always)] |
| 211 |
unsafe fn find_one( |
| 212 |
&self, |
| 213 |
cur: *const u8, |
| 214 |
end: *const u8, |
| 215 |
prev0: &mut V, |
| 216 |
) -> Option<Match> { |
| 217 |
let c = self.candidate(cur, prev0); |
| 218 |
if !c.is_zero() { |
| 219 |
if let Some(m) = self.teddy.verify(cur.sub(1), end, c) { |
| 220 |
return Some(m); |
| 221 |
} |
| 222 |
} |
| 223 |
None |
| 224 |
} |
| 225 |
|
| 226 |
/// See Slim<V, 1>::candidate. |
| 227 |
#[inline(always)] |
| 228 |
unsafe fn candidate(&self, cur: *const u8, prev0: &mut V) -> V { |
| 229 |
let chunk = V::load_unaligned(cur); |
| 230 |
let (res0, res1) = Mask::members2(chunk, self.masks); |
| 231 |
let res0prev0 = res0.shift_in_one_byte(*prev0); |
| 232 |
let res = res0prev0.and(res1); |
| 233 |
*prev0 = res0; |
| 234 |
res |
| 235 |
} |
| 236 |
} |
| 237 |
|
| 238 |
impl<V: Vector> Slim<V, 3> { |
| 239 |
/// See Slim<V, 1>::find. |
| 240 |
#[inline(always)] |
| 241 |
pub(crate) unsafe fn find( |
| 242 |
&self, |
| 243 |
start: *const u8, |
| 244 |
end: *const u8, |
| 245 |
) -> Option<Match> { |
| 246 |
let len = end.distance(start); |
| 247 |
debug_assert!(len >= self.minimum_len()); |
| 248 |
let mut cur = start.add(2); |
| 249 |
let mut prev0 = V::splat(0xFF); |
| 250 |
let mut prev1 = V::splat(0xFF); |
| 251 |
while cur <= end.sub(V::BYTES) { |
| 252 |
if let Some(m) = self.find_one(cur, end, &mut prev0, &mut prev1) { |
| 253 |
return Some(m); |
| 254 |
} |
| 255 |
cur = cur.add(V::BYTES); |
| 256 |
} |
| 257 |
if cur < end { |
| 258 |
cur = end.sub(V::BYTES); |
| 259 |
prev0 = V::splat(0xFF); |
| 260 |
prev1 = V::splat(0xFF); |
| 261 |
if let Some(m) = self.find_one(cur, end, &mut prev0, &mut prev1) { |
| 262 |
return Some(m); |
| 263 |
} |
| 264 |
} |
| 265 |
None |
| 266 |
} |
| 267 |
|
| 268 |
/// See Slim<V, 1>::find_one. |
| 269 |
#[inline(always)] |
| 270 |
unsafe fn find_one( |
| 271 |
&self, |
| 272 |
cur: *const u8, |
| 273 |
end: *const u8, |
| 274 |
prev0: &mut V, |
| 275 |
prev1: &mut V, |
| 276 |
) -> Option<Match> { |
| 277 |
let c = self.candidate(cur, prev0, prev1); |
| 278 |
if !c.is_zero() { |
| 279 |
if let Some(m) = self.teddy.verify(cur.sub(2), end, c) { |
| 280 |
return Some(m); |
| 281 |
} |
| 282 |
} |
| 283 |
None |
| 284 |
} |
| 285 |
|
| 286 |
/// See Slim<V, 1>::candidate. |
| 287 |
#[inline(always)] |
| 288 |
unsafe fn candidate( |
| 289 |
&self, |
| 290 |
cur: *const u8, |
| 291 |
prev0: &mut V, |
| 292 |
prev1: &mut V, |
| 293 |
) -> V { |
| 294 |
let chunk = V::load_unaligned(cur); |
| 295 |
let (res0, res1, res2) = Mask::members3(chunk, self.masks); |
| 296 |
let res0prev0 = res0.shift_in_two_bytes(*prev0); |
| 297 |
let res1prev1 = res1.shift_in_one_byte(*prev1); |
| 298 |
let res = res0prev0.and(res1prev1).and(res2); |
| 299 |
*prev0 = res0; |
| 300 |
*prev1 = res1; |
| 301 |
res |
| 302 |
} |
| 303 |
} |
| 304 |
|
| 305 |
impl<V: Vector> Slim<V, 4> { |
| 306 |
/// See Slim<V, 1>::find. |
| 307 |
#[inline(always)] |
| 308 |
pub(crate) unsafe fn find( |
| 309 |
&self, |
| 310 |
start: *const u8, |
| 311 |
end: *const u8, |
| 312 |
) -> Option<Match> { |
| 313 |
let len = end.distance(start); |
| 314 |
debug_assert!(len >= self.minimum_len()); |
| 315 |
let mut cur = start.add(3); |
| 316 |
let mut prev0 = V::splat(0xFF); |
| 317 |
let mut prev1 = V::splat(0xFF); |
| 318 |
let mut prev2 = V::splat(0xFF); |
| 319 |
while cur <= end.sub(V::BYTES) { |
| 320 |
if let Some(m) = |
| 321 |
self.find_one(cur, end, &mut prev0, &mut prev1, &mut prev2) |
| 322 |
{ |
| 323 |
return Some(m); |
| 324 |
} |
| 325 |
cur = cur.add(V::BYTES); |
| 326 |
} |
| 327 |
if cur < end { |
| 328 |
cur = end.sub(V::BYTES); |
| 329 |
prev0 = V::splat(0xFF); |
| 330 |
prev1 = V::splat(0xFF); |
| 331 |
prev2 = V::splat(0xFF); |
| 332 |
if let Some(m) = |
| 333 |
self.find_one(cur, end, &mut prev0, &mut prev1, &mut prev2) |
| 334 |
{ |
| 335 |
return Some(m); |
| 336 |
} |
| 337 |
} |
| 338 |
None |
| 339 |
} |
| 340 |
|
| 341 |
/// See Slim<V, 1>::find_one. |
| 342 |
#[inline(always)] |
| 343 |
unsafe fn find_one( |
| 344 |
&self, |
| 345 |
cur: *const u8, |
| 346 |
end: *const u8, |
| 347 |
prev0: &mut V, |
| 348 |
prev1: &mut V, |
| 349 |
prev2: &mut V, |
| 350 |
) -> Option<Match> { |
| 351 |
let c = self.candidate(cur, prev0, prev1, prev2); |
| 352 |
if !c.is_zero() { |
| 353 |
if let Some(m) = self.teddy.verify(cur.sub(3), end, c) { |
| 354 |
return Some(m); |
| 355 |
} |
| 356 |
} |
| 357 |
None |
| 358 |
} |
| 359 |
|
| 360 |
/// See Slim<V, 1>::candidate. |
| 361 |
#[inline(always)] |
| 362 |
unsafe fn candidate( |
| 363 |
&self, |
| 364 |
cur: *const u8, |
| 365 |
prev0: &mut V, |
| 366 |
prev1: &mut V, |
| 367 |
prev2: &mut V, |
| 368 |
) -> V { |
| 369 |
let chunk = V::load_unaligned(cur); |
| 370 |
let (res0, res1, res2, res3) = Mask::members4(chunk, self.masks); |
| 371 |
let res0prev0 = res0.shift_in_three_bytes(*prev0); |
| 372 |
let res1prev1 = res1.shift_in_two_bytes(*prev1); |
| 373 |
let res2prev2 = res2.shift_in_one_byte(*prev2); |
| 374 |
let res = res0prev0.and(res1prev1).and(res2prev2).and(res3); |
| 375 |
*prev0 = res0; |
| 376 |
*prev1 = res1; |
| 377 |
*prev2 = res2; |
| 378 |
res |
| 379 |
} |
| 380 |
} |
| 381 |
|
| 382 |
/// A "fat" Teddy implementation that is generic over both the vector type |
| 383 |
/// and the minimum length of the patterns being searched for. |
| 384 |
/// |
| 385 |
/// Only 1, 2, 3 and 4 bytes are supported as minimum lengths. |
| 386 |
#[derive(Clone, Debug)] |
| 387 |
pub(crate) struct Fat<V, const BYTES: usize> { |
| 388 |
/// A generic data structure for doing "fat" Teddy verification. |
| 389 |
teddy: Teddy<16>, |
| 390 |
/// The masks used as inputs to the shuffle operation to generate |
| 391 |
/// candidates (which are fed into the verification routines). |
| 392 |
masks: [Mask<V>; BYTES], |
| 393 |
} |
| 394 |
|
| 395 |
impl<V: FatVector, const BYTES: usize> Fat<V, BYTES> { |
| 396 |
/// Create a new "fat" Teddy searcher for the given patterns. |
| 397 |
/// |
| 398 |
/// # Panics |
| 399 |
/// |
| 400 |
/// This panics when `BYTES` is any value other than 1, 2, 3 or 4. |
| 401 |
/// |
| 402 |
/// # Safety |
| 403 |
/// |
| 404 |
/// Callers must ensure that this is okay to call in the current target for |
| 405 |
/// the current CPU. |
| 406 |
#[inline(always)] |
| 407 |
pub(crate) unsafe fn new(patterns: Arc<Patterns>) -> Fat<V, BYTES> { |
| 408 |
assert!( |
| 409 |
1 <= BYTES && BYTES <= 4, |
| 410 |
"only 1, 2, 3 or 4 bytes are supported" |
| 411 |
); |
| 412 |
let teddy = Teddy::new(patterns); |
| 413 |
let masks = FatMaskBuilder::from_teddy(&teddy); |
| 414 |
Fat { teddy, masks } |
| 415 |
} |
| 416 |
|
| 417 |
/// Returns the approximate total amount of heap used by this type, in |
| 418 |
/// units of bytes. |
| 419 |
#[inline(always)] |
| 420 |
pub(crate) fn memory_usage(&self) -> usize { |
| 421 |
self.teddy.memory_usage() |
| 422 |
} |
| 423 |
|
| 424 |
/// Returns the minimum length, in bytes, that a haystack must be in order |
| 425 |
/// to use it with this searcher. |
| 426 |
#[inline(always)] |
| 427 |
pub(crate) fn minimum_len(&self) -> usize { |
| 428 |
V::Half::BYTES + (BYTES - 1) |
| 429 |
} |
| 430 |
} |
| 431 |
|
| 432 |
impl<V: FatVector> Fat<V, 1> { |
| 433 |
/// Look for an occurrences of the patterns in this finder in the haystack |
| 434 |
/// given by the `start` and `end` pointers. |
| 435 |
/// |
| 436 |
/// If no match could be found, then `None` is returned. |
| 437 |
/// |
| 438 |
/// # Safety |
| 439 |
/// |
| 440 |
/// The given pointers representing the haystack must be valid to read |
| 441 |
/// from. They must also point to a region of memory that is at least the |
| 442 |
/// minimum length required by this searcher. |
| 443 |
/// |
| 444 |
/// Callers must ensure that this is okay to call in the current target for |
| 445 |
/// the current CPU. |
| 446 |
#[inline(always)] |
| 447 |
pub(crate) unsafe fn find( |
| 448 |
&self, |
| 449 |
start: *const u8, |
| 450 |
end: *const u8, |
| 451 |
) -> Option<Match> { |
| 452 |
let len = end.distance(start); |
| 453 |
debug_assert!(len >= self.minimum_len()); |
| 454 |
let mut cur = start; |
| 455 |
while cur <= end.sub(V::Half::BYTES) { |
| 456 |
if let Some(m) = self.find_one(cur, end) { |
| 457 |
return Some(m); |
| 458 |
} |
| 459 |
cur = cur.add(V::Half::BYTES); |
| 460 |
} |
| 461 |
if cur < end { |
| 462 |
cur = end.sub(V::Half::BYTES); |
| 463 |
if let Some(m) = self.find_one(cur, end) { |
| 464 |
return Some(m); |
| 465 |
} |
| 466 |
} |
| 467 |
None |
| 468 |
} |
| 469 |
|
| 470 |
/// Look for a match starting at the `V::BYTES` at and after `cur`. If |
| 471 |
/// there isn't one, then `None` is returned. |
| 472 |
/// |
| 473 |
/// # Safety |
| 474 |
/// |
| 475 |
/// The given pointers representing the haystack must be valid to read |
| 476 |
/// from. They must also point to a region of memory that is at least the |
| 477 |
/// minimum length required by this searcher. |
| 478 |
/// |
| 479 |
/// Callers must ensure that this is okay to call in the current target for |
| 480 |
/// the current CPU. |
| 481 |
#[inline(always)] |
| 482 |
unsafe fn find_one( |
| 483 |
&self, |
| 484 |
cur: *const u8, |
| 485 |
end: *const u8, |
| 486 |
) -> Option<Match> { |
| 487 |
let c = self.candidate(cur); |
| 488 |
if !c.is_zero() { |
| 489 |
if let Some(m) = self.teddy.verify(cur, end, c) { |
| 490 |
return Some(m); |
| 491 |
} |
| 492 |
} |
| 493 |
None |
| 494 |
} |
| 495 |
|
| 496 |
/// Look for a candidate match (represented as a vector) starting at the |
| 497 |
/// `V::BYTES` at and after `cur`. If there isn't one, then a vector with |
| 498 |
/// all bits set to zero is returned. |
| 499 |
/// |
| 500 |
/// # Safety |
| 501 |
/// |
| 502 |
/// The given pointer representing the haystack must be valid to read |
| 503 |
/// from. |
| 504 |
/// |
| 505 |
/// Callers must ensure that this is okay to call in the current target for |
| 506 |
/// the current CPU. |
| 507 |
#[inline(always)] |
| 508 |
unsafe fn candidate(&self, cur: *const u8) -> V { |
| 509 |
let chunk = V::load_half_unaligned(cur); |
| 510 |
Mask::members1(chunk, self.masks) |
| 511 |
} |
| 512 |
} |
| 513 |
|
| 514 |
impl<V: FatVector> Fat<V, 2> { |
| 515 |
/// See `Fat<V, 1>::find`. |
| 516 |
#[inline(always)] |
| 517 |
pub(crate) unsafe fn find( |
| 518 |
&self, |
| 519 |
start: *const u8, |
| 520 |
end: *const u8, |
| 521 |
) -> Option<Match> { |
| 522 |
let len = end.distance(start); |
| 523 |
debug_assert!(len >= self.minimum_len()); |
| 524 |
let mut cur = start.add(1); |
| 525 |
let mut prev0 = V::splat(0xFF); |
| 526 |
while cur <= end.sub(V::Half::BYTES) { |
| 527 |
if let Some(m) = self.find_one(cur, end, &mut prev0) { |
| 528 |
return Some(m); |
| 529 |
} |
| 530 |
cur = cur.add(V::Half::BYTES); |
| 531 |
} |
| 532 |
if cur < end { |
| 533 |
cur = end.sub(V::Half::BYTES); |
| 534 |
prev0 = V::splat(0xFF); |
| 535 |
if let Some(m) = self.find_one(cur, end, &mut prev0) { |
| 536 |
return Some(m); |
| 537 |
} |
| 538 |
} |
| 539 |
None |
| 540 |
} |
| 541 |
|
| 542 |
/// See `Fat<V, 1>::find_one`. |
| 543 |
#[inline(always)] |
| 544 |
unsafe fn find_one( |
| 545 |
&self, |
| 546 |
cur: *const u8, |
| 547 |
end: *const u8, |
| 548 |
prev0: &mut V, |
| 549 |
) -> Option<Match> { |
| 550 |
let c = self.candidate(cur, prev0); |
| 551 |
if !c.is_zero() { |
| 552 |
if let Some(m) = self.teddy.verify(cur.sub(1), end, c) { |
| 553 |
return Some(m); |
| 554 |
} |
| 555 |
} |
| 556 |
None |
| 557 |
} |
| 558 |
|
| 559 |
/// See `Fat<V, 1>::candidate`. |
| 560 |
#[inline(always)] |
| 561 |
unsafe fn candidate(&self, cur: *const u8, prev0: &mut V) -> V { |
| 562 |
let chunk = V::load_half_unaligned(cur); |
| 563 |
let (res0, res1) = Mask::members2(chunk, self.masks); |
| 564 |
let res0prev0 = res0.half_shift_in_one_byte(*prev0); |
| 565 |
let res = res0prev0.and(res1); |
| 566 |
*prev0 = res0; |
| 567 |
res |
| 568 |
} |
| 569 |
} |
| 570 |
|
| 571 |
impl<V: FatVector> Fat<V, 3> { |
| 572 |
/// See `Fat<V, 1>::find`. |
| 573 |
#[inline(always)] |
| 574 |
pub(crate) unsafe fn find( |
| 575 |
&self, |
| 576 |
start: *const u8, |
| 577 |
end: *const u8, |
| 578 |
) -> Option<Match> { |
| 579 |
let len = end.distance(start); |
| 580 |
debug_assert!(len >= self.minimum_len()); |
| 581 |
let mut cur = start.add(2); |
| 582 |
let mut prev0 = V::splat(0xFF); |
| 583 |
let mut prev1 = V::splat(0xFF); |
| 584 |
while cur <= end.sub(V::Half::BYTES) { |
| 585 |
if let Some(m) = self.find_one(cur, end, &mut prev0, &mut prev1) { |
| 586 |
return Some(m); |
| 587 |
} |
| 588 |
cur = cur.add(V::Half::BYTES); |
| 589 |
} |
| 590 |
if cur < end { |
| 591 |
cur = end.sub(V::Half::BYTES); |
| 592 |
prev0 = V::splat(0xFF); |
| 593 |
prev1 = V::splat(0xFF); |
| 594 |
if let Some(m) = self.find_one(cur, end, &mut prev0, &mut prev1) { |
| 595 |
return Some(m); |
| 596 |
} |
| 597 |
} |
| 598 |
None |
| 599 |
} |
| 600 |
|
| 601 |
/// See `Fat<V, 1>::find_one`. |
| 602 |
#[inline(always)] |
| 603 |
unsafe fn find_one( |
| 604 |
&self, |
| 605 |
cur: *const u8, |
| 606 |
end: *const u8, |
| 607 |
prev0: &mut V, |
| 608 |
prev1: &mut V, |
| 609 |
) -> Option<Match> { |
| 610 |
let c = self.candidate(cur, prev0, prev1); |
| 611 |
if !c.is_zero() { |
| 612 |
if let Some(m) = self.teddy.verify(cur.sub(2), end, c) { |
| 613 |
return Some(m); |
| 614 |
} |
| 615 |
} |
| 616 |
None |
| 617 |
} |
| 618 |
|
| 619 |
/// See `Fat<V, 1>::candidate`. |
| 620 |
#[inline(always)] |
| 621 |
unsafe fn candidate( |
| 622 |
&self, |
| 623 |
cur: *const u8, |
| 624 |
prev0: &mut V, |
| 625 |
prev1: &mut V, |
| 626 |
) -> V { |
| 627 |
let chunk = V::load_half_unaligned(cur); |
| 628 |
let (res0, res1, res2) = Mask::members3(chunk, self.masks); |
| 629 |
let res0prev0 = res0.half_shift_in_two_bytes(*prev0); |
| 630 |
let res1prev1 = res1.half_shift_in_one_byte(*prev1); |
| 631 |
let res = res0prev0.and(res1prev1).and(res2); |
| 632 |
*prev0 = res0; |
| 633 |
*prev1 = res1; |
| 634 |
res |
| 635 |
} |
| 636 |
} |
| 637 |
|
| 638 |
impl<V: FatVector> Fat<V, 4> { |
| 639 |
/// See `Fat<V, 1>::find`. |
| 640 |
#[inline(always)] |
| 641 |
pub(crate) unsafe fn find( |
| 642 |
&self, |
| 643 |
start: *const u8, |
| 644 |
end: *const u8, |
| 645 |
) -> Option<Match> { |
| 646 |
let len = end.distance(start); |
| 647 |
debug_assert!(len >= self.minimum_len()); |
| 648 |
let mut cur = start.add(3); |
| 649 |
let mut prev0 = V::splat(0xFF); |
| 650 |
let mut prev1 = V::splat(0xFF); |
| 651 |
let mut prev2 = V::splat(0xFF); |
| 652 |
while cur <= end.sub(V::Half::BYTES) { |
| 653 |
if let Some(m) = |
| 654 |
self.find_one(cur, end, &mut prev0, &mut prev1, &mut prev2) |
| 655 |
{ |
| 656 |
return Some(m); |
| 657 |
} |
| 658 |
cur = cur.add(V::Half::BYTES); |
| 659 |
} |
| 660 |
if cur < end { |
| 661 |
cur = end.sub(V::Half::BYTES); |
| 662 |
prev0 = V::splat(0xFF); |
| 663 |
prev1 = V::splat(0xFF); |
| 664 |
prev2 = V::splat(0xFF); |
| 665 |
if let Some(m) = |
| 666 |
self.find_one(cur, end, &mut prev0, &mut prev1, &mut prev2) |
| 667 |
{ |
| 668 |
return Some(m); |
| 669 |
} |
| 670 |
} |
| 671 |
None |
| 672 |
} |
| 673 |
|
| 674 |
/// See `Fat<V, 1>::find_one`. |
| 675 |
#[inline(always)] |
| 676 |
unsafe fn find_one( |
| 677 |
&self, |
| 678 |
cur: *const u8, |
| 679 |
end: *const u8, |
| 680 |
prev0: &mut V, |
| 681 |
prev1: &mut V, |
| 682 |
prev2: &mut V, |
| 683 |
) -> Option<Match> { |
| 684 |
let c = self.candidate(cur, prev0, prev1, prev2); |
| 685 |
if !c.is_zero() { |
| 686 |
if let Some(m) = self.teddy.verify(cur.sub(3), end, c) { |
| 687 |
return Some(m); |
| 688 |
} |
| 689 |
} |
| 690 |
None |
| 691 |
} |
| 692 |
|
| 693 |
/// See `Fat<V, 1>::candidate`. |
| 694 |
#[inline(always)] |
| 695 |
unsafe fn candidate( |
| 696 |
&self, |
| 697 |
cur: *const u8, |
| 698 |
prev0: &mut V, |
| 699 |
prev1: &mut V, |
| 700 |
prev2: &mut V, |
| 701 |
) -> V { |
| 702 |
let chunk = V::load_half_unaligned(cur); |
| 703 |
let (res0, res1, res2, res3) = Mask::members4(chunk, self.masks); |
| 704 |
let res0prev0 = res0.half_shift_in_three_bytes(*prev0); |
| 705 |
let res1prev1 = res1.half_shift_in_two_bytes(*prev1); |
| 706 |
let res2prev2 = res2.half_shift_in_one_byte(*prev2); |
| 707 |
let res = res0prev0.and(res1prev1).and(res2prev2).and(res3); |
| 708 |
*prev0 = res0; |
| 709 |
*prev1 = res1; |
| 710 |
*prev2 = res2; |
| 711 |
res |
| 712 |
} |
| 713 |
} |
| 714 |
|
| 715 |
/// The common elements of all "slim" and "fat" Teddy search implementations. |
| 716 |
/// |
| 717 |
/// Essentially, this contains the patterns and the buckets. Namely, it |
| 718 |
/// contains enough to implement the verification step after candidates are |
| 719 |
/// identified via the shuffle masks. |
| 720 |
/// |
| 721 |
/// It is generic over the number of buckets used. In general, the number of |
| 722 |
/// buckets is either 8 (for "slim" Teddy) or 16 (for "fat" Teddy). The generic |
| 723 |
/// parameter isn't really meant to be instantiated for any value other than |
| 724 |
/// 8 or 16, although it is technically possible. The main hiccup is that there |
| 725 |
/// is some bit-shifting done in the critical part of verification that could |
| 726 |
/// be quite expensive if `N` is not a multiple of 2. |
| 727 |
#[derive(Clone, Debug)] |
| 728 |
struct Teddy<const BUCKETS: usize> { |
| 729 |
/// The patterns we are searching for. |
| 730 |
/// |
| 731 |
/// A pattern string can be found by its `PatternID`. |
| 732 |
patterns: Arc<Patterns>, |
| 733 |
/// The allocation of patterns in buckets. This only contains the IDs of |
| 734 |
/// patterns. In order to do full verification, callers must provide the |
| 735 |
/// actual patterns when using Teddy. |
| 736 |
buckets: [Vec<PatternID>; BUCKETS], |
| 737 |
// N.B. The above representation is very simple, but it definitely results |
| 738 |
// in ping-ponging between different allocations during verification. I've |
| 739 |
// tried experimenting with other representations that flatten the pattern |
| 740 |
// strings into a single allocation, but it doesn't seem to help much. |
| 741 |
// Probably everything is small enough to fit into cache anyway, and so the |
| 742 |
// pointer chasing isn't a big deal? |
| 743 |
// |
| 744 |
// One other avenue I haven't explored is some kind of hashing trick |
| 745 |
// that let's us do another high-confidence check before launching into |
| 746 |
// `memcmp`. |
| 747 |
} |
| 748 |
|
| 749 |
impl<const BUCKETS: usize> Teddy<BUCKETS> { |
| 750 |
/// Create a new generic data structure for Teddy verification. |
| 751 |
fn new(patterns: Arc<Patterns>) -> Teddy<BUCKETS> { |
| 752 |
assert_ne!(0, patterns.len(), "Teddy requires at least one pattern"); |
| 753 |
assert_ne!( |
| 754 |
0, |
| 755 |
patterns.minimum_len(), |
| 756 |
"Teddy does not support zero-length patterns" |
| 757 |
); |
| 758 |
assert!( |
| 759 |
BUCKETS == 8 || BUCKETS == 16, |
| 760 |
"Teddy only supports 8 or 16 buckets" |
| 761 |
); |
| 762 |
// MSRV(1.63): Use core::array::from_fn below instead of allocating a |
| 763 |
// superfluous outer Vec. Not a big deal (especially given the BTreeMap |
| 764 |
// allocation below), but nice to not do it. |
| 765 |
let buckets = |
| 766 |
<[Vec<PatternID>; BUCKETS]>::try_from(vec![vec![]; BUCKETS]) |
| 767 |
.unwrap(); |
| 768 |
let mut t = Teddy { patterns, buckets }; |
| 769 |
|
| 770 |
let mut map: BTreeMap<Box<[u8]>, usize> = BTreeMap::new(); |
| 771 |
for (id, pattern) in t.patterns.iter() { |
| 772 |
// We try to be slightly clever in how we assign patterns into |
| 773 |
// buckets. Generally speaking, we want patterns with the same |
| 774 |
// prefix to be in the same bucket, since it minimizes the amount |
| 775 |
// of time we spend churning through buckets in the verification |
| 776 |
// step. |
| 777 |
// |
| 778 |
// So we could assign patterns with the same N-prefix (where N is |
| 779 |
// the size of the mask, which is one of {1, 2, 3}) to the same |
| 780 |
// bucket. However, case insensitive searches are fairly common, so |
| 781 |
// we'd for example, ideally want to treat `abc` and `ABC` as if |
| 782 |
// they shared the same prefix. ASCII has the nice property that |
| 783 |
// the lower 4 bits of A and a are the same, so we therefore group |
| 784 |
// patterns with the same low-nybble-N-prefix into the same bucket. |
| 785 |
// |
| 786 |
// MOREOVER, this is actually necessary for correctness! In |
| 787 |
// particular, by grouping patterns with the same prefix into the |
| 788 |
// same bucket, we ensure that we preserve correct leftmost-first |
| 789 |
// and leftmost-longest match semantics. In addition to the fact |
| 790 |
// that `patterns.iter()` iterates in the correct order, this |
| 791 |
// guarantees that all possible ambiguous matches will occur in |
| 792 |
// the same bucket. The verification routine could be adjusted to |
| 793 |
// support correct leftmost match semantics regardless of bucket |
| 794 |
// allocation, but that results in a performance hit. It's much |
| 795 |
// nicer to be able to just stop as soon as a match is found. |
| 796 |
let lonybs = pattern.low_nybbles(t.mask_len()); |
| 797 |
if let Some(&bucket) = map.get(&lonybs) { |
| 798 |
t.buckets[bucket].push(id); |
| 799 |
} else { |
| 800 |
// N.B. We assign buckets in reverse because it shouldn't have |
| 801 |
// any influence on performance, but it does make it harder to |
| 802 |
// get leftmost match semantics accidentally correct. |
| 803 |
let bucket = (BUCKETS - 1) - (id.as_usize() % BUCKETS); |
| 804 |
t.buckets[bucket].push(id); |
| 805 |
map.insert(lonybs, bucket); |
| 806 |
} |
| 807 |
} |
| 808 |
t |
| 809 |
} |
| 810 |
|
| 811 |
/// Verify whether there are any matches starting at or after `cur` in the |
| 812 |
/// haystack. The candidate chunk given should correspond to 8-bit bitsets |
| 813 |
/// for N buckets. |
| 814 |
/// |
| 815 |
/// # Safety |
| 816 |
/// |
| 817 |
/// The given pointers representing the haystack must be valid to read |
| 818 |
/// from. |
| 819 |
#[inline(always)] |
| 820 |
unsafe fn verify64( |
| 821 |
&self, |
| 822 |
cur: *const u8, |
| 823 |
end: *const u8, |
| 824 |
mut candidate_chunk: u64, |
| 825 |
) -> Option<Match> { |
| 826 |
while candidate_chunk != 0 { |
| 827 |
let bit = candidate_chunk.trailing_zeros().as_usize(); |
| 828 |
candidate_chunk &= !(1 << bit); |
| 829 |
|
| 830 |
let cur = cur.add(bit / BUCKETS); |
| 831 |
let bucket = bit % BUCKETS; |
| 832 |
if let Some(m) = self.verify_bucket(cur, end, bucket) { |
| 833 |
return Some(m); |
| 834 |
} |
| 835 |
} |
| 836 |
None |
| 837 |
} |
| 838 |
|
| 839 |
/// Verify whether there are any matches starting at `at` in the given |
| 840 |
/// `haystack` corresponding only to patterns in the given bucket. |
| 841 |
/// |
| 842 |
/// # Safety |
| 843 |
/// |
| 844 |
/// The given pointers representing the haystack must be valid to read |
| 845 |
/// from. |
| 846 |
/// |
| 847 |
/// The bucket index must be less than or equal to `self.buckets.len()`. |
| 848 |
#[inline(always)] |
| 849 |
unsafe fn verify_bucket( |
| 850 |
&self, |
| 851 |
cur: *const u8, |
| 852 |
end: *const u8, |
| 853 |
bucket: usize, |
| 854 |
) -> Option<Match> { |
| 855 |
debug_assert!(bucket < self.buckets.len()); |
| 856 |
// SAFETY: The caller must ensure that the bucket index is correct. |
| 857 |
for pid in self.buckets.get_unchecked(bucket).iter().copied() { |
| 858 |
// SAFETY: This is safe because we are guaranteed that every |
| 859 |
// index in a Teddy bucket is a valid index into `pats`, by |
| 860 |
// construction. |
| 861 |
debug_assert!(pid.as_usize() < self.patterns.len()); |
| 862 |
let pat = self.patterns.get_unchecked(pid); |
| 863 |
if pat.is_prefix_raw(cur, end) { |
| 864 |
let start = cur; |
| 865 |
let end = start.add(pat.len()); |
| 866 |
return Some(Match { pid, start, end }); |
| 867 |
} |
| 868 |
} |
| 869 |
None |
| 870 |
} |
| 871 |
|
| 872 |
/// Returns the total number of masks required by the patterns in this |
| 873 |
/// Teddy searcher. |
| 874 |
/// |
| 875 |
/// Basically, the mask length corresponds to the type of Teddy searcher |
| 876 |
/// to use: a 1-byte, 2-byte, 3-byte or 4-byte searcher. The bigger the |
| 877 |
/// better, typically, since searching for longer substrings usually |
| 878 |
/// decreases the rate of false positives. Therefore, the number of masks |
| 879 |
/// needed is the length of the shortest pattern in this searcher. If the |
| 880 |
/// length of the shortest pattern (in bytes) is bigger than 4, then the |
| 881 |
/// mask length is 4 since there are no Teddy searchers for more than 4 |
| 882 |
/// bytes. |
| 883 |
fn mask_len(&self) -> usize { |
| 884 |
core::cmp::min(4, self.patterns.minimum_len()) |
| 885 |
} |
| 886 |
|
| 887 |
/// Returns the approximate total amount of heap used by this type, in |
| 888 |
/// units of bytes. |
| 889 |
fn memory_usage(&self) -> usize { |
| 890 |
// This is an upper bound rather than a precise accounting. No |
| 891 |
// particular reason, other than it's probably very close to actual |
| 892 |
// memory usage in practice. |
| 893 |
self.patterns.len() * core::mem::size_of::<PatternID>() |
| 894 |
} |
| 895 |
} |
| 896 |
|
| 897 |
impl Teddy<8> { |
| 898 |
/// Runs the verification routine for "slim" Teddy. |
| 899 |
/// |
| 900 |
/// The candidate given should be a collection of 8-bit bitsets (one bitset |
| 901 |
/// per lane), where the ith bit is set in the jth lane if and only if the |
| 902 |
/// byte occurring at `at + j` in `cur` is in the bucket `i`. |
| 903 |
/// |
| 904 |
/// # Safety |
| 905 |
/// |
| 906 |
/// Callers must ensure that this is okay to call in the current target for |
| 907 |
/// the current CPU. |
| 908 |
/// |
| 909 |
/// The given pointers must be valid to read from. |
| 910 |
#[inline(always)] |
| 911 |
unsafe fn verify<V: Vector>( |
| 912 |
&self, |
| 913 |
mut cur: *const u8, |
| 914 |
end: *const u8, |
| 915 |
candidate: V, |
| 916 |
) -> Option<Match> { |
| 917 |
debug_assert!(!candidate.is_zero()); |
| 918 |
// Convert the candidate into 64-bit chunks, and then verify each of |
| 919 |
// those chunks. |
| 920 |
candidate.for_each_64bit_lane( |
| 921 |
#[inline(always)] |
| 922 |
|_, chunk| { |
| 923 |
let result = self.verify64(cur, end, chunk); |
| 924 |
cur = cur.add(8); |
| 925 |
result |
| 926 |
}, |
| 927 |
) |
| 928 |
} |
| 929 |
} |
| 930 |
|
| 931 |
impl Teddy<16> { |
| 932 |
/// Runs the verification routine for "fat" Teddy. |
| 933 |
/// |
| 934 |
/// The candidate given should be a collection of 8-bit bitsets (one bitset |
| 935 |
/// per lane), where the ith bit is set in the jth lane if and only if the |
| 936 |
/// byte occurring at `at + (j < 16 ? j : j - 16)` in `cur` is in the |
| 937 |
/// bucket `j < 16 ? i : i + 8`. |
| 938 |
/// |
| 939 |
/// # Safety |
| 940 |
/// |
| 941 |
/// Callers must ensure that this is okay to call in the current target for |
| 942 |
/// the current CPU. |
| 943 |
/// |
| 944 |
/// The given pointers must be valid to read from. |
| 945 |
#[inline(always)] |
| 946 |
unsafe fn verify<V: FatVector>( |
| 947 |
&self, |
| 948 |
mut cur: *const u8, |
| 949 |
end: *const u8, |
| 950 |
candidate: V, |
| 951 |
) -> Option<Match> { |
| 952 |
// This is a bit tricky, but we basically want to convert our |
| 953 |
// candidate, which looks like this (assuming a 256-bit vector): |
| 954 |
// |
| 955 |
// a31 a30 ... a17 a16 a15 a14 ... a01 a00 |
| 956 |
// |
| 957 |
// where each a(i) is an 8-bit bitset corresponding to the activated |
| 958 |
// buckets, to this |
| 959 |
// |
| 960 |
// a31 a15 a30 a14 a29 a13 ... a18 a02 a17 a01 a16 a00 |
| 961 |
// |
| 962 |
// Namely, for Fat Teddy, the high 128-bits of the candidate correspond |
| 963 |
// to the same bytes in the haystack in the low 128-bits (so we only |
| 964 |
// scan 16 bytes at a time), but are for buckets 8-15 instead of 0-7. |
| 965 |
// |
| 966 |
// The verification routine wants to look at all potentially matching |
| 967 |
// buckets before moving on to the next lane. So for example, both |
| 968 |
// a16 and a00 both correspond to the first byte in our window; a00 |
| 969 |
// contains buckets 0-7 and a16 contains buckets 8-15. Specifically, |
| 970 |
// a16 should be checked before a01. So the transformation shown above |
| 971 |
// allows us to use our normal verification procedure with one small |
| 972 |
// change: we treat each bitset as 16 bits instead of 8 bits. |
| 973 |
debug_assert!(!candidate.is_zero()); |
| 974 |
|
| 975 |
// Swap the 128-bit lanes in the candidate vector. |
| 976 |
let swapped = candidate.swap_halves(); |
| 977 |
// Interleave the bytes from the low 128-bit lanes, starting with |
| 978 |
// cand first. |
| 979 |
let r1 = candidate.interleave_low_8bit_lanes(swapped); |
| 980 |
// Interleave the bytes from the high 128-bit lanes, starting with |
| 981 |
// cand first. |
| 982 |
let r2 = candidate.interleave_high_8bit_lanes(swapped); |
| 983 |
// Now just take the 2 low 64-bit integers from both r1 and r2. We |
| 984 |
// can drop the high 64-bit integers because they are a mirror image |
| 985 |
// of the low 64-bit integers. All we care about are the low 128-bit |
| 986 |
// lanes of r1 and r2. Combined, they contain all our 16-bit bitsets |
| 987 |
// laid out in the desired order, as described above. |
| 988 |
r1.for_each_low_64bit_lane( |
| 989 |
r2, |
| 990 |
#[inline(always)] |
| 991 |
|_, chunk| { |
| 992 |
let result = self.verify64(cur, end, chunk); |
| 993 |
cur = cur.add(4); |
| 994 |
result |
| 995 |
}, |
| 996 |
) |
| 997 |
} |
| 998 |
} |
| 999 |
|
| 1000 |
/// A vector generic mask for the low and high nybbles in a set of patterns. |
| 1001 |
/// Each 8-bit lane `j` in a vector corresponds to a bitset where the `i`th bit |
| 1002 |
/// is set if and only if the nybble `j` is in the bucket `i` at a particular |
| 1003 |
/// position. |
| 1004 |
/// |
| 1005 |
/// This is slightly tweaked dependending on whether Slim or Fat Teddy is being |
| 1006 |
/// used. For Slim Teddy, the bitsets in the lower half are the same as the |
| 1007 |
/// bitsets in the higher half, so that we can search `V::BYTES` bytes at a |
| 1008 |
/// time. (Remember, the nybbles in the haystack are used as indices into these |
| 1009 |
/// masks, and 256-bit shuffles only operate on 128-bit lanes.) |
| 1010 |
/// |
| 1011 |
/// For Fat Teddy, the bitsets are not repeated, but instead, the high half |
| 1012 |
/// bits correspond to an addition 8 buckets. So that a bitset `00100010` has |
| 1013 |
/// buckets 1 and 5 set if it's in the lower half, but has buckets 9 and 13 set |
| 1014 |
/// if it's in the higher half. |
| 1015 |
#[derive(Clone, Copy, Debug)] |
| 1016 |
struct Mask<V> { |
| 1017 |
lo: V, |
| 1018 |
hi: V, |
| 1019 |
} |
| 1020 |
|
| 1021 |
impl<V: Vector> Mask<V> { |
| 1022 |
/// Return a candidate for Teddy (fat or slim) that is searching for 1-byte |
| 1023 |
/// candidates. |
| 1024 |
/// |
| 1025 |
/// If a candidate is returned, it will be a collection of 8-bit bitsets |
| 1026 |
/// (one bitset per lane), where the ith bit is set in the jth lane if and |
| 1027 |
/// only if the byte occurring at the jth lane in `chunk` is in the bucket |
| 1028 |
/// `i`. If no candidate is found, then the vector returned will have all |
| 1029 |
/// lanes set to zero. |
| 1030 |
/// |
| 1031 |
/// `chunk` should correspond to a `V::BYTES` window of the haystack (where |
| 1032 |
/// the least significant byte corresponds to the start of the window). For |
| 1033 |
/// fat Teddy, the haystack window length should be `V::BYTES / 2`, with |
| 1034 |
/// the window repeated in each half of the vector. |
| 1035 |
/// |
| 1036 |
/// `mask1` should correspond to a low/high mask for the first byte of all |
| 1037 |
/// patterns that are being searched. |
| 1038 |
#[inline(always)] |
| 1039 |
unsafe fn members1(chunk: V, masks: [Mask<V>; 1]) -> V { |
| 1040 |
let lomask = V::splat(0xF); |
| 1041 |
let hlo = chunk.and(lomask); |
| 1042 |
let hhi = chunk.shift_8bit_lane_right::<4>().and(lomask); |
| 1043 |
let locand = masks[0].lo.shuffle_bytes(hlo); |
| 1044 |
let hicand = masks[0].hi.shuffle_bytes(hhi); |
| 1045 |
locand.and(hicand) |
| 1046 |
} |
| 1047 |
|
| 1048 |
/// Return a candidate for Teddy (fat or slim) that is searching for 2-byte |
| 1049 |
/// candidates. |
| 1050 |
/// |
| 1051 |
/// If candidates are returned, each will be a collection of 8-bit bitsets |
| 1052 |
/// (one bitset per lane), where the ith bit is set in the jth lane if and |
| 1053 |
/// only if the byte occurring at the jth lane in `chunk` is in the bucket |
| 1054 |
/// `i`. Each candidate returned corresponds to the first and second bytes |
| 1055 |
/// of the patterns being searched. If no candidate is found, then all of |
| 1056 |
/// the lanes will be set to zero in at least one of the vectors returned. |
| 1057 |
/// |
| 1058 |
/// `chunk` should correspond to a `V::BYTES` window of the haystack (where |
| 1059 |
/// the least significant byte corresponds to the start of the window). For |
| 1060 |
/// fat Teddy, the haystack window length should be `V::BYTES / 2`, with |
| 1061 |
/// the window repeated in each half of the vector. |
| 1062 |
/// |
| 1063 |
/// The masks should correspond to the masks computed for the first and |
| 1064 |
/// second bytes of all patterns that are being searched. |
| 1065 |
#[inline(always)] |
| 1066 |
unsafe fn members2(chunk: V, masks: [Mask<V>; 2]) -> (V, V) { |
| 1067 |
let lomask = V::splat(0xF); |
| 1068 |
let hlo = chunk.and(lomask); |
| 1069 |
let hhi = chunk.shift_8bit_lane_right::<4>().and(lomask); |
| 1070 |
|
| 1071 |
let locand1 = masks[0].lo.shuffle_bytes(hlo); |
| 1072 |
let hicand1 = masks[0].hi.shuffle_bytes(hhi); |
| 1073 |
let cand1 = locand1.and(hicand1); |
| 1074 |
|
| 1075 |
let locand2 = masks[1].lo.shuffle_bytes(hlo); |
| 1076 |
let hicand2 = masks[1].hi.shuffle_bytes(hhi); |
| 1077 |
let cand2 = locand2.and(hicand2); |
| 1078 |
|
| 1079 |
(cand1, cand2) |
| 1080 |
} |
| 1081 |
|
| 1082 |
/// Return a candidate for Teddy (fat or slim) that is searching for 3-byte |
| 1083 |
/// candidates. |
| 1084 |
/// |
| 1085 |
/// If candidates are returned, each will be a collection of 8-bit bitsets |
| 1086 |
/// (one bitset per lane), where the ith bit is set in the jth lane if and |
| 1087 |
/// only if the byte occurring at the jth lane in `chunk` is in the bucket |
| 1088 |
/// `i`. Each candidate returned corresponds to the first, second and third |
| 1089 |
/// bytes of the patterns being searched. If no candidate is found, then |
| 1090 |
/// all of the lanes will be set to zero in at least one of the vectors |
| 1091 |
/// returned. |
| 1092 |
/// |
| 1093 |
/// `chunk` should correspond to a `V::BYTES` window of the haystack (where |
| 1094 |
/// the least significant byte corresponds to the start of the window). For |
| 1095 |
/// fat Teddy, the haystack window length should be `V::BYTES / 2`, with |
| 1096 |
/// the window repeated in each half of the vector. |
| 1097 |
/// |
| 1098 |
/// The masks should correspond to the masks computed for the first, second |
| 1099 |
/// and third bytes of all patterns that are being searched. |
| 1100 |
#[inline(always)] |
| 1101 |
unsafe fn members3(chunk: V, masks: [Mask<V>; 3]) -> (V, V, V) { |
| 1102 |
let lomask = V::splat(0xF); |
| 1103 |
let hlo = chunk.and(lomask); |
| 1104 |
let hhi = chunk.shift_8bit_lane_right::<4>().and(lomask); |
| 1105 |
|
| 1106 |
let locand1 = masks[0].lo.shuffle_bytes(hlo); |
| 1107 |
let hicand1 = masks[0].hi.shuffle_bytes(hhi); |
| 1108 |
let cand1 = locand1.and(hicand1); |
| 1109 |
|
| 1110 |
let locand2 = masks[1].lo.shuffle_bytes(hlo); |
| 1111 |
let hicand2 = masks[1].hi.shuffle_bytes(hhi); |
| 1112 |
let cand2 = locand2.and(hicand2); |
| 1113 |
|
| 1114 |
let locand3 = masks[2].lo.shuffle_bytes(hlo); |
| 1115 |
let hicand3 = masks[2].hi.shuffle_bytes(hhi); |
| 1116 |
let cand3 = locand3.and(hicand3); |
| 1117 |
|
| 1118 |
(cand1, cand2, cand3) |
| 1119 |
} |
| 1120 |
|
| 1121 |
/// Return a candidate for Teddy (fat or slim) that is searching for 4-byte |
| 1122 |
/// candidates. |
| 1123 |
/// |
| 1124 |
/// If candidates are returned, each will be a collection of 8-bit bitsets |
| 1125 |
/// (one bitset per lane), where the ith bit is set in the jth lane if and |
| 1126 |
/// only if the byte occurring at the jth lane in `chunk` is in the bucket |
| 1127 |
/// `i`. Each candidate returned corresponds to the first, second, third |
| 1128 |
/// and fourth bytes of the patterns being searched. If no candidate is |
| 1129 |
/// found, then all of the lanes will be set to zero in at least one of the |
| 1130 |
/// vectors returned. |
| 1131 |
/// |
| 1132 |
/// `chunk` should correspond to a `V::BYTES` window of the haystack (where |
| 1133 |
/// the least significant byte corresponds to the start of the window). For |
| 1134 |
/// fat Teddy, the haystack window length should be `V::BYTES / 2`, with |
| 1135 |
/// the window repeated in each half of the vector. |
| 1136 |
/// |
| 1137 |
/// The masks should correspond to the masks computed for the first, |
| 1138 |
/// second, third and fourth bytes of all patterns that are being searched. |
| 1139 |
#[inline(always)] |
| 1140 |
unsafe fn members4(chunk: V, masks: [Mask<V>; 4]) -> (V, V, V, V) { |
| 1141 |
let lomask = V::splat(0xF); |
| 1142 |
let hlo = chunk.and(lomask); |
| 1143 |
let hhi = chunk.shift_8bit_lane_right::<4>().and(lomask); |
| 1144 |
|
| 1145 |
let locand1 = masks[0].lo.shuffle_bytes(hlo); |
| 1146 |
let hicand1 = masks[0].hi.shuffle_bytes(hhi); |
| 1147 |
let cand1 = locand1.and(hicand1); |
| 1148 |
|
| 1149 |
let locand2 = masks[1].lo.shuffle_bytes(hlo); |
| 1150 |
let hicand2 = masks[1].hi.shuffle_bytes(hhi); |
| 1151 |
let cand2 = locand2.and(hicand2); |
| 1152 |
|
| 1153 |
let locand3 = masks[2].lo.shuffle_bytes(hlo); |
| 1154 |
let hicand3 = masks[2].hi.shuffle_bytes(hhi); |
| 1155 |
let cand3 = locand3.and(hicand3); |
| 1156 |
|
| 1157 |
let locand4 = masks[3].lo.shuffle_bytes(hlo); |
| 1158 |
let hicand4 = masks[3].hi.shuffle_bytes(hhi); |
| 1159 |
let cand4 = locand4.and(hicand4); |
| 1160 |
|
| 1161 |
(cand1, cand2, cand3, cand4) |
| 1162 |
} |
| 1163 |
} |
| 1164 |
|
| 1165 |
/// Represents the low and high nybble masks that will be used during |
| 1166 |
/// search. Each mask is 32 bytes wide, although only the first 16 bytes are |
| 1167 |
/// used for 128-bit vectors. |
| 1168 |
/// |
| 1169 |
/// Each byte in the mask corresponds to a 8-bit bitset, where bit `i` is set |
| 1170 |
/// if and only if the corresponding nybble is in the ith bucket. The index of |
| 1171 |
/// the byte (0-15, inclusive) corresponds to the nybble. |
| 1172 |
/// |
| 1173 |
/// Each mask is used as the target of a shuffle, where the indices for the |
| 1174 |
/// shuffle are taken from the haystack. AND'ing the shuffles for both the |
| 1175 |
/// low and high masks together also results in 8-bit bitsets, but where bit |
| 1176 |
/// `i` is set if and only if the correspond *byte* is in the ith bucket. |
| 1177 |
#[derive(Clone, Default)] |
| 1178 |
struct SlimMaskBuilder { |
| 1179 |
lo: [u8; 32], |
| 1180 |
hi: [u8; 32], |
| 1181 |
} |
| 1182 |
|
| 1183 |
impl SlimMaskBuilder { |
| 1184 |
/// Update this mask by adding the given byte to the given bucket. The |
| 1185 |
/// given bucket must be in the range 0-7. |
| 1186 |
/// |
| 1187 |
/// # Panics |
| 1188 |
/// |
| 1189 |
/// When `bucket >= 8`. |
| 1190 |
fn add(&mut self, bucket: usize, byte: u8) { |
| 1191 |
assert!(bucket < 8); |
| 1192 |
|
| 1193 |
let bucket = u8::try_from(bucket).unwrap(); |
| 1194 |
let byte_lo = usize::from(byte & 0xF); |
| 1195 |
let byte_hi = usize::from((byte >> 4) & 0xF); |
| 1196 |
// When using 256-bit vectors, we need to set this bucket assignment in |
| 1197 |
// the low and high 128-bit portions of the mask. This allows us to |
| 1198 |
// process 32 bytes at a time. Namely, AVX2 shuffles operate on each |
| 1199 |
// of the 128-bit lanes, rather than the full 256-bit vector at once. |
| 1200 |
self.lo[byte_lo] |= 1 << bucket; |
| 1201 |
self.lo[byte_lo + 16] |= 1 << bucket; |
| 1202 |
self.hi[byte_hi] |= 1 << bucket; |
| 1203 |
self.hi[byte_hi + 16] |= 1 << bucket; |
| 1204 |
} |
| 1205 |
|
| 1206 |
/// Turn this builder into a vector mask. |
| 1207 |
/// |
| 1208 |
/// # Panics |
| 1209 |
/// |
| 1210 |
/// When `V` represents a vector bigger than what `MaskBytes` can contain. |
| 1211 |
/// |
| 1212 |
/// # Safety |
| 1213 |
/// |
| 1214 |
/// Callers must ensure that this is okay to call in the current target for |
| 1215 |
/// the current CPU. |
| 1216 |
#[inline(always)] |
| 1217 |
unsafe fn build<V: Vector>(&self) -> Mask<V> { |
| 1218 |
assert!(V::BYTES <= self.lo.len()); |
| 1219 |
assert!(V::BYTES <= self.hi.len()); |
| 1220 |
Mask { |
| 1221 |
lo: V::load_unaligned(self.lo[..].as_ptr()), |
| 1222 |
hi: V::load_unaligned(self.hi[..].as_ptr()), |
| 1223 |
} |
| 1224 |
} |
| 1225 |
|
| 1226 |
/// A convenience function for building `N` vector masks from a slim |
| 1227 |
/// `Teddy` value. |
| 1228 |
/// |
| 1229 |
/// # Panics |
| 1230 |
/// |
| 1231 |
/// When `V` represents a vector bigger than what `MaskBytes` can contain. |
| 1232 |
/// |
| 1233 |
/// # Safety |
| 1234 |
/// |
| 1235 |
/// Callers must ensure that this is okay to call in the current target for |
| 1236 |
/// the current CPU. |
| 1237 |
#[inline(always)] |
| 1238 |
unsafe fn from_teddy<const BYTES: usize, V: Vector>( |
| 1239 |
teddy: &Teddy<8>, |
| 1240 |
) -> [Mask<V>; BYTES] { |
| 1241 |
// MSRV(1.63): Use core::array::from_fn to just build the array here |
| 1242 |
// instead of creating a vector and turning it into an array. |
| 1243 |
let mut mask_builders = vec![SlimMaskBuilder::default(); BYTES]; |
| 1244 |
for (bucket_index, bucket) in teddy.buckets.iter().enumerate() { |
| 1245 |
for pid in bucket.iter().copied() { |
| 1246 |
let pat = teddy.patterns.get(pid); |
| 1247 |
for (i, builder) in mask_builders.iter_mut().enumerate() { |
| 1248 |
builder.add(bucket_index, pat.bytes()[i]); |
| 1249 |
} |
| 1250 |
} |
| 1251 |
} |
| 1252 |
let array = |
| 1253 |
<[SlimMaskBuilder; BYTES]>::try_from(mask_builders).unwrap(); |
| 1254 |
array.map(|builder| builder.build()) |
| 1255 |
} |
| 1256 |
} |
| 1257 |
|
| 1258 |
impl Debug for SlimMaskBuilder { |
| 1259 |
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { |
| 1260 |
let (mut parts_lo, mut parts_hi) = (vec![], vec![]); |
| 1261 |
for i in 0..32 { |
| 1262 |
parts_lo.push(format!("{:02}: {:08b}", i, self.lo[i])); |
| 1263 |
parts_hi.push(format!("{:02}: {:08b}", i, self.hi[i])); |
| 1264 |
} |
| 1265 |
f.debug_struct("SlimMaskBuilder") |
| 1266 |
.field("lo", &parts_lo) |
| 1267 |
.field("hi", &parts_hi) |
| 1268 |
.finish() |
| 1269 |
} |
| 1270 |
} |
| 1271 |
|
| 1272 |
/// Represents the low and high nybble masks that will be used during "fat" |
| 1273 |
/// Teddy search. |
| 1274 |
/// |
| 1275 |
/// Each mask is 32 bytes wide, and at the time of writing, only 256-bit vectors |
| 1276 |
/// support fat Teddy. |
| 1277 |
/// |
| 1278 |
/// A fat Teddy mask is like a slim Teddy mask, except that instead of |
| 1279 |
/// repeating the bitsets in the high and low 128-bits in 256-bit vectors, the |
| 1280 |
/// high and low 128-bit halves each represent distinct buckets. (Bringing the |
| 1281 |
/// total to 16 instead of 8.) This permits spreading the patterns out a bit |
| 1282 |
/// more and thus putting less pressure on verification to be fast. |
| 1283 |
/// |
| 1284 |
/// Each byte in the mask corresponds to a 8-bit bitset, where bit `i` is set |
| 1285 |
/// if and only if the corresponding nybble is in the ith bucket. The index of |
| 1286 |
/// the byte (0-15, inclusive) corresponds to the nybble. |
| 1287 |
#[derive(Clone, Copy, Default)] |
| 1288 |
struct FatMaskBuilder { |
| 1289 |
lo: [u8; 32], |
| 1290 |
hi: [u8; 32], |
| 1291 |
} |
| 1292 |
|
| 1293 |
impl FatMaskBuilder { |
| 1294 |
/// Update this mask by adding the given byte to the given bucket. The |
| 1295 |
/// given bucket must be in the range 0-15. |
| 1296 |
/// |
| 1297 |
/// # Panics |
| 1298 |
/// |
| 1299 |
/// When `bucket >= 16`. |
| 1300 |
fn add(&mut self, bucket: usize, byte: u8) { |
| 1301 |
assert!(bucket < 16); |
| 1302 |
|
| 1303 |
let bucket = u8::try_from(bucket).unwrap(); |
| 1304 |
let byte_lo = usize::from(byte & 0xF); |
| 1305 |
let byte_hi = usize::from((byte >> 4) & 0xF); |
| 1306 |
// Unlike slim teddy, fat teddy only works with AVX2. For fat teddy, |
| 1307 |
// the high 128 bits of our mask correspond to buckets 8-15, while the |
| 1308 |
// low 128 bits correspond to buckets 0-7. |
| 1309 |
if bucket < 8 { |
| 1310 |
self.lo[byte_lo] |= 1 << bucket; |
| 1311 |
self.hi[byte_hi] |= 1 << bucket; |
| 1312 |
} else { |
| 1313 |
self.lo[byte_lo + 16] |= 1 << (bucket % 8); |
| 1314 |
self.hi[byte_hi + 16] |= 1 << (bucket % 8); |
| 1315 |
} |
| 1316 |
} |
| 1317 |
|
| 1318 |
/// Turn this builder into a vector mask. |
| 1319 |
/// |
| 1320 |
/// # Panics |
| 1321 |
/// |
| 1322 |
/// When `V` represents a vector bigger than what `MaskBytes` can contain. |
| 1323 |
/// |
| 1324 |
/// # Safety |
| 1325 |
/// |
| 1326 |
/// Callers must ensure that this is okay to call in the current target for |
| 1327 |
/// the current CPU. |
| 1328 |
#[inline(always)] |
| 1329 |
unsafe fn build<V: Vector>(&self) -> Mask<V> { |
| 1330 |
assert!(V::BYTES <= self.lo.len()); |
| 1331 |
assert!(V::BYTES <= self.hi.len()); |
| 1332 |
Mask { |
| 1333 |
lo: V::load_unaligned(self.lo[..].as_ptr()), |
| 1334 |
hi: V::load_unaligned(self.hi[..].as_ptr()), |
| 1335 |
} |
| 1336 |
} |
| 1337 |
|
| 1338 |
/// A convenience function for building `N` vector masks from a fat |
| 1339 |
/// `Teddy` value. |
| 1340 |
/// |
| 1341 |
/// # Panics |
| 1342 |
/// |
| 1343 |
/// When `V` represents a vector bigger than what `MaskBytes` can contain. |
| 1344 |
/// |
| 1345 |
/// # Safety |
| 1346 |
/// |
| 1347 |
/// Callers must ensure that this is okay to call in the current target for |
| 1348 |
/// the current CPU. |
| 1349 |
#[inline(always)] |
| 1350 |
unsafe fn from_teddy<const BYTES: usize, V: Vector>( |
| 1351 |
teddy: &Teddy<16>, |
| 1352 |
) -> [Mask<V>; BYTES] { |
| 1353 |
// MSRV(1.63): Use core::array::from_fn to just build the array here |
| 1354 |
// instead of creating a vector and turning it into an array. |
| 1355 |
let mut mask_builders = vec![FatMaskBuilder::default(); BYTES]; |
| 1356 |
for (bucket_index, bucket) in teddy.buckets.iter().enumerate() { |
| 1357 |
for pid in bucket.iter().copied() { |
| 1358 |
let pat = teddy.patterns.get(pid); |
| 1359 |
for (i, builder) in mask_builders.iter_mut().enumerate() { |
| 1360 |
builder.add(bucket_index, pat.bytes()[i]); |
| 1361 |
} |
| 1362 |
} |
| 1363 |
} |
| 1364 |
let array = |
| 1365 |
<[FatMaskBuilder; BYTES]>::try_from(mask_builders).unwrap(); |
| 1366 |
array.map(|builder| builder.build()) |
| 1367 |
} |
| 1368 |
} |
| 1369 |
|
| 1370 |
impl Debug for FatMaskBuilder { |
| 1371 |
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result { |
| 1372 |
let (mut parts_lo, mut parts_hi) = (vec![], vec![]); |
| 1373 |
for i in 0..32 { |
| 1374 |
parts_lo.push(format!("{:02}: {:08b}", i, self.lo[i])); |
| 1375 |
parts_hi.push(format!("{:02}: {:08b}", i, self.hi[i])); |
| 1376 |
} |
| 1377 |
f.debug_struct("FatMaskBuilder") |
| 1378 |
.field("lo", &parts_lo) |
| 1379 |
.field("hi", &parts_hi) |
| 1380 |
.finish() |
| 1381 |
} |
| 1382 |
} |