-
Notifications
You must be signed in to change notification settings - Fork 97
/
memchr.rs
1214 lines (1143 loc) · 47.9 KB
/
memchr.rs
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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*!
Generic crate-internal routines for the `memchr` family of functions.
*/
// What follows is a vector algorithm generic over the specific vector
// type to detect the position of one, two or three needles in a haystack.
// From what I know, this is a "classic" algorithm, although I don't
// believe it has been published in any peer reviewed journal. I believe
// it can be found in places like glibc and Go's standard library. It
// appears to be well known and is elaborated on in more detail here:
// https://gms.tf/stdfind-and-memchr-optimizations.html
//
// While the routine below is fairly long and perhaps intimidating, the basic
// idea is actually very simple and can be expressed straight-forwardly in
// pseudo code. The psuedo code below is written for 128 bit vectors, but the
// actual code below works for anything that implements the Vector trait.
//
// needle = (n1 << 15) | (n1 << 14) | ... | (n1 << 1) | n1
// // Note: shift amount is in bytes
//
// while i <= haystack.len() - 16:
// // A 16 byte vector. Each byte in chunk corresponds to a byte in
// // the haystack.
// chunk = haystack[i:i+16]
// // Compare bytes in needle with bytes in chunk. The result is a 16
// // byte chunk where each byte is 0xFF if the corresponding bytes
// // in needle and chunk were equal, or 0x00 otherwise.
// eqs = cmpeq(needle, chunk)
// // Return a 32 bit integer where the most significant 16 bits
// // are always 0 and the lower 16 bits correspond to whether the
// // most significant bit in the correspond byte in `eqs` is set.
// // In other words, `mask as u16` has bit i set if and only if
// // needle[i] == chunk[i].
// mask = movemask(eqs)
//
// // Mask is 0 if there is no match, and non-zero otherwise.
// if mask != 0:
// // trailing_zeros tells us the position of the least significant
// // bit that is set.
// return i + trailing_zeros(mask)
//
// // haystack length may not be a multiple of 16, so search the rest.
// while i < haystack.len():
// if haystack[i] == n1:
// return i
//
// // No match found.
// return NULL
//
// In fact, we could loosely translate the above code to Rust line-for-line
// and it would be a pretty fast algorithm. But, we pull out all the stops
// to go as fast as possible:
//
// 1. We use aligned loads. That is, we do some finagling to make sure our
// primary loop not only proceeds in increments of 16 bytes, but that
// the address of haystack's pointer that we dereference is aligned to
// 16 bytes. 16 is a magic number here because it is the size of SSE2
// 128-bit vector. (For the AVX2 algorithm, 32 is the magic number.)
// Therefore, to get aligned loads, our pointer's address must be evenly
// divisible by 16.
// 2. Our primary loop proceeds 64 bytes at a time instead of 16. It's
// kind of like loop unrolling, but we combine the equality comparisons
// using a vector OR such that we only need to extract a single mask to
// determine whether a match exists or not. If so, then we do some
// book-keeping to determine the precise location but otherwise mush on.
// 3. We use our "chunk" comparison routine in as many places as possible,
// even if it means using unaligned loads. In particular, if haystack
// starts with an unaligned address, then we do an unaligned load to
// search the first 16 bytes. We then start our primary loop at the
// smallest subsequent aligned address, which will actually overlap with
// previously searched bytes. But we're OK with that. We do a similar
// dance at the end of our primary loop. Finally, to avoid a
// byte-at-a-time loop at the end, we do a final 16 byte unaligned load
// that may overlap with a previous load. This is OK because it converts
// a loop into a small number of very fast vector instructions. The overlap
// is OK because we know the place where the overlap occurs does not
// contain a match.
//
// And that's pretty all there is to it. Note that since the below is
// generic and since it's meant to be inlined into routines with a
// `#[target_feature(enable = "...")]` annotation, we must mark all routines as
// both unsafe and `#[inline(always)]`.
//
// The fact that the code below is generic does somewhat inhibit us. For
// example, I've noticed that introducing an unlineable `#[cold]` function to
// handle the match case in the loop generates tighter assembly, but there is
// no way to do this in the generic code below because the generic code doesn't
// know what `target_feature` annotation to apply to the unlineable function.
// We could make such functions part of the `Vector` trait, but we instead live
// with the slightly sub-optimal codegen for now since it doesn't seem to have
// a noticeable perf difference.
use crate::{
ext::Pointer,
vector::{MoveMask, Vector},
};
/// Finds all occurrences of a single byte in a haystack.
#[derive(Clone, Copy, Debug)]
pub(crate) struct One<V> {
s1: u8,
v1: V,
}
impl<V: Vector> One<V> {
/// The number of bytes we examine per each iteration of our search loop.
const LOOP_SIZE: usize = 4 * V::BYTES;
/// Create a new searcher that finds occurrences of the byte given.
#[inline(always)]
pub(crate) unsafe fn new(needle: u8) -> One<V> {
One { s1: needle, v1: V::splat(needle) }
}
/// Returns the needle given to `One::new`.
#[inline(always)]
pub(crate) fn needle1(&self) -> u8 {
self.s1
}
/// Return a pointer to the first occurrence of the needle in the given
/// haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn find_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::first_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
// Search a possibly unaligned chunk at `start`. This covers any part
// of the haystack prior to where aligned loads can start.
if let Some(cur) = self.search_chunk(start, topos) {
return Some(cur);
}
// Set `cur` to the first V-aligned pointer greater than `start`.
let mut cur = start.add(V::BYTES - (start.as_usize() & V::ALIGN));
debug_assert!(cur > start && end.sub(V::BYTES) >= start);
if len >= Self::LOOP_SIZE {
while cur <= end.sub(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(1 * V::BYTES));
let c = V::load_aligned(cur.add(2 * V::BYTES));
let d = V::load_aligned(cur.add(3 * V::BYTES));
let eqa = self.v1.cmpeq(a);
let eqb = self.v1.cmpeq(b);
let eqc = self.v1.cmpeq(c);
let eqd = self.v1.cmpeq(d);
let or1 = eqa.or(eqb);
let or2 = eqc.or(eqd);
let or3 = or1.or(or2);
if or3.movemask_will_have_non_zero() {
let mask = eqa.movemask();
if mask.has_non_zero() {
return Some(cur.add(topos(mask)));
}
let mask = eqb.movemask();
if mask.has_non_zero() {
return Some(cur.add(1 * V::BYTES).add(topos(mask)));
}
let mask = eqc.movemask();
if mask.has_non_zero() {
return Some(cur.add(2 * V::BYTES).add(topos(mask)));
}
let mask = eqd.movemask();
debug_assert!(mask.has_non_zero());
return Some(cur.add(3 * V::BYTES).add(topos(mask)));
}
cur = cur.add(Self::LOOP_SIZE);
}
}
// Handle any leftovers after the aligned loop above. We use unaligned
// loads here, but I believe we are guaranteed that they are aligned
// since `cur` is aligned.
while cur <= end.sub(V::BYTES) {
debug_assert!(end.distance(cur) >= V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
cur = cur.add(V::BYTES);
}
// Finally handle any remaining bytes less than the size of V. In this
// case, our pointer may indeed be unaligned and the load may overlap
// with the previous one. But that's okay since we know the previous
// load didn't lead to a match (otherwise we wouldn't be here).
if cur < end {
debug_assert!(end.distance(cur) < V::BYTES);
cur = cur.sub(V::BYTES - end.distance(cur));
debug_assert_eq!(end.distance(cur), V::BYTES);
return self.search_chunk(cur, topos);
}
None
}
/// Return a pointer to the last occurrence of the needle in the given
/// haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn rfind_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::last_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
if let Some(cur) = self.search_chunk(end.sub(V::BYTES), topos) {
return Some(cur);
}
let mut cur = end.sub(end.as_usize() & V::ALIGN);
debug_assert!(start <= cur && cur <= end);
if len >= Self::LOOP_SIZE {
while cur >= start.add(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
cur = cur.sub(Self::LOOP_SIZE);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(1 * V::BYTES));
let c = V::load_aligned(cur.add(2 * V::BYTES));
let d = V::load_aligned(cur.add(3 * V::BYTES));
let eqa = self.v1.cmpeq(a);
let eqb = self.v1.cmpeq(b);
let eqc = self.v1.cmpeq(c);
let eqd = self.v1.cmpeq(d);
let or1 = eqa.or(eqb);
let or2 = eqc.or(eqd);
let or3 = or1.or(or2);
if or3.movemask_will_have_non_zero() {
let mask = eqd.movemask();
if mask.has_non_zero() {
return Some(cur.add(3 * V::BYTES).add(topos(mask)));
}
let mask = eqc.movemask();
if mask.has_non_zero() {
return Some(cur.add(2 * V::BYTES).add(topos(mask)));
}
let mask = eqb.movemask();
if mask.has_non_zero() {
return Some(cur.add(1 * V::BYTES).add(topos(mask)));
}
let mask = eqa.movemask();
debug_assert!(mask.has_non_zero());
return Some(cur.add(topos(mask)));
}
}
}
while cur >= start.add(V::BYTES) {
debug_assert!(cur.distance(start) >= V::BYTES);
cur = cur.sub(V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
}
if cur > start {
debug_assert!(cur.distance(start) < V::BYTES);
return self.search_chunk(start, topos);
}
None
}
/// Return a count of all matching bytes in the given haystack.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn count_raw(
&self,
start: *const u8,
end: *const u8,
) -> usize {
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let confirm = |b| b == self.needle1();
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
// Set `cur` to the first V-aligned pointer greater than `start`.
let mut cur = start.add(V::BYTES - (start.as_usize() & V::ALIGN));
// Count any matching bytes before we start our aligned loop.
let mut count = count_byte_by_byte(start, cur, confirm);
debug_assert!(cur > start && end.sub(V::BYTES) >= start);
if len >= Self::LOOP_SIZE {
while cur <= end.sub(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(1 * V::BYTES));
let c = V::load_aligned(cur.add(2 * V::BYTES));
let d = V::load_aligned(cur.add(3 * V::BYTES));
let eqa = self.v1.cmpeq(a);
let eqb = self.v1.cmpeq(b);
let eqc = self.v1.cmpeq(c);
let eqd = self.v1.cmpeq(d);
count += eqa.movemask().count_ones();
count += eqb.movemask().count_ones();
count += eqc.movemask().count_ones();
count += eqd.movemask().count_ones();
cur = cur.add(Self::LOOP_SIZE);
}
}
// Handle any leftovers after the aligned loop above. We use unaligned
// loads here, but I believe we are guaranteed that they are aligned
// since `cur` is aligned.
while cur <= end.sub(V::BYTES) {
debug_assert!(end.distance(cur) >= V::BYTES);
let chunk = V::load_unaligned(cur);
count += self.v1.cmpeq(chunk).movemask().count_ones();
cur = cur.add(V::BYTES);
}
// And finally count any leftovers that weren't caught above.
count += count_byte_by_byte(cur, end, confirm);
count
}
/// Search `V::BYTES` starting at `cur` via an unaligned load.
///
/// `mask_to_offset` should be a function that converts a `movemask` to
/// an offset such that `cur.add(offset)` corresponds to a pointer to the
/// match location if one is found. Generally it is expected to use either
/// `mask_to_first_offset` or `mask_to_last_offset`, depending on whether
/// one is implementing a forward or reverse search, respectively.
///
/// # Safety
///
/// `cur` must be a valid pointer and it must be valid to do an unaligned
/// load of size `V::BYTES` at `cur`.
#[inline(always)]
unsafe fn search_chunk(
&self,
cur: *const u8,
mask_to_offset: impl Fn(V::Mask) -> usize,
) -> Option<*const u8> {
let chunk = V::load_unaligned(cur);
let mask = self.v1.cmpeq(chunk).movemask();
if mask.has_non_zero() {
Some(cur.add(mask_to_offset(mask)))
} else {
None
}
}
}
/// Finds all occurrences of two bytes in a haystack.
///
/// That is, this reports matches of one of two possible bytes. For example,
/// searching for `a` or `b` in `afoobar` would report matches at offsets `0`,
/// `4` and `5`.
#[derive(Clone, Copy, Debug)]
pub(crate) struct Two<V> {
s1: u8,
s2: u8,
v1: V,
v2: V,
}
impl<V: Vector> Two<V> {
/// The number of bytes we examine per each iteration of our search loop.
const LOOP_SIZE: usize = 2 * V::BYTES;
/// Create a new searcher that finds occurrences of the byte given.
#[inline(always)]
pub(crate) unsafe fn new(needle1: u8, needle2: u8) -> Two<V> {
Two {
s1: needle1,
s2: needle2,
v1: V::splat(needle1),
v2: V::splat(needle2),
}
}
/// Returns the first needle given to `Two::new`.
#[inline(always)]
pub(crate) fn needle1(&self) -> u8 {
self.s1
}
/// Returns the second needle given to `Two::new`.
#[inline(always)]
pub(crate) fn needle2(&self) -> u8 {
self.s2
}
/// Return a pointer to the first occurrence of one of the needles in the
/// given haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn find_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::first_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
// Search a possibly unaligned chunk at `start`. This covers any part
// of the haystack prior to where aligned loads can start.
if let Some(cur) = self.search_chunk(start, topos) {
return Some(cur);
}
// Set `cur` to the first V-aligned pointer greater than `start`.
let mut cur = start.add(V::BYTES - (start.as_usize() & V::ALIGN));
debug_assert!(cur > start && end.sub(V::BYTES) >= start);
if len >= Self::LOOP_SIZE {
while cur <= end.sub(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(V::BYTES));
let eqa1 = self.v1.cmpeq(a);
let eqb1 = self.v1.cmpeq(b);
let eqa2 = self.v2.cmpeq(a);
let eqb2 = self.v2.cmpeq(b);
let or1 = eqa1.or(eqb1);
let or2 = eqa2.or(eqb2);
let or3 = or1.or(or2);
if or3.movemask_will_have_non_zero() {
let mask = eqa1.movemask().or(eqa2.movemask());
if mask.has_non_zero() {
return Some(cur.add(topos(mask)));
}
let mask = eqb1.movemask().or(eqb2.movemask());
debug_assert!(mask.has_non_zero());
return Some(cur.add(V::BYTES).add(topos(mask)));
}
cur = cur.add(Self::LOOP_SIZE);
}
}
// Handle any leftovers after the aligned loop above. We use unaligned
// loads here, but I believe we are guaranteed that they are aligned
// since `cur` is aligned.
while cur <= end.sub(V::BYTES) {
debug_assert!(end.distance(cur) >= V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
cur = cur.add(V::BYTES);
}
// Finally handle any remaining bytes less than the size of V. In this
// case, our pointer may indeed be unaligned and the load may overlap
// with the previous one. But that's okay since we know the previous
// load didn't lead to a match (otherwise we wouldn't be here).
if cur < end {
debug_assert!(end.distance(cur) < V::BYTES);
cur = cur.sub(V::BYTES - end.distance(cur));
debug_assert_eq!(end.distance(cur), V::BYTES);
return self.search_chunk(cur, topos);
}
None
}
/// Return a pointer to the last occurrence of the needle in the given
/// haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn rfind_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::last_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
if let Some(cur) = self.search_chunk(end.sub(V::BYTES), topos) {
return Some(cur);
}
let mut cur = end.sub(end.as_usize() & V::ALIGN);
debug_assert!(start <= cur && cur <= end);
if len >= Self::LOOP_SIZE {
while cur >= start.add(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
cur = cur.sub(Self::LOOP_SIZE);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(V::BYTES));
let eqa1 = self.v1.cmpeq(a);
let eqb1 = self.v1.cmpeq(b);
let eqa2 = self.v2.cmpeq(a);
let eqb2 = self.v2.cmpeq(b);
let or1 = eqa1.or(eqb1);
let or2 = eqa2.or(eqb2);
let or3 = or1.or(or2);
if or3.movemask_will_have_non_zero() {
let mask = eqb1.movemask().or(eqb2.movemask());
if mask.has_non_zero() {
return Some(cur.add(V::BYTES).add(topos(mask)));
}
let mask = eqa1.movemask().or(eqa2.movemask());
debug_assert!(mask.has_non_zero());
return Some(cur.add(topos(mask)));
}
}
}
while cur >= start.add(V::BYTES) {
debug_assert!(cur.distance(start) >= V::BYTES);
cur = cur.sub(V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
}
if cur > start {
debug_assert!(cur.distance(start) < V::BYTES);
return self.search_chunk(start, topos);
}
None
}
/// Search `V::BYTES` starting at `cur` via an unaligned load.
///
/// `mask_to_offset` should be a function that converts a `movemask` to
/// an offset such that `cur.add(offset)` corresponds to a pointer to the
/// match location if one is found. Generally it is expected to use either
/// `mask_to_first_offset` or `mask_to_last_offset`, depending on whether
/// one is implementing a forward or reverse search, respectively.
///
/// # Safety
///
/// `cur` must be a valid pointer and it must be valid to do an unaligned
/// load of size `V::BYTES` at `cur`.
#[inline(always)]
unsafe fn search_chunk(
&self,
cur: *const u8,
mask_to_offset: impl Fn(V::Mask) -> usize,
) -> Option<*const u8> {
let chunk = V::load_unaligned(cur);
let eq1 = self.v1.cmpeq(chunk);
let eq2 = self.v2.cmpeq(chunk);
let mask = eq1.or(eq2).movemask();
if mask.has_non_zero() {
let mask1 = eq1.movemask();
let mask2 = eq2.movemask();
Some(cur.add(mask_to_offset(mask1.or(mask2))))
} else {
None
}
}
}
/// Finds all occurrences of two bytes in a haystack.
///
/// That is, this reports matches of one of two possible bytes. For example,
/// searching for `a` or `b` in `afoobar` would report matches at offsets `0`,
/// `4` and `5`.
#[derive(Clone, Copy, Debug)]
pub(crate) struct Three<V> {
s1: u8,
s2: u8,
s3: u8,
v1: V,
v2: V,
v3: V,
}
impl<V: Vector> Three<V> {
/// The number of bytes we examine per each iteration of our search loop.
const LOOP_SIZE: usize = 2 * V::BYTES;
/// Create a new searcher that finds occurrences of the byte given.
#[inline(always)]
pub(crate) unsafe fn new(
needle1: u8,
needle2: u8,
needle3: u8,
) -> Three<V> {
Three {
s1: needle1,
s2: needle2,
s3: needle3,
v1: V::splat(needle1),
v2: V::splat(needle2),
v3: V::splat(needle3),
}
}
/// Returns the first needle given to `Three::new`.
#[inline(always)]
pub(crate) fn needle1(&self) -> u8 {
self.s1
}
/// Returns the second needle given to `Three::new`.
#[inline(always)]
pub(crate) fn needle2(&self) -> u8 {
self.s2
}
/// Returns the third needle given to `Three::new`.
#[inline(always)]
pub(crate) fn needle3(&self) -> u8 {
self.s3
}
/// Return a pointer to the first occurrence of one of the needles in the
/// given haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn find_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::first_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
// Search a possibly unaligned chunk at `start`. This covers any part
// of the haystack prior to where aligned loads can start.
if let Some(cur) = self.search_chunk(start, topos) {
return Some(cur);
}
// Set `cur` to the first V-aligned pointer greater than `start`.
let mut cur = start.add(V::BYTES - (start.as_usize() & V::ALIGN));
debug_assert!(cur > start && end.sub(V::BYTES) >= start);
if len >= Self::LOOP_SIZE {
while cur <= end.sub(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(V::BYTES));
let eqa1 = self.v1.cmpeq(a);
let eqb1 = self.v1.cmpeq(b);
let eqa2 = self.v2.cmpeq(a);
let eqb2 = self.v2.cmpeq(b);
let eqa3 = self.v3.cmpeq(a);
let eqb3 = self.v3.cmpeq(b);
let or1 = eqa1.or(eqb1);
let or2 = eqa2.or(eqb2);
let or3 = eqa3.or(eqb3);
let or4 = or1.or(or2);
let or5 = or3.or(or4);
if or5.movemask_will_have_non_zero() {
let mask = eqa1
.movemask()
.or(eqa2.movemask())
.or(eqa3.movemask());
if mask.has_non_zero() {
return Some(cur.add(topos(mask)));
}
let mask = eqb1
.movemask()
.or(eqb2.movemask())
.or(eqb3.movemask());
debug_assert!(mask.has_non_zero());
return Some(cur.add(V::BYTES).add(topos(mask)));
}
cur = cur.add(Self::LOOP_SIZE);
}
}
// Handle any leftovers after the aligned loop above. We use unaligned
// loads here, but I believe we are guaranteed that they are aligned
// since `cur` is aligned.
while cur <= end.sub(V::BYTES) {
debug_assert!(end.distance(cur) >= V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
cur = cur.add(V::BYTES);
}
// Finally handle any remaining bytes less than the size of V. In this
// case, our pointer may indeed be unaligned and the load may overlap
// with the previous one. But that's okay since we know the previous
// load didn't lead to a match (otherwise we wouldn't be here).
if cur < end {
debug_assert!(end.distance(cur) < V::BYTES);
cur = cur.sub(V::BYTES - end.distance(cur));
debug_assert_eq!(end.distance(cur), V::BYTES);
return self.search_chunk(cur, topos);
}
None
}
/// Return a pointer to the last occurrence of the needle in the given
/// haystack. If no such occurrence exists, then `None` is returned.
///
/// When a match is found, the pointer returned is guaranteed to be
/// `>= start` and `< end`.
///
/// # Safety
///
/// * It must be the case that `start < end` and that the distance between
/// them is at least equal to `V::BYTES`. That is, it must always be valid
/// to do at least an unaligned load of `V` at `start`.
/// * Both `start` and `end` must be valid for reads.
/// * Both `start` and `end` must point to an initialized value.
/// * Both `start` and `end` must point to the same allocated object and
/// must either be in bounds or at most one byte past the end of the
/// allocated object.
/// * Both `start` and `end` must be _derived from_ a pointer to the same
/// object.
/// * The distance between `start` and `end` must not overflow `isize`.
/// * The distance being in bounds must not rely on "wrapping around" the
/// address space.
#[inline(always)]
pub(crate) unsafe fn rfind_raw(
&self,
start: *const u8,
end: *const u8,
) -> Option<*const u8> {
// If we want to support vectors bigger than 256 bits, we probably
// need to move up to using a u64 for the masks used below. Currently
// they are 32 bits, which means we're SOL for vectors that need masks
// bigger than 32 bits. Overall unclear until there's a use case.
debug_assert!(V::BYTES <= 32, "vector cannot be bigger than 32 bytes");
let topos = V::Mask::last_offset;
let len = end.distance(start);
debug_assert!(
len >= V::BYTES,
"haystack has length {}, but must be at least {}",
len,
V::BYTES
);
if let Some(cur) = self.search_chunk(end.sub(V::BYTES), topos) {
return Some(cur);
}
let mut cur = end.sub(end.as_usize() & V::ALIGN);
debug_assert!(start <= cur && cur <= end);
if len >= Self::LOOP_SIZE {
while cur >= start.add(Self::LOOP_SIZE) {
debug_assert_eq!(0, cur.as_usize() % V::BYTES);
cur = cur.sub(Self::LOOP_SIZE);
let a = V::load_aligned(cur);
let b = V::load_aligned(cur.add(V::BYTES));
let eqa1 = self.v1.cmpeq(a);
let eqb1 = self.v1.cmpeq(b);
let eqa2 = self.v2.cmpeq(a);
let eqb2 = self.v2.cmpeq(b);
let eqa3 = self.v3.cmpeq(a);
let eqb3 = self.v3.cmpeq(b);
let or1 = eqa1.or(eqb1);
let or2 = eqa2.or(eqb2);
let or3 = eqa3.or(eqb3);
let or4 = or1.or(or2);
let or5 = or3.or(or4);
if or5.movemask_will_have_non_zero() {
let mask = eqb1
.movemask()
.or(eqb2.movemask())
.or(eqb3.movemask());
if mask.has_non_zero() {
return Some(cur.add(V::BYTES).add(topos(mask)));
}
let mask = eqa1
.movemask()
.or(eqa2.movemask())
.or(eqa3.movemask());
debug_assert!(mask.has_non_zero());
return Some(cur.add(topos(mask)));
}
}
}
while cur >= start.add(V::BYTES) {
debug_assert!(cur.distance(start) >= V::BYTES);
cur = cur.sub(V::BYTES);
if let Some(cur) = self.search_chunk(cur, topos) {
return Some(cur);
}
}
if cur > start {
debug_assert!(cur.distance(start) < V::BYTES);
return self.search_chunk(start, topos);
}
None
}
/// Search `V::BYTES` starting at `cur` via an unaligned load.
///
/// `mask_to_offset` should be a function that converts a `movemask` to
/// an offset such that `cur.add(offset)` corresponds to a pointer to the
/// match location if one is found. Generally it is expected to use either
/// `mask_to_first_offset` or `mask_to_last_offset`, depending on whether
/// one is implementing a forward or reverse search, respectively.
///
/// # Safety
///
/// `cur` must be a valid pointer and it must be valid to do an unaligned
/// load of size `V::BYTES` at `cur`.
#[inline(always)]
unsafe fn search_chunk(
&self,
cur: *const u8,
mask_to_offset: impl Fn(V::Mask) -> usize,
) -> Option<*const u8> {
let chunk = V::load_unaligned(cur);
let eq1 = self.v1.cmpeq(chunk);
let eq2 = self.v2.cmpeq(chunk);
let eq3 = self.v3.cmpeq(chunk);
let mask = eq1.or(eq2).or(eq3).movemask();
if mask.has_non_zero() {
let mask1 = eq1.movemask();
let mask2 = eq2.movemask();
let mask3 = eq3.movemask();
Some(cur.add(mask_to_offset(mask1.or(mask2).or(mask3))))
} else {
None
}
}
}
/// An iterator over all occurrences of a set of bytes in a haystack.
///
/// This iterator implements the routines necessary to provide a
/// `DoubleEndedIterator` impl, which means it can also be used to find
/// occurrences in reverse order.
///
/// The lifetime parameters are as follows:
///
/// * `'h` refers to the lifetime of the haystack being searched.
///
/// This type is intended to be used to implement all iterators for the
/// `memchr` family of functions. It handles a tiny bit of marginally tricky
/// raw pointer math, but otherwise expects the caller to provide `find_raw`
/// and `rfind_raw` routines for each call of `next` and `next_back`,
/// respectively.
#[derive(Clone, Debug)]
pub(crate) struct Iter<'h> {
/// The original starting point into the haystack. We use this to convert