summaryrefslogtreecommitdiffstats
path: root/vendor/regex/src/dfa.rs
blob: 4aee8039c6c1ac8dc75ceebe17b05c8ccbb74b33 (plain)
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
/*!
The DFA matching engine.

A DFA provides faster matching because the engine is in exactly one state at
any point in time. In the NFA, there may be multiple active states, and
considerable CPU cycles are spent shuffling them around. In finite automata
speak, the DFA follows epsilon transitions in the regex far less than the NFA.

A DFA is a classic trade off between time and space. The NFA is slower, but
its memory requirements are typically small and predictable. The DFA is faster,
but given the right regex and the right input, the number of states in the
DFA can grow exponentially. To mitigate this space problem, we do two things:

1. We implement an *online* DFA. That is, the DFA is constructed from the NFA
   during a search. When a new state is computed, it is stored in a cache so
   that it may be reused. An important consequence of this implementation
   is that states that are never reached for a particular input are never
   computed. (This is impossible in an "offline" DFA which needs to compute
   all possible states up front.)
2. If the cache gets too big, we wipe it and continue matching.

In pathological cases, a new state can be created for every byte of input.
(e.g., The regex `(a|b)*a(a|b){20}` on a long sequence of a's and b's.)
In this case, performance regresses to slightly slower than the full NFA
simulation, in large part because the cache becomes useless. If the cache
is wiped too frequently, the DFA quits and control falls back to one of the
NFA simulations.

Because of the "lazy" nature of this DFA, the inner matching loop is
considerably more complex than one might expect out of a DFA. A number of
tricks are employed to make it fast. Tread carefully.

N.B. While this implementation is heavily commented, Russ Cox's series of
articles on regexes is strongly recommended: https://swtch.com/~rsc/regexp/
(As is the DFA implementation in RE2, which heavily influenced this
implementation.)
*/

use std::collections::HashMap;
use std::fmt;
use std::iter::repeat;
use std::mem;
use std::sync::Arc;

use crate::exec::ProgramCache;
use crate::prog::{Inst, Program};
use crate::sparse::SparseSet;

/// Return true if and only if the given program can be executed by a DFA.
///
/// Generally, a DFA is always possible. A pathological case where it is not
/// possible is if the number of NFA states exceeds `u32::MAX`, in which case,
/// this function will return false.
///
/// This function will also return false if the given program has any Unicode
/// instructions (Char or Ranges) since the DFA operates on bytes only.
pub fn can_exec(insts: &Program) -> bool {
    use crate::prog::Inst::*;
    // If for some reason we manage to allocate a regex program with more
    // than i32::MAX instructions, then we can't execute the DFA because we
    // use 32 bit instruction pointer deltas for memory savings.
    // If i32::MAX is the largest positive delta,
    // then -i32::MAX == i32::MIN + 1 is the largest negative delta,
    // and we are OK to use 32 bits.
    if insts.dfa_size_limit == 0 || insts.len() > ::std::i32::MAX as usize {
        return false;
    }
    for inst in insts {
        match *inst {
            Char(_) | Ranges(_) => return false,
            EmptyLook(_) | Match(_) | Save(_) | Split(_) | Bytes(_) => {}
        }
    }
    true
}

/// A reusable cache of DFA states.
///
/// This cache is reused between multiple invocations of the same regex
/// program. (It is not shared simultaneously between threads. If there is
/// contention, then new caches are created.)
#[derive(Debug)]
pub struct Cache {
    /// Group persistent DFA related cache state together. The sparse sets
    /// listed below are used as scratch space while computing uncached states.
    inner: CacheInner,
    /// qcur and qnext are ordered sets with constant time
    /// addition/membership/clearing-whole-set and linear time iteration. They
    /// are used to manage the sets of NFA states in DFA states when computing
    /// cached DFA states. In particular, the order of the NFA states matters
    /// for leftmost-first style matching. Namely, when computing a cached
    /// state, the set of NFA states stops growing as soon as the first Match
    /// instruction is observed.
    qcur: SparseSet,
    qnext: SparseSet,
}

/// `CacheInner` is logically just a part of Cache, but groups together fields
/// that aren't passed as function parameters throughout search. (This split
/// is mostly an artifact of the borrow checker. It is happily paid.)
#[derive(Debug)]
struct CacheInner {
    /// A cache of pre-compiled DFA states, keyed by the set of NFA states
    /// and the set of empty-width flags set at the byte in the input when the
    /// state was observed.
    ///
    /// A StatePtr is effectively a `*State`, but to avoid various inconvenient
    /// things, we just pass indexes around manually. The performance impact of
    /// this is probably an instruction or two in the inner loop. However, on
    /// 64 bit, each StatePtr is half the size of a *State.
    compiled: StateMap,
    /// The transition table.
    ///
    /// The transition table is laid out in row-major order, where states are
    /// rows and the transitions for each state are columns. At a high level,
    /// given state `s` and byte `b`, the next state can be found at index
    /// `s * 256 + b`.
    ///
    /// This is, of course, a lie. A StatePtr is actually a pointer to the
    /// *start* of a row in this table. When indexing in the DFA's inner loop,
    /// this removes the need to multiply the StatePtr by the stride. Yes, it
    /// matters. This reduces the number of states we can store, but: the
    /// stride is rarely 256 since we define transitions in terms of
    /// *equivalence classes* of bytes. Each class corresponds to a set of
    /// bytes that never discriminate a distinct path through the DFA from each
    /// other.
    trans: Transitions,
    /// A set of cached start states, which are limited to the number of
    /// permutations of flags set just before the initial byte of input. (The
    /// index into this vec is a `EmptyFlags`.)
    ///
    /// N.B. A start state can be "dead" (i.e., no possible match), so we
    /// represent it with a StatePtr.
    start_states: Vec<StatePtr>,
    /// Stack scratch space used to follow epsilon transitions in the NFA.
    /// (This permits us to avoid recursion.)
    ///
    /// The maximum stack size is the number of NFA states.
    stack: Vec<InstPtr>,
    /// The total number of times this cache has been flushed by the DFA
    /// because of space constraints.
    flush_count: u64,
    /// The total heap size of the DFA's cache. We use this to determine when
    /// we should flush the cache.
    size: usize,
    /// Scratch space used when building instruction pointer lists for new
    /// states. This helps amortize allocation.
    insts_scratch_space: Vec<u8>,
}

/// The transition table.
///
/// It is laid out in row-major order, with states as rows and byte class
/// transitions as columns.
///
/// The transition table is responsible for producing valid `StatePtrs`. A
/// `StatePtr` points to the start of a particular row in this table. When
/// indexing to find the next state this allows us to avoid a multiplication
/// when computing an index into the table.
#[derive(Clone)]
struct Transitions {
    /// The table.
    table: Vec<StatePtr>,
    /// The stride.
    num_byte_classes: usize,
}

/// Fsm encapsulates the actual execution of the DFA.
#[derive(Debug)]
pub struct Fsm<'a> {
    /// prog contains the NFA instruction opcodes. DFA execution uses either
    /// the `dfa` instructions or the `dfa_reverse` instructions from
    /// `exec::ExecReadOnly`. (It never uses `ExecReadOnly.nfa`, which may have
    /// Unicode opcodes that cannot be executed by the DFA.)
    prog: &'a Program,
    /// The start state. We record it here because the pointer may change
    /// when the cache is wiped.
    start: StatePtr,
    /// The current position in the input.
    at: usize,
    /// Should we quit after seeing the first match? e.g., When the caller
    /// uses `is_match` or `shortest_match`.
    quit_after_match: bool,
    /// The last state that matched.
    ///
    /// When no match has occurred, this is set to STATE_UNKNOWN.
    ///
    /// This is only useful when matching regex sets. The last match state
    /// is useful because it contains all of the match instructions seen,
    /// thereby allowing us to enumerate which regexes in the set matched.
    last_match_si: StatePtr,
    /// The input position of the last cache flush. We use this to determine
    /// if we're thrashing in the cache too often. If so, the DFA quits so
    /// that we can fall back to the NFA algorithm.
    last_cache_flush: usize,
    /// All cached DFA information that is persisted between searches.
    cache: &'a mut CacheInner,
}

/// The result of running the DFA.
///
/// Generally, the result is either a match or not a match, but sometimes the
/// DFA runs too slowly because the cache size is too small. In that case, it
/// gives up with the intent of falling back to the NFA algorithm.
///
/// The DFA can also give up if it runs out of room to create new states, or if
/// it sees non-ASCII bytes in the presence of a Unicode word boundary.
#[derive(Clone, Debug)]
pub enum Result<T> {
    Match(T),
    NoMatch(usize),
    Quit,
}

impl<T> Result<T> {
    /// Returns true if this result corresponds to a match.
    pub fn is_match(&self) -> bool {
        match *self {
            Result::Match(_) => true,
            Result::NoMatch(_) | Result::Quit => false,
        }
    }

    /// Maps the given function onto T and returns the result.
    ///
    /// If this isn't a match, then this is a no-op.
    #[cfg(feature = "perf-literal")]
    pub fn map<U, F: FnMut(T) -> U>(self, mut f: F) -> Result<U> {
        match self {
            Result::Match(t) => Result::Match(f(t)),
            Result::NoMatch(x) => Result::NoMatch(x),
            Result::Quit => Result::Quit,
        }
    }

    /// Sets the non-match position.
    ///
    /// If this isn't a non-match, then this is a no-op.
    fn set_non_match(self, at: usize) -> Result<T> {
        match self {
            Result::NoMatch(_) => Result::NoMatch(at),
            r => r,
        }
    }
}

/// `State` is a DFA state. It contains an ordered set of NFA states (not
/// necessarily complete) and a smattering of flags.
///
/// The flags are packed into the first byte of data.
///
/// States don't carry their transitions. Instead, transitions are stored in
/// a single row-major table.
///
/// Delta encoding is used to store the instruction pointers.
/// The first instruction pointer is stored directly starting
/// at data[1], and each following pointer is stored as an offset
/// to the previous one. If a delta is in the range -127..127,
/// it is packed into a single byte; Otherwise the byte 128 (-128 as an i8)
/// is coded as a flag, followed by 4 bytes encoding the delta.
#[derive(Clone, Eq, Hash, PartialEq)]
struct State {
    data: Arc<[u8]>,
}

/// `InstPtr` is a 32 bit pointer into a sequence of opcodes (i.e., it indexes
/// an NFA state).
///
/// Throughout this library, this is usually set to `usize`, but we force a
/// `u32` here for the DFA to save on space.
type InstPtr = u32;

/// Adds ip to data using delta encoding with respect to prev.
///
/// After completion, `data` will contain `ip` and `prev` will be set to `ip`.
fn push_inst_ptr(data: &mut Vec<u8>, prev: &mut InstPtr, ip: InstPtr) {
    let delta = (ip as i32) - (*prev as i32);
    write_vari32(data, delta);
    *prev = ip;
}

struct InstPtrs<'a> {
    base: usize,
    data: &'a [u8],
}

impl<'a> Iterator for InstPtrs<'a> {
    type Item = usize;

    fn next(&mut self) -> Option<usize> {
        if self.data.is_empty() {
            return None;
        }
        let (delta, nread) = read_vari32(self.data);
        let base = self.base as i32 + delta;
        debug_assert!(base >= 0);
        debug_assert!(nread > 0);
        self.data = &self.data[nread..];
        self.base = base as usize;
        Some(self.base)
    }
}

impl State {
    fn flags(&self) -> StateFlags {
        StateFlags(self.data[0])
    }

    fn inst_ptrs(&self) -> InstPtrs<'_> {
        InstPtrs { base: 0, data: &self.data[1..] }
    }
}

/// `StatePtr` is a 32 bit pointer to the start of a row in the transition
/// table.
///
/// It has many special values. There are two types of special values:
/// sentinels and flags.
///
/// Sentinels corresponds to special states that carry some kind of
/// significance. There are three such states: unknown, dead and quit states.
///
/// Unknown states are states that haven't been computed yet. They indicate
/// that a transition should be filled in that points to either an existing
/// cached state or a new state altogether. In general, an unknown state means
/// "follow the NFA's epsilon transitions."
///
/// Dead states are states that can never lead to a match, no matter what
/// subsequent input is observed. This means that the DFA should quit
/// immediately and return the longest match it has found thus far.
///
/// Quit states are states that imply the DFA is not capable of matching the
/// regex correctly. Currently, this is only used when a Unicode word boundary
/// exists in the regex *and* a non-ASCII byte is observed.
///
/// The other type of state pointer is a state pointer with special flag bits.
/// There are two flags: a start flag and a match flag. The lower bits of both
/// kinds always contain a "valid" `StatePtr` (indicated by the `STATE_MAX`
/// mask).
///
/// The start flag means that the state is a start state, and therefore may be
/// subject to special prefix scanning optimizations.
///
/// The match flag means that the state is a match state, and therefore the
/// current position in the input (while searching) should be recorded.
///
/// The above exists mostly in the service of making the inner loop fast.
/// In particular, the inner *inner* loop looks something like this:
///
/// ```ignore
/// while state <= STATE_MAX and i < len(text):
///     state = state.next[i]
/// ```
///
/// This is nice because it lets us execute a lazy DFA as if it were an
/// entirely offline DFA (i.e., with very few instructions). The loop will
/// quit only when we need to examine a case that needs special attention.
type StatePtr = u32;

/// An unknown state means that the state has not been computed yet, and that
/// the only way to progress is to compute it.
const STATE_UNKNOWN: StatePtr = 1 << 31;

/// A dead state means that the state has been computed and it is known that
/// once it is entered, no future match can ever occur.
const STATE_DEAD: StatePtr = STATE_UNKNOWN + 1;

/// A quit state means that the DFA came across some input that it doesn't
/// know how to process correctly. The DFA should quit and another matching
/// engine should be run in its place.
const STATE_QUIT: StatePtr = STATE_DEAD + 1;

/// A start state is a state that the DFA can start in.
///
/// Note that start states have their lower bits set to a state pointer.
const STATE_START: StatePtr = 1 << 30;

/// A match state means that the regex has successfully matched.
///
/// Note that match states have their lower bits set to a state pointer.
const STATE_MATCH: StatePtr = 1 << 29;

/// The maximum state pointer. This is useful to mask out the "valid" state
/// pointer from a state with the "start" or "match" bits set.
///
/// It doesn't make sense to use this with unknown, dead or quit state
/// pointers, since those pointers are sentinels and never have their lower
/// bits set to anything meaningful.
const STATE_MAX: StatePtr = STATE_MATCH - 1;

/// Byte is a u8 in spirit, but a u16 in practice so that we can represent the
/// special EOF sentinel value.
#[derive(Copy, Clone, Debug)]
struct Byte(u16);

/// A set of flags for zero-width assertions.
#[derive(Clone, Copy, Eq, Debug, Default, Hash, PartialEq)]
struct EmptyFlags {
    start: bool,
    end: bool,
    start_line: bool,
    end_line: bool,
    word_boundary: bool,
    not_word_boundary: bool,
}

/// A set of flags describing various configurations of a DFA state. This is
/// represented by a `u8` so that it is compact.
#[derive(Clone, Copy, Eq, Default, Hash, PartialEq)]
struct StateFlags(u8);

impl Cache {
    /// Create new empty cache for the DFA engine.
    pub fn new(prog: &Program) -> Self {
        // We add 1 to account for the special EOF byte.
        let num_byte_classes = (prog.byte_classes[255] as usize + 1) + 1;
        let starts = vec![STATE_UNKNOWN; 256];
        let mut cache = Cache {
            inner: CacheInner {
                compiled: StateMap::new(num_byte_classes),
                trans: Transitions::new(num_byte_classes),
                start_states: starts,
                stack: vec![],
                flush_count: 0,
                size: 0,
                insts_scratch_space: vec![],
            },
            qcur: SparseSet::new(prog.insts.len()),
            qnext: SparseSet::new(prog.insts.len()),
        };
        cache.inner.reset_size();
        cache
    }
}

impl CacheInner {
    /// Resets the cache size to account for fixed costs, such as the program
    /// and stack sizes.
    fn reset_size(&mut self) {
        self.size = (self.start_states.len() * mem::size_of::<StatePtr>())
            + (self.stack.len() * mem::size_of::<InstPtr>());
    }
}

impl<'a> Fsm<'a> {
    #[cfg_attr(feature = "perf-inline", inline(always))]
    pub fn forward(
        prog: &'a Program,
        cache: &ProgramCache,
        quit_after_match: bool,
        text: &[u8],
        at: usize,
    ) -> Result<usize> {
        let mut cache = cache.borrow_mut();
        let cache = &mut cache.dfa;
        let mut dfa = Fsm {
            prog: prog,
            start: 0, // filled in below
            at: at,
            quit_after_match: quit_after_match,
            last_match_si: STATE_UNKNOWN,
            last_cache_flush: at,
            cache: &mut cache.inner,
        };
        let (empty_flags, state_flags) = dfa.start_flags(text, at);
        dfa.start =
            match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
                None => return Result::Quit,
                Some(STATE_DEAD) => return Result::NoMatch(at),
                Some(si) => si,
            };
        debug_assert!(dfa.start != STATE_UNKNOWN);
        dfa.exec_at(&mut cache.qcur, &mut cache.qnext, text)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    pub fn reverse(
        prog: &'a Program,
        cache: &ProgramCache,
        quit_after_match: bool,
        text: &[u8],
        at: usize,
    ) -> Result<usize> {
        let mut cache = cache.borrow_mut();
        let cache = &mut cache.dfa_reverse;
        let mut dfa = Fsm {
            prog: prog,
            start: 0, // filled in below
            at: at,
            quit_after_match: quit_after_match,
            last_match_si: STATE_UNKNOWN,
            last_cache_flush: at,
            cache: &mut cache.inner,
        };
        let (empty_flags, state_flags) = dfa.start_flags_reverse(text, at);
        dfa.start =
            match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
                None => return Result::Quit,
                Some(STATE_DEAD) => return Result::NoMatch(at),
                Some(si) => si,
            };
        debug_assert!(dfa.start != STATE_UNKNOWN);
        dfa.exec_at_reverse(&mut cache.qcur, &mut cache.qnext, text)
    }

    #[cfg_attr(feature = "perf-inline", inline(always))]
    pub fn forward_many(
        prog: &'a Program,
        cache: &ProgramCache,
        matches: &mut [bool],
        text: &[u8],
        at: usize,
    ) -> Result<usize> {
        debug_assert!(matches.len() == prog.matches.len());
        let mut cache = cache.borrow_mut();
        let cache = &mut cache.dfa;
        let mut dfa = Fsm {
            prog: prog,
            start: 0, // filled in below
            at: at,
            quit_after_match: false,
            last_match_si: STATE_UNKNOWN,
            last_cache_flush: at,
            cache: &mut cache.inner,
        };
        let (empty_flags, state_flags) = dfa.start_flags(text, at);
        dfa.start =
            match dfa.start_state(&mut cache.qcur, empty_flags, state_flags) {
                None => return Result::Quit,
                Some(STATE_DEAD) => return Result::NoMatch(at),
                Some(si) => si,
            };
        debug_assert!(dfa.start != STATE_UNKNOWN);
        let result = dfa.exec_at(&mut cache.qcur, &mut cache.qnext, text);
        if result.is_match() {
            if matches.len() == 1 {
                matches[0] = true;
            } else {
                debug_assert!(dfa.last_match_si != STATE_UNKNOWN);
                debug_assert!(dfa.last_match_si != STATE_DEAD);
                for ip in dfa.state(dfa.last_match_si).inst_ptrs() {
                    if let Inst::Match(slot) = dfa.prog[ip] {
                        matches[slot] = true;
                    }
                }
            }
        }
        result
    }

    /// Executes the DFA on a forward NFA.
    ///
    /// {qcur,qnext} are scratch ordered sets which may be non-empty.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn exec_at(
        &mut self,
        qcur: &mut SparseSet,
        qnext: &mut SparseSet,
        text: &[u8],
    ) -> Result<usize> {
        // For the most part, the DFA is basically:
        //
        //   last_match = null
        //   while current_byte != EOF:
        //     si = current_state.next[current_byte]
        //     if si is match
        //       last_match = si
        //   return last_match
        //
        // However, we need to deal with a few things:
        //
        //   1. This is an *online* DFA, so the current state's next list
        //      may not point to anywhere yet, so we must go out and compute
        //      them. (They are then cached into the current state's next list
        //      to avoid re-computation.)
        //   2. If we come across a state that is known to be dead (i.e., never
        //      leads to a match), then we can quit early.
        //   3. If the caller just wants to know if a match occurs, then we
        //      can quit as soon as we know we have a match. (Full leftmost
        //      first semantics require continuing on.)
        //   4. If we're in the start state, then we can use a pre-computed set
        //      of prefix literals to skip quickly along the input.
        //   5. After the input is exhausted, we run the DFA on one symbol
        //      that stands for EOF. This is useful for handling empty width
        //      assertions.
        //   6. We can't actually do state.next[byte]. Instead, we have to do
        //      state.next[byte_classes[byte]], which permits us to keep the
        //      'next' list very small.
        //
        // Since there's a bunch of extra stuff we need to consider, we do some
        // pretty hairy tricks to get the inner loop to run as fast as
        // possible.
        debug_assert!(!self.prog.is_reverse);

        // The last match is the currently known ending match position. It is
        // reported as an index to the most recent byte that resulted in a
        // transition to a match state and is always stored in capture slot `1`
        // when searching forwards. Its maximum value is `text.len()`.
        let mut result = Result::NoMatch(self.at);
        let (mut prev_si, mut next_si) = (self.start, self.start);
        let mut at = self.at;
        while at < text.len() {
            // This is the real inner loop. We take advantage of special bits
            // set in the state pointer to determine whether a state is in the
            // "common" case or not. Specifically, the common case is a
            // non-match non-start non-dead state that has already been
            // computed. So long as we remain in the common case, this inner
            // loop will chew through the input.
            //
            // We also unroll the loop 4 times to amortize the cost of checking
            // whether we've consumed the entire input. We are also careful
            // to make sure that `prev_si` always represents the previous state
            // and `next_si` always represents the next state after the loop
            // exits, even if it isn't always true inside the loop.
            while next_si <= STATE_MAX && at < text.len() {
                // Argument for safety is in the definition of next_si.
                prev_si = unsafe { self.next_si(next_si, text, at) };
                at += 1;
                if prev_si > STATE_MAX || at + 2 >= text.len() {
                    mem::swap(&mut prev_si, &mut next_si);
                    break;
                }
                next_si = unsafe { self.next_si(prev_si, text, at) };
                at += 1;
                if next_si > STATE_MAX {
                    break;
                }
                prev_si = unsafe { self.next_si(next_si, text, at) };
                at += 1;
                if prev_si > STATE_MAX {
                    mem::swap(&mut prev_si, &mut next_si);
                    break;
                }
                next_si = unsafe { self.next_si(prev_si, text, at) };
                at += 1;
            }
            if next_si & STATE_MATCH > 0 {
                // A match state is outside of the common case because it needs
                // special case analysis. In particular, we need to record the
                // last position as having matched and possibly quit the DFA if
                // we don't need to keep matching.
                next_si &= !STATE_MATCH;
                result = Result::Match(at - 1);
                if self.quit_after_match {
                    return result;
                }
                self.last_match_si = next_si;
                prev_si = next_si;

                // This permits short-circuiting when matching a regex set.
                // In particular, if this DFA state contains only match states,
                // then it's impossible to extend the set of matches since
                // match states are final. Therefore, we can quit.
                if self.prog.matches.len() > 1 {
                    let state = self.state(next_si);
                    let just_matches =
                        state.inst_ptrs().all(|ip| self.prog[ip].is_match());
                    if just_matches {
                        return result;
                    }
                }

                // Another inner loop! If the DFA stays in this particular
                // match state, then we can rip through all of the input
                // very quickly, and only recording the match location once
                // we've left this particular state.
                let cur = at;
                while (next_si & !STATE_MATCH) == prev_si
                    && at + 2 < text.len()
                {
                    // Argument for safety is in the definition of next_si.
                    next_si = unsafe {
                        self.next_si(next_si & !STATE_MATCH, text, at)
                    };
                    at += 1;
                }
                if at > cur {
                    result = Result::Match(at - 2);
                }
            } else if next_si & STATE_START > 0 {
                // A start state isn't in the common case because we may
                // want to do quick prefix scanning. If the program doesn't
                // have a detected prefix, then start states are actually
                // considered common and this case is never reached.
                debug_assert!(self.has_prefix());
                next_si &= !STATE_START;
                prev_si = next_si;
                at = match self.prefix_at(text, at) {
                    None => return Result::NoMatch(text.len()),
                    Some(i) => i,
                };
            } else if next_si >= STATE_UNKNOWN {
                if next_si == STATE_QUIT {
                    return Result::Quit;
                }
                // Finally, this corresponds to the case where the transition
                // entered a state that can never lead to a match or a state
                // that hasn't been computed yet. The latter being the "slow"
                // path.
                let byte = Byte::byte(text[at - 1]);
                // We no longer care about the special bits in the state
                // pointer.
                prev_si &= STATE_MAX;
                // Record where we are. This is used to track progress for
                // determining whether we should quit if we've flushed the
                // cache too much.
                self.at = at;
                next_si = match self.next_state(qcur, qnext, prev_si, byte) {
                    None => return Result::Quit,
                    Some(STATE_DEAD) => return result.set_non_match(at),
                    Some(si) => si,
                };
                debug_assert!(next_si != STATE_UNKNOWN);
                if next_si & STATE_MATCH > 0 {
                    next_si &= !STATE_MATCH;
                    result = Result::Match(at - 1);
                    if self.quit_after_match {
                        return result;
                    }
                    self.last_match_si = next_si;
                }
                prev_si = next_si;
            } else {
                prev_si = next_si;
            }
        }

        // Run the DFA once more on the special EOF sentinel value.
        // We don't care about the special bits in the state pointer any more,
        // so get rid of them.
        prev_si &= STATE_MAX;
        prev_si = match self.next_state(qcur, qnext, prev_si, Byte::eof()) {
            None => return Result::Quit,
            Some(STATE_DEAD) => return result.set_non_match(text.len()),
            Some(si) => si & !STATE_START,
        };
        debug_assert!(prev_si != STATE_UNKNOWN);
        if prev_si & STATE_MATCH > 0 {
            prev_si &= !STATE_MATCH;
            self.last_match_si = prev_si;
            result = Result::Match(text.len());
        }
        result
    }

    /// Executes the DFA on a reverse NFA.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn exec_at_reverse(
        &mut self,
        qcur: &mut SparseSet,
        qnext: &mut SparseSet,
        text: &[u8],
    ) -> Result<usize> {
        // The comments in `exec_at` above mostly apply here too. The main
        // difference is that we move backwards over the input and we look for
        // the longest possible match instead of the leftmost-first match.
        //
        // N.B. The code duplication here is regrettable. Efforts to improve
        // it without sacrificing performance are welcome. ---AG
        debug_assert!(self.prog.is_reverse);
        let mut result = Result::NoMatch(self.at);
        let (mut prev_si, mut next_si) = (self.start, self.start);
        let mut at = self.at;
        while at > 0 {
            while next_si <= STATE_MAX && at > 0 {
                // Argument for safety is in the definition of next_si.
                at -= 1;
                prev_si = unsafe { self.next_si(next_si, text, at) };
                if prev_si > STATE_MAX || at <= 4 {
                    mem::swap(&mut prev_si, &mut next_si);
                    break;
                }
                at -= 1;
                next_si = unsafe { self.next_si(prev_si, text, at) };
                if next_si > STATE_MAX {
                    break;
                }
                at -= 1;
                prev_si = unsafe { self.next_si(next_si, text, at) };
                if prev_si > STATE_MAX {
                    mem::swap(&mut prev_si, &mut next_si);
                    break;
                }
                at -= 1;
                next_si = unsafe { self.next_si(prev_si, text, at) };
            }
            if next_si & STATE_MATCH > 0 {
                next_si &= !STATE_MATCH;
                result = Result::Match(at + 1);
                if self.quit_after_match {
                    return result;
                }
                self.last_match_si = next_si;
                prev_si = next_si;
                let cur = at;
                while (next_si & !STATE_MATCH) == prev_si && at >= 2 {
                    // Argument for safety is in the definition of next_si.
                    at -= 1;
                    next_si = unsafe {
                        self.next_si(next_si & !STATE_MATCH, text, at)
                    };
                }
                if at < cur {
                    result = Result::Match(at + 2);
                }
            } else if next_si >= STATE_UNKNOWN {
                if next_si == STATE_QUIT {
                    return Result::Quit;
                }
                let byte = Byte::byte(text[at]);
                prev_si &= STATE_MAX;
                self.at = at;
                next_si = match self.next_state(qcur, qnext, prev_si, byte) {
                    None => return Result::Quit,
                    Some(STATE_DEAD) => return result.set_non_match(at),
                    Some(si) => si,
                };
                debug_assert!(next_si != STATE_UNKNOWN);
                if next_si & STATE_MATCH > 0 {
                    next_si &= !STATE_MATCH;
                    result = Result::Match(at + 1);
                    if self.quit_after_match {
                        return result;
                    }
                    self.last_match_si = next_si;
                }
                prev_si = next_si;
            } else {
                prev_si = next_si;
            }
        }

        // Run the DFA once more on the special EOF sentinel value.
        prev_si = match self.next_state(qcur, qnext, prev_si, Byte::eof()) {
            None => return Result::Quit,
            Some(STATE_DEAD) => return result.set_non_match(0),
            Some(si) => si,
        };
        debug_assert!(prev_si != STATE_UNKNOWN);
        if prev_si & STATE_MATCH > 0 {
            prev_si &= !STATE_MATCH;
            self.last_match_si = prev_si;
            result = Result::Match(0);
        }
        result
    }

    /// next_si transitions to the next state, where the transition input
    /// corresponds to text[i].
    ///
    /// This elides bounds checks, and is therefore not safe.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    unsafe fn next_si(&self, si: StatePtr, text: &[u8], i: usize) -> StatePtr {
        // What is the argument for safety here?
        // We have three unchecked accesses that could possibly violate safety:
        //
        //   1. The given byte of input (`text[i]`).
        //   2. The class of the byte of input (`classes[text[i]]`).
        //   3. The transition for the class (`trans[si + cls]`).
        //
        // (1) is only safe when calling next_si is guarded by
        // `i < text.len()`.
        //
        // (2) is the easiest case to guarantee since `text[i]` is always a
        // `u8` and `self.prog.byte_classes` always has length `u8::MAX`.
        // (See `ByteClassSet.byte_classes` in `compile.rs`.)
        //
        // (3) is only safe if (1)+(2) are safe. Namely, the transitions
        // of every state are defined to have length equal to the number of
        // byte classes in the program. Therefore, a valid class leads to a
        // valid transition. (All possible transitions are valid lookups, even
        // if it points to a state that hasn't been computed yet.) (3) also
        // relies on `si` being correct, but StatePtrs should only ever be
        // retrieved from the transition table, which ensures they are correct.
        debug_assert!(i < text.len());
        let b = *text.get_unchecked(i);
        debug_assert!((b as usize) < self.prog.byte_classes.len());
        let cls = *self.prog.byte_classes.get_unchecked(b as usize);
        self.cache.trans.next_unchecked(si, cls as usize)
    }

    /// Computes the next state given the current state and the current input
    /// byte (which may be EOF).
    ///
    /// If STATE_DEAD is returned, then there is no valid state transition.
    /// This implies that no permutation of future input can lead to a match
    /// state.
    ///
    /// STATE_UNKNOWN can never be returned.
    fn exec_byte(
        &mut self,
        qcur: &mut SparseSet,
        qnext: &mut SparseSet,
        mut si: StatePtr,
        b: Byte,
    ) -> Option<StatePtr> {
        use crate::prog::Inst::*;

        // Initialize a queue with the current DFA state's NFA states.
        qcur.clear();
        for ip in self.state(si).inst_ptrs() {
            qcur.insert(ip);
        }

        // Before inspecting the current byte, we may need to also inspect
        // whether the position immediately preceding the current byte
        // satisfies the empty assertions found in the current state.
        //
        // We only need to do this step if there are any empty assertions in
        // the current state.
        let is_word_last = self.state(si).flags().is_word();
        let is_word = b.is_ascii_word();
        if self.state(si).flags().has_empty() {
            // Compute the flags immediately preceding the current byte.
            // This means we only care about the "end" or "end line" flags.
            // (The "start" flags are computed immediately following the
            // current byte and are handled below.)
            let mut flags = EmptyFlags::default();
            if b.is_eof() {
                flags.end = true;
                flags.end_line = true;
            } else if b.as_byte().map_or(false, |b| b == b'\n') {
                flags.end_line = true;
            }
            if is_word_last == is_word {
                flags.not_word_boundary = true;
            } else {
                flags.word_boundary = true;
            }
            // Now follow epsilon transitions from every NFA state, but make
            // sure we only follow transitions that satisfy our flags.
            qnext.clear();
            for &ip in &*qcur {
                self.follow_epsilons(usize_to_u32(ip), qnext, flags);
            }
            mem::swap(qcur, qnext);
        }

        // Now we set flags for immediately after the current byte. Since start
        // states are processed separately, and are the only states that can
        // have the StartText flag set, we therefore only need to worry about
        // the StartLine flag here.
        //
        // We do also keep track of whether this DFA state contains a NFA state
        // that is a matching state. This is precisely how we delay the DFA
        // matching by one byte in order to process the special EOF sentinel
        // byte. Namely, if this DFA state containing a matching NFA state,
        // then it is the *next* DFA state that is marked as a match.
        let mut empty_flags = EmptyFlags::default();
        let mut state_flags = StateFlags::default();
        empty_flags.start_line = b.as_byte().map_or(false, |b| b == b'\n');
        if b.is_ascii_word() {
            state_flags.set_word();
        }
        // Now follow all epsilon transitions again, but only after consuming
        // the current byte.
        qnext.clear();
        for &ip in &*qcur {
            match self.prog[ip as usize] {
                // These states never happen in a byte-based program.
                Char(_) | Ranges(_) => unreachable!(),
                // These states are handled when following epsilon transitions.
                Save(_) | Split(_) | EmptyLook(_) => {}
                Match(_) => {
                    state_flags.set_match();
                    if !self.continue_past_first_match() {
                        break;
                    } else if self.prog.matches.len() > 1
                        && !qnext.contains(ip as usize)
                    {
                        // If we are continuing on to find other matches,
                        // then keep a record of the match states we've seen.
                        qnext.insert(ip);
                    }
                }
                Bytes(ref inst) => {
                    if b.as_byte().map_or(false, |b| inst.matches(b)) {
                        self.follow_epsilons(
                            inst.goto as InstPtr,
                            qnext,
                            empty_flags,
                        );
                    }
                }
            }
        }

        let cache = if b.is_eof() && self.prog.matches.len() > 1 {
            // If we're processing the last byte of the input and we're
            // matching a regex set, then make the next state contain the
            // previous states transitions. We do this so that the main
            // matching loop can extract all of the match instructions.
            mem::swap(qcur, qnext);
            // And don't cache this state because it's totally bunk.
            false
        } else {
            true
        };

        // We've now built up the set of NFA states that ought to comprise the
        // next DFA state, so try to find it in the cache, and if it doesn't
        // exist, cache it.
        //
        // N.B. We pass `&mut si` here because the cache may clear itself if
        // it has gotten too full. When that happens, the location of the
        // current state may change.
        let mut next =
            match self.cached_state(qnext, state_flags, Some(&mut si)) {
                None => return None,
                Some(next) => next,
            };
        if (self.start & !STATE_START) == next {
            // Start states can never be match states since all matches are
            // delayed by one byte.
            debug_assert!(!self.state(next).flags().is_match());
            next = self.start_ptr(next);
        }
        if next <= STATE_MAX && self.state(next).flags().is_match() {
            next |= STATE_MATCH;
        }
        debug_assert!(next != STATE_UNKNOWN);
        // And now store our state in the current state's next list.
        if cache {
            let cls = self.byte_class(b);
            self.cache.trans.set_next(si, cls, next);
        }
        Some(next)
    }

    /// Follows the epsilon transitions starting at (and including) `ip`. The
    /// resulting states are inserted into the ordered set `q`.
    ///
    /// Conditional epsilon transitions (i.e., empty width assertions) are only
    /// followed if they are satisfied by the given flags, which should
    /// represent the flags set at the current location in the input.
    ///
    /// If the current location corresponds to the empty string, then only the
    /// end line and/or end text flags may be set. If the current location
    /// corresponds to a real byte in the input, then only the start line
    /// and/or start text flags may be set.
    ///
    /// As an exception to the above, when finding the initial state, any of
    /// the above flags may be set:
    ///
    /// If matching starts at the beginning of the input, then start text and
    /// start line should be set. If the input is empty, then end text and end
    /// line should also be set.
    ///
    /// If matching starts after the beginning of the input, then only start
    /// line should be set if the preceding byte is `\n`. End line should never
    /// be set in this case. (Even if the following byte is a `\n`, it will
    /// be handled in a subsequent DFA state.)
    fn follow_epsilons(
        &mut self,
        ip: InstPtr,
        q: &mut SparseSet,
        flags: EmptyFlags,
    ) {
        use crate::prog::EmptyLook::*;
        use crate::prog::Inst::*;

        // We need to traverse the NFA to follow epsilon transitions, so avoid
        // recursion with an explicit stack.
        self.cache.stack.push(ip);
        while let Some(mut ip) = self.cache.stack.pop() {
            // Try to munch through as many states as possible without
            // pushes/pops to the stack.
            loop {
                // Don't visit states we've already added.
                if q.contains(ip as usize) {
                    break;
                }
                q.insert(ip as usize);
                match self.prog[ip as usize] {
                    Char(_) | Ranges(_) => unreachable!(),
                    Match(_) | Bytes(_) => {
                        break;
                    }
                    EmptyLook(ref inst) => {
                        // Only follow empty assertion states if our flags
                        // satisfy the assertion.
                        match inst.look {
                            StartLine if flags.start_line => {
                                ip = inst.goto as InstPtr;
                            }
                            EndLine if flags.end_line => {
                                ip = inst.goto as InstPtr;
                            }
                            StartText if flags.start => {
                                ip = inst.goto as InstPtr;
                            }
                            EndText if flags.end => {
                                ip = inst.goto as InstPtr;
                            }
                            WordBoundaryAscii if flags.word_boundary => {
                                ip = inst.goto as InstPtr;
                            }
                            NotWordBoundaryAscii
                                if flags.not_word_boundary =>
                            {
                                ip = inst.goto as InstPtr;
                            }
                            WordBoundary if flags.word_boundary => {
                                ip = inst.goto as InstPtr;
                            }
                            NotWordBoundary if flags.not_word_boundary => {
                                ip = inst.goto as InstPtr;
                            }
                            StartLine | EndLine | StartText | EndText
                            | WordBoundaryAscii | NotWordBoundaryAscii
                            | WordBoundary | NotWordBoundary => {
                                break;
                            }
                        }
                    }
                    Save(ref inst) => {
                        ip = inst.goto as InstPtr;
                    }
                    Split(ref inst) => {
                        self.cache.stack.push(inst.goto2 as InstPtr);
                        ip = inst.goto1 as InstPtr;
                    }
                }
            }
        }
    }

    /// Find a previously computed state matching the given set of instructions
    /// and is_match bool.
    ///
    /// The given set of instructions should represent a single state in the
    /// NFA along with all states reachable without consuming any input.
    ///
    /// The is_match bool should be true if and only if the preceding DFA state
    /// contains an NFA matching state. The cached state produced here will
    /// then signify a match. (This enables us to delay a match by one byte,
    /// in order to account for the EOF sentinel byte.)
    ///
    /// If the cache is full, then it is wiped before caching a new state.
    ///
    /// The current state should be specified if it exists, since it will need
    /// to be preserved if the cache clears itself. (Start states are
    /// always saved, so they should not be passed here.) It takes a mutable
    /// pointer to the index because if the cache is cleared, the state's
    /// location may change.
    fn cached_state(
        &mut self,
        q: &SparseSet,
        mut state_flags: StateFlags,
        current_state: Option<&mut StatePtr>,
    ) -> Option<StatePtr> {
        // If we couldn't come up with a non-empty key to represent this state,
        // then it is dead and can never lead to a match.
        //
        // Note that inst_flags represent the set of empty width assertions
        // in q. We use this as an optimization in exec_byte to determine when
        // we should follow epsilon transitions at the empty string preceding
        // the current byte.
        let key = match self.cached_state_key(q, &mut state_flags) {
            None => return Some(STATE_DEAD),
            Some(v) => v,
        };
        // In the cache? Cool. Done.
        if let Some(si) = self.cache.compiled.get_ptr(&key) {
            return Some(si);
        }
        // If the cache has gotten too big, wipe it.
        if self.approximate_size() > self.prog.dfa_size_limit
            && !self.clear_cache_and_save(current_state)
        {
            // Ooops. DFA is giving up.
            return None;
        }
        // Allocate room for our state and add it.
        self.add_state(key)
    }

    /// Produces a key suitable for describing a state in the DFA cache.
    ///
    /// The key invariant here is that equivalent keys are produced for any two
    /// sets of ordered NFA states (and toggling of whether the previous NFA
    /// states contain a match state) that do not discriminate a match for any
    /// input.
    ///
    /// Specifically, q should be an ordered set of NFA states and is_match
    /// should be true if and only if the previous NFA states contained a match
    /// state.
    fn cached_state_key(
        &mut self,
        q: &SparseSet,
        state_flags: &mut StateFlags,
    ) -> Option<State> {
        use crate::prog::Inst::*;

        // We need to build up enough information to recognize pre-built states
        // in the DFA. Generally speaking, this includes every instruction
        // except for those which are purely epsilon transitions, e.g., the
        // Save and Split instructions.
        //
        // Empty width assertions are also epsilon transitions, but since they
        // are conditional, we need to make them part of a state's key in the
        // cache.

        let mut insts =
            mem::replace(&mut self.cache.insts_scratch_space, vec![]);
        insts.clear();
        // Reserve 1 byte for flags.
        insts.push(0);

        let mut prev = 0;
        for &ip in q {
            let ip = usize_to_u32(ip);
            match self.prog[ip as usize] {
                Char(_) | Ranges(_) => unreachable!(),
                Save(_) | Split(_) => {}
                Bytes(_) => push_inst_ptr(&mut insts, &mut prev, ip),
                EmptyLook(_) => {
                    state_flags.set_empty();
                    push_inst_ptr(&mut insts, &mut prev, ip)
                }
                Match(_) => {
                    push_inst_ptr(&mut insts, &mut prev, ip);
                    if !self.continue_past_first_match() {
                        break;
                    }
                }
            }
        }
        // If we couldn't transition to any other instructions and we didn't
        // see a match when expanding NFA states previously, then this is a
        // dead state and no amount of additional input can transition out
        // of this state.
        let opt_state = if insts.len() == 1 && !state_flags.is_match() {
            None
        } else {
            let StateFlags(f) = *state_flags;
            insts[0] = f;
            Some(State { data: Arc::from(&*insts) })
        };
        self.cache.insts_scratch_space = insts;
        opt_state
    }

    /// Clears the cache, but saves and restores current_state if it is not
    /// none.
    ///
    /// The current state must be provided here in case its location in the
    /// cache changes.
    ///
    /// This returns false if the cache is not cleared and the DFA should
    /// give up.
    fn clear_cache_and_save(
        &mut self,
        current_state: Option<&mut StatePtr>,
    ) -> bool {
        if self.cache.compiled.is_empty() {
            // Nothing to clear...
            return true;
        }
        match current_state {
            None => self.clear_cache(),
            Some(si) => {
                let cur = self.state(*si).clone();
                if !self.clear_cache() {
                    return false;
                }
                // The unwrap is OK because we just cleared the cache and
                // therefore know that the next state pointer won't exceed
                // STATE_MAX.
                *si = self.restore_state(cur).unwrap();
                true
            }
        }
    }

    /// Wipes the state cache, but saves and restores the current start state.
    ///
    /// This returns false if the cache is not cleared and the DFA should
    /// give up.
    fn clear_cache(&mut self) -> bool {
        // Bail out of the DFA if we're moving too "slowly."
        // A heuristic from RE2: assume the DFA is too slow if it is processing
        // 10 or fewer bytes per state.
        // Additionally, we permit the cache to be flushed a few times before
        // caling it quits.
        let nstates = self.cache.compiled.len();
        if self.cache.flush_count >= 3
            && self.at >= self.last_cache_flush
            && (self.at - self.last_cache_flush) <= 10 * nstates
        {
            return false;
        }
        // Update statistics tracking cache flushes.
        self.last_cache_flush = self.at;
        self.cache.flush_count += 1;

        // OK, actually flush the cache.
        let start = self.state(self.start & !STATE_START).clone();
        let last_match = if self.last_match_si <= STATE_MAX {
            Some(self.state(self.last_match_si).clone())
        } else {
            None
        };
        self.cache.reset_size();
        self.cache.trans.clear();
        self.cache.compiled.clear();
        for s in &mut self.cache.start_states {
            *s = STATE_UNKNOWN;
        }
        // The unwraps are OK because we just cleared the cache and therefore
        // know that the next state pointer won't exceed STATE_MAX.
        let start_ptr = self.restore_state(start).unwrap();
        self.start = self.start_ptr(start_ptr);
        if let Some(last_match) = last_match {
            self.last_match_si = self.restore_state(last_match).unwrap();
        }
        true
    }

    /// Restores the given state back into the cache, and returns a pointer
    /// to it.
    fn restore_state(&mut self, state: State) -> Option<StatePtr> {
        // If we've already stored this state, just return a pointer to it.
        // None will be the wiser.
        if let Some(si) = self.cache.compiled.get_ptr(&state) {
            return Some(si);
        }
        self.add_state(state)
    }

    /// Returns the next state given the current state si and current byte
    /// b. {qcur,qnext} are used as scratch space for storing ordered NFA
    /// states.
    ///
    /// This tries to fetch the next state from the cache, but if that fails,
    /// it computes the next state, caches it and returns a pointer to it.
    ///
    /// The pointer can be to a real state, or it can be STATE_DEAD.
    /// STATE_UNKNOWN cannot be returned.
    ///
    /// None is returned if a new state could not be allocated (i.e., the DFA
    /// ran out of space and thinks it's running too slowly).
    fn next_state(
        &mut self,
        qcur: &mut SparseSet,
        qnext: &mut SparseSet,
        si: StatePtr,
        b: Byte,
    ) -> Option<StatePtr> {
        if si == STATE_DEAD {
            return Some(STATE_DEAD);
        }
        match self.cache.trans.next(si, self.byte_class(b)) {
            STATE_UNKNOWN => self.exec_byte(qcur, qnext, si, b),
            STATE_QUIT => None,
            nsi => Some(nsi),
        }
    }

    /// Computes and returns the start state, where searching begins at
    /// position `at` in `text`. If the state has already been computed,
    /// then it is pulled from the cache. If the state hasn't been cached,
    /// then it is computed, cached and a pointer to it is returned.
    ///
    /// This may return STATE_DEAD but never STATE_UNKNOWN.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn start_state(
        &mut self,
        q: &mut SparseSet,
        empty_flags: EmptyFlags,
        state_flags: StateFlags,
    ) -> Option<StatePtr> {
        // Compute an index into our cache of start states based on the set
        // of empty/state flags set at the current position in the input. We
        // don't use every flag since not all flags matter. For example, since
        // matches are delayed by one byte, start states can never be match
        // states.
        let flagi = {
            (((empty_flags.start as u8) << 0)
                | ((empty_flags.end as u8) << 1)
                | ((empty_flags.start_line as u8) << 2)
                | ((empty_flags.end_line as u8) << 3)
                | ((empty_flags.word_boundary as u8) << 4)
                | ((empty_flags.not_word_boundary as u8) << 5)
                | ((state_flags.is_word() as u8) << 6)) as usize
        };
        match self.cache.start_states[flagi] {
            STATE_UNKNOWN => {}
            si => return Some(si),
        }
        q.clear();
        let start = usize_to_u32(self.prog.start);
        self.follow_epsilons(start, q, empty_flags);
        // Start states can never be match states because we delay every match
        // by one byte. Given an empty string and an empty match, the match
        // won't actually occur until the DFA processes the special EOF
        // sentinel byte.
        let sp = match self.cached_state(q, state_flags, None) {
            None => return None,
            Some(sp) => self.start_ptr(sp),
        };
        self.cache.start_states[flagi] = sp;
        Some(sp)
    }

    /// Computes the set of starting flags for the given position in text.
    ///
    /// This should only be used when executing the DFA forwards over the
    /// input.
    fn start_flags(&self, text: &[u8], at: usize) -> (EmptyFlags, StateFlags) {
        let mut empty_flags = EmptyFlags::default();
        let mut state_flags = StateFlags::default();
        empty_flags.start = at == 0;
        empty_flags.end = text.is_empty();
        empty_flags.start_line = at == 0 || text[at - 1] == b'\n';
        empty_flags.end_line = text.is_empty();

        let is_word_last = at > 0 && Byte::byte(text[at - 1]).is_ascii_word();
        let is_word = at < text.len() && Byte::byte(text[at]).is_ascii_word();
        if is_word_last {
            state_flags.set_word();
        }
        if is_word == is_word_last {
            empty_flags.not_word_boundary = true;
        } else {
            empty_flags.word_boundary = true;
        }
        (empty_flags, state_flags)
    }

    /// Computes the set of starting flags for the given position in text.
    ///
    /// This should only be used when executing the DFA in reverse over the
    /// input.
    fn start_flags_reverse(
        &self,
        text: &[u8],
        at: usize,
    ) -> (EmptyFlags, StateFlags) {
        let mut empty_flags = EmptyFlags::default();
        let mut state_flags = StateFlags::default();
        empty_flags.start = at == text.len();
        empty_flags.end = text.is_empty();
        empty_flags.start_line = at == text.len() || text[at] == b'\n';
        empty_flags.end_line = text.is_empty();

        let is_word_last =
            at < text.len() && Byte::byte(text[at]).is_ascii_word();
        let is_word = at > 0 && Byte::byte(text[at - 1]).is_ascii_word();
        if is_word_last {
            state_flags.set_word();
        }
        if is_word == is_word_last {
            empty_flags.not_word_boundary = true;
        } else {
            empty_flags.word_boundary = true;
        }
        (empty_flags, state_flags)
    }

    /// Returns a reference to a State given a pointer to it.
    fn state(&self, si: StatePtr) -> &State {
        self.cache.compiled.get_state(si).unwrap()
    }

    /// Adds the given state to the DFA.
    ///
    /// This allocates room for transitions out of this state in
    /// self.cache.trans. The transitions can be set with the returned
    /// StatePtr.
    ///
    /// If None is returned, then the state limit was reached and the DFA
    /// should quit.
    fn add_state(&mut self, state: State) -> Option<StatePtr> {
        // This will fail if the next state pointer exceeds STATE_PTR. In
        // practice, the cache limit will prevent us from ever getting here,
        // but maybe callers will set the cache size to something ridiculous...
        let si = match self.cache.trans.add() {
            None => return None,
            Some(si) => si,
        };
        // If the program has a Unicode word boundary, then set any transitions
        // for non-ASCII bytes to STATE_QUIT. If the DFA stumbles over such a
        // transition, then it will quit and an alternative matching engine
        // will take over.
        if self.prog.has_unicode_word_boundary {
            for b in 128..256 {
                let cls = self.byte_class(Byte::byte(b as u8));
                self.cache.trans.set_next(si, cls, STATE_QUIT);
            }
        }
        // Finally, put our actual state on to our heap of states and index it
        // so we can find it later.
        self.cache.size += self.cache.trans.state_heap_size()
            + state.data.len()
            + (2 * mem::size_of::<State>())
            + mem::size_of::<StatePtr>();
        self.cache.compiled.insert(state, si);
        // Transition table and set of states and map should all be in sync.
        debug_assert!(
            self.cache.compiled.len() == self.cache.trans.num_states()
        );
        Some(si)
    }

    /// Quickly finds the next occurrence of any literal prefixes in the regex.
    /// If there are no literal prefixes, then the current position is
    /// returned. If there are literal prefixes and one could not be found,
    /// then None is returned.
    ///
    /// This should only be called when the DFA is in a start state.
    fn prefix_at(&self, text: &[u8], at: usize) -> Option<usize> {
        self.prog.prefixes.find(&text[at..]).map(|(s, _)| at + s)
    }

    /// Returns the number of byte classes required to discriminate transitions
    /// in each state.
    ///
    /// invariant: num_byte_classes() == len(State.next)
    fn num_byte_classes(&self) -> usize {
        // We add 1 to account for the special EOF byte.
        (self.prog.byte_classes[255] as usize + 1) + 1
    }

    /// Given an input byte or the special EOF sentinel, return its
    /// corresponding byte class.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn byte_class(&self, b: Byte) -> usize {
        match b.as_byte() {
            None => self.num_byte_classes() - 1,
            Some(b) => self.u8_class(b),
        }
    }

    /// Like byte_class, but explicitly for u8s.
    #[cfg_attr(feature = "perf-inline", inline(always))]
    fn u8_class(&self, b: u8) -> usize {
        self.prog.byte_classes[b as usize] as usize
    }

    /// Returns true if the DFA should continue searching past the first match.
    ///
    /// Leftmost first semantics in the DFA are preserved by not following NFA
    /// transitions after the first match is seen.
    ///
    /// On occasion, we want to avoid leftmost first semantics to find either
    /// the longest match (for reverse search) or all possible matches (for
    /// regex sets).
    fn continue_past_first_match(&self) -> bool {
        self.prog.is_reverse || self.prog.matches.len() > 1
    }

    /// Returns true if there is a prefix we can quickly search for.
    fn has_prefix(&self) -> bool {
        !self.prog.is_reverse
            && !self.prog.prefixes.is_empty()
            && !self.prog.is_anchored_start
    }

    /// Sets the STATE_START bit in the given state pointer if and only if
    /// we have a prefix to scan for.
    ///
    /// If there's no prefix, then it's a waste to treat the start state
    /// specially.
    fn start_ptr(&self, si: StatePtr) -> StatePtr {
        if self.has_prefix() {
            si | STATE_START
        } else {
            si
        }
    }

    /// Approximate size returns the approximate heap space currently used by
    /// the DFA. It is used to determine whether the DFA's state cache needs to
    /// be wiped. Namely, it is possible that for certain regexes on certain
    /// inputs, a new state could be created for every byte of input. (This is
    /// bad for memory use, so we bound it with a cache.)
    fn approximate_size(&self) -> usize {
        self.cache.size + self.prog.approximate_size()
    }
}

/// An abstraction for representing a map of states. The map supports two
/// different ways of state lookup. One is fast constant time access via a
/// state pointer. The other is a hashmap lookup based on the DFA's
/// constituent NFA states.
///
/// A DFA state internally uses an Arc such that we only need to store the
/// set of NFA states on the heap once, even though we support looking up
/// states by two different means. A more natural way to express this might
/// use raw pointers, but an Arc is safe and effectively achieves the same
/// thing.
#[derive(Debug)]
struct StateMap {
    /// The keys are not actually static but rely on always pointing to a
    /// buffer in `states` which will never be moved except when clearing
    /// the map or on drop, in which case the keys of this map will be
    /// removed before
    map: HashMap<State, StatePtr>,
    /// Our set of states. Note that `StatePtr / num_byte_classes` indexes
    /// this Vec rather than just a `StatePtr`.
    states: Vec<State>,
    /// The number of byte classes in the DFA. Used to index `states`.
    num_byte_classes: usize,
}

impl StateMap {
    fn new(num_byte_classes: usize) -> StateMap {
        StateMap {
            map: HashMap::new(),
            states: vec![],
            num_byte_classes: num_byte_classes,
        }
    }

    fn len(&self) -> usize {
        self.states.len()
    }

    fn is_empty(&self) -> bool {
        self.states.is_empty()
    }

    fn get_ptr(&self, state: &State) -> Option<StatePtr> {
        self.map.get(state).cloned()
    }

    fn get_state(&self, si: StatePtr) -> Option<&State> {
        self.states.get(si as usize / self.num_byte_classes)
    }

    fn insert(&mut self, state: State, si: StatePtr) {
        self.map.insert(state.clone(), si);
        self.states.push(state);
    }

    fn clear(&mut self) {
        self.map.clear();
        self.states.clear();
    }
}

impl Transitions {
    /// Create a new transition table.
    ///
    /// The number of byte classes corresponds to the stride. Every state will
    /// have `num_byte_classes` slots for transitions.
    fn new(num_byte_classes: usize) -> Transitions {
        Transitions { table: vec![], num_byte_classes: num_byte_classes }
    }

    /// Returns the total number of states currently in this table.
    fn num_states(&self) -> usize {
        self.table.len() / self.num_byte_classes
    }

    /// Allocates room for one additional state and returns a pointer to it.
    ///
    /// If there's no more room, None is returned.
    fn add(&mut self) -> Option<StatePtr> {
        let si = self.table.len();
        if si > STATE_MAX as usize {
            return None;
        }
        self.table.extend(repeat(STATE_UNKNOWN).take(self.num_byte_classes));
        Some(usize_to_u32(si))
    }

    /// Clears the table of all states.
    fn clear(&mut self) {
        self.table.clear();
    }

    /// Sets the transition from (si, cls) to next.
    fn set_next(&mut self, si: StatePtr, cls: usize, next: StatePtr) {
        self.table[si as usize + cls] = next;
    }

    /// Returns the transition corresponding to (si, cls).
    fn next(&self, si: StatePtr, cls: usize) -> StatePtr {
        self.table[si as usize + cls]
    }

    /// The heap size, in bytes, of a single state in the transition table.
    fn state_heap_size(&self) -> usize {
        self.num_byte_classes * mem::size_of::<StatePtr>()
    }

    /// Like `next`, but uses unchecked access and is therefore not safe.
    unsafe fn next_unchecked(&self, si: StatePtr, cls: usize) -> StatePtr {
        debug_assert!((si as usize) < self.table.len());
        debug_assert!(cls < self.num_byte_classes);
        *self.table.get_unchecked(si as usize + cls)
    }
}

impl StateFlags {
    fn is_match(&self) -> bool {
        self.0 & 0b0000000_1 > 0
    }

    fn set_match(&mut self) {
        self.0 |= 0b0000000_1;
    }

    fn is_word(&self) -> bool {
        self.0 & 0b000000_1_0 > 0
    }

    fn set_word(&mut self) {
        self.0 |= 0b000000_1_0;
    }

    fn has_empty(&self) -> bool {
        self.0 & 0b00000_1_00 > 0
    }

    fn set_empty(&mut self) {
        self.0 |= 0b00000_1_00;
    }
}

impl Byte {
    fn byte(b: u8) -> Self {
        Byte(b as u16)
    }
    fn eof() -> Self {
        Byte(256)
    }
    fn is_eof(&self) -> bool {
        self.0 == 256
    }

    fn is_ascii_word(&self) -> bool {
        let b = match self.as_byte() {
            None => return false,
            Some(b) => b,
        };
        match b {
            b'A'..=b'Z' | b'a'..=b'z' | b'0'..=b'9' | b'_' => true,
            _ => false,
        }
    }

    fn as_byte(&self) -> Option<u8> {
        if self.is_eof() {
            None
        } else {
            Some(self.0 as u8)
        }
    }
}

impl fmt::Debug for State {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        let ips: Vec<usize> = self.inst_ptrs().collect();
        f.debug_struct("State")
            .field("flags", &self.flags())
            .field("insts", &ips)
            .finish()
    }
}

impl fmt::Debug for Transitions {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        let mut fmtd = f.debug_map();
        for si in 0..self.num_states() {
            let s = si * self.num_byte_classes;
            let e = s + self.num_byte_classes;
            fmtd.entry(&si.to_string(), &TransitionsRow(&self.table[s..e]));
        }
        fmtd.finish()
    }
}

struct TransitionsRow<'a>(&'a [StatePtr]);

impl<'a> fmt::Debug for TransitionsRow<'a> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        let mut fmtd = f.debug_map();
        for (b, si) in self.0.iter().enumerate() {
            match *si {
                STATE_UNKNOWN => {}
                STATE_DEAD => {
                    fmtd.entry(&vb(b as usize), &"DEAD");
                }
                si => {
                    fmtd.entry(&vb(b as usize), &si.to_string());
                }
            }
        }
        fmtd.finish()
    }
}

impl fmt::Debug for StateFlags {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("StateFlags")
            .field("is_match", &self.is_match())
            .field("is_word", &self.is_word())
            .field("has_empty", &self.has_empty())
            .finish()
    }
}

/// Helper function for formatting a byte as a nice-to-read escaped string.
fn vb(b: usize) -> String {
    use std::ascii::escape_default;

    if b > ::std::u8::MAX as usize {
        "EOF".to_owned()
    } else {
        let escaped = escape_default(b as u8).collect::<Vec<u8>>();
        String::from_utf8_lossy(&escaped).into_owned()
    }
}

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

#[allow(dead_code)] // useful for debugging
fn show_state_ptr(si: StatePtr) -> String {
    let mut s = format!("{:?}", si & STATE_MAX);
    if si == STATE_UNKNOWN {
        s = format!("{} (unknown)", s);
    }
    if si == STATE_DEAD {
        s = format!("{} (dead)", s);
    }
    if si == STATE_QUIT {
        s = format!("{} (quit)", s);
    }
    if si & STATE_START > 0 {
        s = format!("{} (start)", s);
    }
    if si & STATE_MATCH > 0 {
        s = format!("{} (match)", s);
    }
    s
}

/// https://developers.google.com/protocol-buffers/docs/encoding#varints
fn write_vari32(data: &mut Vec<u8>, n: i32) {
    let mut un = (n as u32) << 1;
    if n < 0 {
        un = !un;
    }
    write_varu32(data, un)
}

/// https://developers.google.com/protocol-buffers/docs/encoding#varints
fn read_vari32(data: &[u8]) -> (i32, usize) {
    let (un, i) = read_varu32(data);
    let mut n = (un >> 1) as i32;
    if un & 1 != 0 {
        n = !n;
    }
    (n, i)
}

/// https://developers.google.com/protocol-buffers/docs/encoding#varints
fn write_varu32(data: &mut Vec<u8>, mut n: u32) {
    while n >= 0b1000_0000 {
        data.push((n as u8) | 0b1000_0000);
        n >>= 7;
    }
    data.push(n as u8);
}

/// https://developers.google.com/protocol-buffers/docs/encoding#varints
fn read_varu32(data: &[u8]) -> (u32, usize) {
    let mut n: u32 = 0;
    let mut shift: u32 = 0;
    for (i, &b) in data.iter().enumerate() {
        if b < 0b1000_0000 {
            return (n | ((b as u32) << shift), i + 1);
        }
        n |= ((b as u32) & 0b0111_1111) << shift;
        shift += 7;
    }
    (0, 0)
}

#[cfg(test)]
mod tests {

    use super::{
        push_inst_ptr, read_vari32, read_varu32, write_vari32, write_varu32,
        State, StateFlags,
    };
    use quickcheck::{quickcheck, Gen, QuickCheck};
    use std::sync::Arc;

    #[test]
    fn prop_state_encode_decode() {
        fn p(mut ips: Vec<u32>, flags: u8) -> bool {
            // It looks like our encoding scheme can't handle instruction
            // pointers at or above 2**31. We should fix that, but it seems
            // unlikely to occur in real code due to the amount of memory
            // required for such a state machine. So for now, we just clamp
            // our test data.
            for ip in &mut ips {
                if *ip >= 1 << 31 {
                    *ip = (1 << 31) - 1;
                }
            }
            let mut data = vec![flags];
            let mut prev = 0;
            for &ip in ips.iter() {
                push_inst_ptr(&mut data, &mut prev, ip);
            }
            let state = State { data: Arc::from(&data[..]) };

            let expected: Vec<usize> =
                ips.into_iter().map(|ip| ip as usize).collect();
            let got: Vec<usize> = state.inst_ptrs().collect();
            expected == got && state.flags() == StateFlags(flags)
        }
        QuickCheck::new()
            .gen(Gen::new(10_000))
            .quickcheck(p as fn(Vec<u32>, u8) -> bool);
    }

    #[test]
    fn prop_read_write_u32() {
        fn p(n: u32) -> bool {
            let mut buf = vec![];
            write_varu32(&mut buf, n);
            let (got, nread) = read_varu32(&buf);
            nread == buf.len() && got == n
        }
        quickcheck(p as fn(u32) -> bool);
    }

    #[test]
    fn prop_read_write_i32() {
        fn p(n: i32) -> bool {
            let mut buf = vec![];
            write_vari32(&mut buf, n);
            let (got, nread) = read_vari32(&buf);
            nread == buf.len() && got == n
        }
        quickcheck(p as fn(i32) -> bool);
    }
}