This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Keep parse pointer on full character boundary
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
afda64e8 84EXTERN_C const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
f7e03a10 89#include "dquote_inline.h"
b992490d 90#include "invlist_inline.h"
1b0f46bf 91#include "unicode_constants.h"
04e98a4d 92
538e84ed
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
95#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 97#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 98#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 99
a687059c
LW
100#ifndef STATIC
101#define STATIC static
102#endif
103
dea37815
KW
104#ifndef MIN
105#define MIN(a,b) ((a) < (b) ? (a) : (b))
106#endif
b1603ef8 107
3f910e62
YO
108/* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
111struct scan_frame;
112typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
118
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
122} scan_frame;
123
f4ae5a27
KW
124/* Certain characters are output as a sequence with the first being a
125 * backslash. */
126#define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
128
129
09b2b2e6 130struct RExC_state_t {
514a91f1
DM
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 133 char *precomp; /* uncompiled string. */
288b8c02 134 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 135 regexp *rx; /* perl core regexp structure */
538e84ed
KW
136 regexp_internal *rxi; /* internal data for regexp object
137 pprivate field */
fac92740 138 char *start; /* Start of input for compile */
830247a4
IZ
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
ea3daa5d 141 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 142 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
143 regnode *emit_bound; /* First regnode outside of the
144 allocated space */
f7c7e32a
DM
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
9a81a976
KW
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
150 scratch in pass1 */
830247a4
IZ
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
153 U32 seen;
ea3daa5d 154 SSize_t size; /* Code size. */
538e84ed
KW
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
157 pattern)*/
158 I32 nestroot; /* root parens we are in - used by
159 accept */
830247a4
IZ
160 I32 extralen;
161 I32 seen_zerolen;
40d049e4
YO
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
02daf0ab
YO
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
e40e74fe
KW
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
171 utf8 */
81714fb9 172 HV *paren_names; /* Paren names */
538e84ed 173
40d049e4
YO
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
4286711a 176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 177 through */
09a65838 178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 179 I32 in_lookbehind;
4624b182 180 I32 contains_locale;
cfafade5 181 I32 contains_i;
bb3f3ed2 182 I32 override_recoding;
b6d67071
KW
183#ifdef EBCDIC
184 I32 recode_x_to_native;
185#endif
9d53c457 186 I32 in_multi_char_class;
3d2bd50a 187 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 188 within pattern */
b1603ef8
DM
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
ee273784 191 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
192 scan_frame *frame_head;
193 scan_frame *frame_last;
194 U32 frame_count;
67cdf558 195 U32 strict;
dc6d7f5c 196#ifdef ADD_TO_REGEXEC
830247a4
IZ
197 char *starttry; /* -Dr: where regtry was called. */
198#define RExC_starttry (pRExC_state->starttry)
199#endif
d24ca0c5 200 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 201#ifdef DEBUGGING
be8e71aa 202 const char *lastparse;
3dab1dad 203 I32 lastnum;
1f1031fe 204 AV *paren_name_list; /* idx -> name */
d9a72fcc 205 U32 study_chunk_recursed_count;
c9f0d54c
YO
206 SV *mysv1;
207 SV *mysv2;
3dab1dad
YO
208#define RExC_lastparse (pRExC_state->lastparse)
209#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 210#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 211#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
212#define RExC_mysv (pRExC_state->mysv1)
213#define RExC_mysv1 (pRExC_state->mysv1)
214#define RExC_mysv2 (pRExC_state->mysv2)
215
3dab1dad 216#endif
512c0f5a 217 bool seen_unfolded_sharp_s;
09b2b2e6 218};
830247a4 219
e2509266 220#define RExC_flags (pRExC_state->flags)
514a91f1 221#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 222#define RExC_precomp (pRExC_state->precomp)
288b8c02 223#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 224#define RExC_rx (pRExC_state->rx)
f8fc2ecf 225#define RExC_rxi (pRExC_state->rxi)
fac92740 226#define RExC_start (pRExC_state->start)
830247a4
IZ
227#define RExC_end (pRExC_state->end)
228#define RExC_parse (pRExC_state->parse)
229#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
230
231/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
232 * EXACTF node, hence was parsed under /di rules. If later in the parse,
233 * something forces the pattern into using /ui rules, the sharp s should be
234 * folded into the sequence 'ss', which takes up more space than previously
235 * calculated. This means that the sizing pass needs to be restarted. (The
236 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
237 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
238 * so there is no need to resize [perl #125990]. */
239#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
240
7122b237 241#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
242#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
243 others */
7122b237 244#endif
830247a4 245#define RExC_emit (pRExC_state->emit)
f7c7e32a 246#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 247#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 248#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
249#define RExC_sawback (pRExC_state->sawback)
250#define RExC_seen (pRExC_state->seen)
251#define RExC_size (pRExC_state->size)
ee273784 252#define RExC_maxlen (pRExC_state->maxlen)
830247a4 253#define RExC_npar (pRExC_state->npar)
e2e6a0f1 254#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
255#define RExC_extralen (pRExC_state->extralen)
256#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 257#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 258#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 259#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
260#define RExC_open_parens (pRExC_state->open_parens)
261#define RExC_close_parens (pRExC_state->close_parens)
262#define RExC_opend (pRExC_state->opend)
81714fb9 263#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
264#define RExC_recurse (pRExC_state->recurse)
265#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 266#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
267#define RExC_study_chunk_recursed_bytes \
268 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 269#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 270#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 271#define RExC_contains_i (pRExC_state->contains_i)
9d53c457 272#define RExC_override_recoding (pRExC_state->override_recoding)
b6d67071
KW
273#ifdef EBCDIC
274# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275#endif
9d53c457 276#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
277#define RExC_frame_head (pRExC_state->frame_head)
278#define RExC_frame_last (pRExC_state->frame_last)
279#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 280#define RExC_strict (pRExC_state->strict)
830247a4 281
99807a43
HS
282/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
283 * a flag to disable back-off on the fixed/floating substrings - if it's
284 * a high complexity pattern we assume the benefit of avoiding a full match
285 * is worth the cost of checking for the substrings even if they rarely help.
286 */
287#define RExC_naughty (pRExC_state->naughty)
288#define TOO_NAUGHTY (10)
289#define MARK_NAUGHTY(add) \
290 if (RExC_naughty < TOO_NAUGHTY) \
291 RExC_naughty += (add)
292#define MARK_NAUGHTY_EXP(exp, add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 295
a687059c
LW
296#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
297#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 298 ((*s) == '{' && regcurly(s)))
a687059c
LW
299
300/*
301 * Flags to be passed up and down.
302 */
a687059c 303#define WORST 0 /* Worst case. */
a3b492c3 304#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 305
e64f369d 306/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
307 * character. (There needs to be a case: in the switch statement in regexec.c
308 * for any node marked SIMPLE.) Note that this is not the same thing as
309 * REGNODE_SIMPLE */
fda99bee 310#define SIMPLE 0x02
e64f369d 311#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
312#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
313#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
314#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
315#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
316 calcuate sizes as UTF-8 */
a687059c 317
3dab1dad
YO
318#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
319
07be1b83
YO
320/* whether trie related optimizations are enabled */
321#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
322#define TRIE_STUDY_OPT
786e8c11 323#define FULL_TRIE_STUDY
07be1b83
YO
324#define TRIE_STCLASS
325#endif
1de06328
YO
326
327
40d049e4
YO
328
329#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
330#define PBITVAL(paren) (1 << ((paren) & 7))
331#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
332#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
333#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
334
82a6ada4 335#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 336 if (!UTF) { \
82a6ada4
KW
337 assert(PASS1); \
338 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
339 return NULL; \
340 } \
82a6ada4 341 } STMT_END
40d049e4 342
512c0f5a
KW
343/* Change from /d into /u rules, and restart the parse if we've already seen
344 * something whose size would increase as a result, by setting *flagp and
345 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
346 * we've change to /u during the parse. */
347#define REQUIRE_UNI_RULES(flagp, restart_retval) \
348 STMT_START { \
349 if (DEPENDS_SEMANTICS) { \
350 assert(PASS1); \
351 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
352 RExC_uni_semantics = 1; \
353 if (RExC_seen_unfolded_sharp_s) { \
354 *flagp |= RESTART_PASS1; \
355 return restart_retval; \
356 } \
357 } \
358 } STMT_END
359
f19b1a63
KW
360/* This converts the named class defined in regcomp.h to its equivalent class
361 * number defined in handy.h. */
362#define namedclass_to_classnum(class) ((int) ((class) / 2))
363#define classnum_to_namedclass(classnum) ((classnum) * 2)
364
de92f5e6
KW
365#define _invlist_union_complement_2nd(a, b, output) \
366 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
367#define _invlist_intersection_complement_2nd(a, b, output) \
368 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
369
1de06328
YO
370/* About scan_data_t.
371
372 During optimisation we recurse through the regexp program performing
373 various inplace (keyhole style) optimisations. In addition study_chunk
374 and scan_commit populate this data structure with information about
538e84ed 375 what strings MUST appear in the pattern. We look for the longest
3b753521 376 string that must appear at a fixed location, and we look for the
1de06328
YO
377 longest string that may appear at a floating location. So for instance
378 in the pattern:
538e84ed 379
1de06328 380 /FOO[xX]A.*B[xX]BAR/
538e84ed 381
1de06328
YO
382 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
383 strings (because they follow a .* construct). study_chunk will identify
384 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 385
1de06328 386 The strings can be composites, for instance
538e84ed 387
1de06328 388 /(f)(o)(o)/
538e84ed 389
1de06328 390 will result in a composite fixed substring 'foo'.
538e84ed 391
1de06328 392 For each string some basic information is maintained:
538e84ed 393
1de06328
YO
394 - offset or min_offset
395 This is the position the string must appear at, or not before.
396 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
397 characters must match before the string we are searching for.
398 Likewise when combined with minlenp and the length of the string it
538e84ed 399 tells us how many characters must appear after the string we have
1de06328 400 found.
538e84ed 401
1de06328
YO
402 - max_offset
403 Only used for floating strings. This is the rightmost point that
ea3daa5d 404 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 405 string can occur infinitely far to the right.
538e84ed 406
1de06328 407 - minlenp
2d608413
KW
408 A pointer to the minimum number of characters of the pattern that the
409 string was found inside. This is important as in the case of positive
538e84ed 410 lookahead or positive lookbehind we can have multiple patterns
1de06328 411 involved. Consider
538e84ed 412
1de06328 413 /(?=FOO).*F/
538e84ed 414
1de06328
YO
415 The minimum length of the pattern overall is 3, the minimum length
416 of the lookahead part is 3, but the minimum length of the part that
538e84ed 417 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 418 minimum length for the F is 1. This is important as the minimum length
538e84ed 419 is used to determine offsets in front of and behind the string being
1de06328 420 looked for. Since strings can be composites this is the length of the
486ec47a 421 pattern at the time it was committed with a scan_commit. Note that
1de06328 422 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 423 are not known until the full pattern has been compiled, thus the
1de06328 424 pointer to the value.
538e84ed 425
1de06328 426 - lookbehind
538e84ed 427
1de06328 428 In the case of lookbehind the string being searched for can be
538e84ed 429 offset past the start point of the final matching string.
1de06328
YO
430 If this value was just blithely removed from the min_offset it would
431 invalidate some of the calculations for how many chars must match
432 before or after (as they are derived from min_offset and minlen and
538e84ed 433 the length of the string being searched for).
1de06328
YO
434 When the final pattern is compiled and the data is moved from the
435 scan_data_t structure into the regexp structure the information
538e84ed
KW
436 about lookbehind is factored in, with the information that would
437 have been lost precalculated in the end_shift field for the
1de06328
YO
438 associated string.
439
440 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 441 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
442
443*/
2c2d71f5
JH
444
445typedef struct scan_data_t {
1de06328
YO
446 /*I32 len_min; unused */
447 /*I32 len_delta; unused */
49f55535 448 SSize_t pos_min;
ea3daa5d 449 SSize_t pos_delta;
2c2d71f5 450 SV *last_found;
ea3daa5d 451 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 452 SSize_t last_start_min;
ea3daa5d 453 SSize_t last_start_max;
1de06328
YO
454 SV **longest; /* Either &l_fixed, or &l_float. */
455 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 456 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 457 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
458 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
459 SV *longest_float; /* longest floating string found in pattern */
49f55535 460 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
461 SSize_t offset_float_max; /* latest point in string it can appear */
462 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 463 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
464 I32 flags;
465 I32 whilem_c;
49f55535 466 SSize_t *last_closep;
b8f7bb16 467 regnode_ssc *start_class;
2c2d71f5
JH
468} scan_data_t;
469
a687059c 470/*
e50aee73 471 * Forward declarations for pregcomp()'s friends.
a687059c 472 */
a0d0e21e 473
27da23d5 474static const scan_data_t zero_scan_data =
1de06328 475 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
476
477#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
478#define SF_BEFORE_SEOL 0x0001
479#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
480#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
481#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
482
44e3dfd2
BF
483#define SF_FIX_SHIFT_EOL (+2)
484#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
485
486#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
487#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
488
489#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
490#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
491#define SF_IS_INF 0x0040
492#define SF_HAS_PAR 0x0080
493#define SF_IN_PAR 0x0100
494#define SF_HAS_EVAL 0x0200
495#define SCF_DO_SUBSTR 0x0400
653099ff
GS
496#define SCF_DO_STCLASS_AND 0x0800
497#define SCF_DO_STCLASS_OR 0x1000
498#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 499#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 500
786e8c11 501#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 502#define SCF_SEEN_ACCEPT 0x8000
688e0391 503#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
504#define SCF_IN_DEFINE 0x20000
505
506
507
07be1b83 508
43fead97 509#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
510
511/* The enums for all these are ordered so things work out correctly */
a62b1201 512#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
513#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
514 == REGEX_DEPENDS_CHARSET)
00b27cfc 515#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
516#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
517 >= REGEX_UNICODE_CHARSET)
518#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
519 == REGEX_ASCII_RESTRICTED_CHARSET)
520#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
521 >= REGEX_ASCII_RESTRICTED_CHARSET)
522#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
523 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 524
43fead97 525#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 526
f2c2a6ab
KW
527/* For programs that want to be strictly Unicode compatible by dying if any
528 * attempt is made to match a non-Unicode code point against a Unicode
529 * property. */
530#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
531
93733859 532#define OOB_NAMEDCLASS -1
b8c5462f 533
8e661ac5
KW
534/* There is no code point that is out-of-bounds, so this is problematic. But
535 * its only current use is to initialize a variable that is always set before
536 * looked at. */
537#define OOB_UNICODE 0xDEADBEEF
538
a0ed51b3
LW
539#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
540#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
541
8615cb43 542
b45f050a
JF
543/* length of regex to show in messages that don't mark a position within */
544#define RegexLengthToShowInErrorMessages 127
545
546/*
547 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
548 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
549 * op/pragma/warn/regcomp.
550 */
7253e4e3
RK
551#define MARKER1 "<-- HERE" /* marker as it appears in the description */
552#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 553
538e84ed
KW
554#define REPORT_LOCATION " in regex; marked by " MARKER1 \
555 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 556
c1d900c3
BF
557#define REPORT_LOCATION_ARGS(offset) \
558 UTF8fARG(UTF, offset, RExC_precomp), \
559 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
560
8a6d8ec6
HS
561/* Used to point after bad bytes for an error message, but avoid skipping
562 * past a nul byte. */
563#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
564
b45f050a
JF
565/*
566 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
567 * arg. Show regex, up to a maximum length. If it's too long, chop and add
568 * "...".
569 */
58e23c8d 570#define _FAIL(code) STMT_START { \
bfed75c6 571 const char *ellipses = ""; \
ccb2c380
MP
572 IV len = RExC_end - RExC_precomp; \
573 \
574 if (!SIZE_ONLY) \
a5e7bc51 575 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
576 if (len > RegexLengthToShowInErrorMessages) { \
577 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
578 len = RegexLengthToShowInErrorMessages - 10; \
579 ellipses = "..."; \
580 } \
58e23c8d 581 code; \
ccb2c380 582} STMT_END
8615cb43 583
58e23c8d 584#define FAIL(msg) _FAIL( \
c1d900c3
BF
585 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
586 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
587
588#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
589 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
590 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 591
b45f050a 592/*
b45f050a
JF
593 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
594 */
ccb2c380 595#define Simple_vFAIL(m) STMT_START { \
8fb0127d
YO
596 const IV offset = \
597 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
ccb2c380 598 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
c1d900c3 599 m, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 600} STMT_END
b45f050a
JF
601
602/*
603 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
604 */
ccb2c380
MP
605#define vFAIL(m) STMT_START { \
606 if (!SIZE_ONLY) \
a5e7bc51 607 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
608 Simple_vFAIL(m); \
609} STMT_END
b45f050a
JF
610
611/*
612 * Like Simple_vFAIL(), but accepts two arguments.
613 */
ccb2c380 614#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 615 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
616 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
617 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 618} STMT_END
b45f050a
JF
619
620/*
621 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
622 */
ccb2c380
MP
623#define vFAIL2(m,a1) STMT_START { \
624 if (!SIZE_ONLY) \
a5e7bc51 625 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
626 Simple_vFAIL2(m, a1); \
627} STMT_END
b45f050a
JF
628
629
630/*
631 * Like Simple_vFAIL(), but accepts three arguments.
632 */
ccb2c380 633#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 634 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
635 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
636 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 637} STMT_END
b45f050a
JF
638
639/*
640 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
641 */
ccb2c380
MP
642#define vFAIL3(m,a1,a2) STMT_START { \
643 if (!SIZE_ONLY) \
a5e7bc51 644 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
645 Simple_vFAIL3(m, a1, a2); \
646} STMT_END
b45f050a
JF
647
648/*
649 * Like Simple_vFAIL(), but accepts four arguments.
650 */
ccb2c380 651#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 652 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
653 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
654 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 655} STMT_END
b45f050a 656
95db3ffa
KW
657#define vFAIL4(m,a1,a2,a3) STMT_START { \
658 if (!SIZE_ONLY) \
659 SAVEFREESV(RExC_rx_sv); \
660 Simple_vFAIL4(m, a1, a2, a3); \
661} STMT_END
662
946095af 663/* A specialized version of vFAIL2 that works with UTF8f */
631041de 664#define vFAIL2utf8f(m, a1) STMT_START { \
ef3f731d
BF
665 const IV offset = RExC_parse - RExC_precomp; \
666 if (!SIZE_ONLY) \
667 SAVEFREESV(RExC_rx_sv); \
668 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
669 REPORT_LOCATION_ARGS(offset)); \
946095af
BF
670} STMT_END
671
3ba22297
KW
672#define vFAIL3utf8f(m, a1, a2) STMT_START { \
673 const IV offset = RExC_parse - RExC_precomp; \
674 if (!SIZE_ONLY) \
675 SAVEFREESV(RExC_rx_sv); \
676 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
677 REPORT_LOCATION_ARGS(offset)); \
678} STMT_END
679
499333dc
KW
680/* These have asserts in them because of [perl #122671] Many warnings in
681 * regcomp.c can occur twice. If they get output in pass1 and later in that
682 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
683 * would get output again. So they should be output in pass2, and these
684 * asserts make sure new warnings follow that paradigm. */
946095af 685
5e0a247b
KW
686/* m is not necessarily a "literal string", in this macro */
687#define reg_warn_non_literal_string(loc, m) STMT_START { \
688 const IV offset = loc - RExC_precomp; \
499333dc 689 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
c1d900c3 690 m, REPORT_LOCATION_ARGS(offset)); \
5e0a247b
KW
691} STMT_END
692
668c081a 693#define ckWARNreg(loc,m) STMT_START { \
a28509cc 694 const IV offset = loc - RExC_precomp; \
499333dc 695 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 696 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
697} STMT_END
698
b927b7e9
KW
699#define vWARN(loc, m) STMT_START { \
700 const IV offset = loc - RExC_precomp; \
701 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
702 REPORT_LOCATION_ARGS(offset)); \
703} STMT_END
704
0d6106aa
KW
705#define vWARN_dep(loc, m) STMT_START { \
706 const IV offset = loc - RExC_precomp; \
499333dc 707 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
c1d900c3 708 REPORT_LOCATION_ARGS(offset)); \
0d6106aa
KW
709} STMT_END
710
147508a2
KW
711#define ckWARNdep(loc,m) STMT_START { \
712 const IV offset = loc - RExC_precomp; \
499333dc 713 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
147508a2 714 m REPORT_LOCATION, \
c1d900c3 715 REPORT_LOCATION_ARGS(offset)); \
147508a2
KW
716} STMT_END
717
668c081a 718#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 719 const IV offset = loc - RExC_precomp; \
499333dc 720 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18 721 m REPORT_LOCATION, \
c1d900c3 722 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
723} STMT_END
724
b23eb183 725#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 726 const IV offset = loc - RExC_precomp; \
499333dc 727 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3 728 m REPORT_LOCATION, \
c1d900c3 729 a1, REPORT_LOCATION_ARGS(offset)); \
2335b3d3
KW
730} STMT_END
731
668c081a 732#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 733 const IV offset = loc - RExC_precomp; \
499333dc 734 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 735 a1, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
736} STMT_END
737
738#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 739 const IV offset = loc - RExC_precomp; \
499333dc 740 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 741 a1, a2, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
742} STMT_END
743
668c081a
NC
744#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
745 const IV offset = loc - RExC_precomp; \
499333dc 746 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 747 a1, a2, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
748} STMT_END
749
ccb2c380 750#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 751 const IV offset = loc - RExC_precomp; \
499333dc 752 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 753 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
754} STMT_END
755
668c081a
NC
756#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
757 const IV offset = loc - RExC_precomp; \
499333dc 758 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 759 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
760} STMT_END
761
ccb2c380 762#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 763 const IV offset = loc - RExC_precomp; \
499333dc 764 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 765 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 766} STMT_END
9d1d55b5 767
538e84ed 768/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
769 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
770 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
771 * Element 0 holds the number n.
07be1b83 772 * Position is 1 indexed.
fac92740 773 */
7122b237
YO
774#ifndef RE_TRACK_PATTERN_OFFSETS
775#define Set_Node_Offset_To_R(node,byte)
776#define Set_Node_Offset(node,byte)
777#define Set_Cur_Node_Offset
778#define Set_Node_Length_To_R(node,len)
779#define Set_Node_Length(node,len)
6a86c6ad 780#define Set_Node_Cur_Length(node,start)
538e84ed
KW
781#define Node_Offset(n)
782#define Node_Length(n)
7122b237
YO
783#define Set_Node_Offset_Length(node,offset,len)
784#define ProgLen(ri) ri->u.proglen
785#define SetProgLen(ri,x) ri->u.proglen = x
786#else
787#define ProgLen(ri) ri->u.offsets[0]
788#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
789#define Set_Node_Offset_To_R(node,byte) STMT_START { \
790 if (! SIZE_ONLY) { \
791 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 792 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 793 if((node) < 0) { \
538e84ed
KW
794 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
795 (int)(node)); \
ccb2c380
MP
796 } else { \
797 RExC_offsets[2*(node)-1] = (byte); \
798 } \
799 } \
800} STMT_END
801
802#define Set_Node_Offset(node,byte) \
803 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
804#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
805
806#define Set_Node_Length_To_R(node,len) STMT_START { \
807 if (! SIZE_ONLY) { \
808 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 809 __LINE__, (int)(node), (int)(len))); \
ccb2c380 810 if((node) < 0) { \
538e84ed
KW
811 Perl_croak(aTHX_ "value of node is %d in Length macro", \
812 (int)(node)); \
ccb2c380
MP
813 } else { \
814 RExC_offsets[2*(node)] = (len); \
815 } \
816 } \
817} STMT_END
818
819#define Set_Node_Length(node,len) \
820 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
821#define Set_Node_Cur_Length(node, start) \
822 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
823
824/* Get offsets and lengths */
825#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
826#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
827
07be1b83
YO
828#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
829 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
830 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
831} STMT_END
7122b237 832#endif
07be1b83
YO
833
834#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
835#define EXPERIMENTAL_INPLACESCAN
f427392e 836#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 837
9e9ecfdf 838#define DEBUG_RExC_seen() \
538e84ed
KW
839 DEBUG_OPTIMISE_MORE_r({ \
840 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
841 \
e384d5c1
YO
842 if (RExC_seen & REG_ZERO_LEN_SEEN) \
843 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
538e84ed 844 \
e384d5c1
YO
845 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
846 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
538e84ed 847 \
e384d5c1
YO
848 if (RExC_seen & REG_GPOS_SEEN) \
849 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
538e84ed 850 \
e384d5c1
YO
851 if (RExC_seen & REG_RECURSE_SEEN) \
852 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
538e84ed 853 \
e384d5c1
YO
854 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
855 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 856 \
e384d5c1
YO
857 if (RExC_seen & REG_VERBARG_SEEN) \
858 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
538e84ed 859 \
e384d5c1
YO
860 if (RExC_seen & REG_CUTGROUP_SEEN) \
861 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
538e84ed 862 \
e384d5c1
YO
863 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
864 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 865 \
e384d5c1
YO
866 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
867 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 868 \
e384d5c1
YO
869 if (RExC_seen & REG_GOSTART_SEEN) \
870 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
538e84ed 871 \
ee273784
YO
872 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
873 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
874 \
538e84ed 875 PerlIO_printf(Perl_debug_log,"\n"); \
9e9ecfdf
YO
876 });
877
fdfb4f21
YO
878#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
879 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
880
881#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
882 if ( ( flags ) ) { \
883 PerlIO_printf(Perl_debug_log, "%s", open_str); \
884 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
885 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
886 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
887 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
888 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
889 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
890 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
891 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
892 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
893 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
894 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
895 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
896 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
897 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
898 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
899 PerlIO_printf(Perl_debug_log, "%s", close_str); \
900 }
901
902
304ee84b
YO
903#define DEBUG_STUDYDATA(str,data,depth) \
904DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 905 PerlIO_printf(Perl_debug_log, \
304ee84b 906 "%*s" str "Pos:%"IVdf"/%"IVdf \
fdfb4f21 907 " Flags: 0x%"UVXf, \
1de06328
YO
908 (int)(depth)*2, "", \
909 (IV)((data)->pos_min), \
910 (IV)((data)->pos_delta), \
fdfb4f21
YO
911 (UV)((data)->flags) \
912 ); \
913 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
914 PerlIO_printf(Perl_debug_log, \
915 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328 916 (IV)((data)->whilem_c), \
304ee84b
YO
917 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
918 is_inf ? "INF " : "" \
1de06328
YO
919 ); \
920 if ((data)->last_found) \
921 PerlIO_printf(Perl_debug_log, \
922 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
923 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
924 SvPVX_const((data)->last_found), \
925 (IV)((data)->last_end), \
926 (IV)((data)->last_start_min), \
927 (IV)((data)->last_start_max), \
928 ((data)->longest && \
929 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
930 SvPVX_const((data)->longest_fixed), \
931 (IV)((data)->offset_fixed), \
932 ((data)->longest && \
933 (data)->longest==&((data)->longest_float)) ? "*" : "", \
934 SvPVX_const((data)->longest_float), \
935 (IV)((data)->offset_float_min), \
936 (IV)((data)->offset_float_max) \
937 ); \
938 PerlIO_printf(Perl_debug_log,"\n"); \
939});
940
8e35b056
KW
941/* is c a control character for which we have a mnemonic? */
942#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
943
549b4e78
KW
944STATIC const char *
945S_cntrl_to_mnemonic(const U8 c)
946{
947 /* Returns the mnemonic string that represents character 'c', if one
948 * exists; NULL otherwise. The only ones that exist for the purposes of
949 * this routine are a few control characters */
950
951 switch (c) {
952 case '\a': return "\\a";
953 case '\b': return "\\b";
954 case ESC_NATIVE: return "\\e";
955 case '\f': return "\\f";
956 case '\n': return "\\n";
957 case '\r': return "\\r";
958 case '\t': return "\\t";
959 }
960
961 return NULL;
962}
963
653099ff 964/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 965 Update the longest found anchored substring and the longest found
653099ff
GS
966 floating substrings if needed. */
967
4327152a 968STATIC void
ea3daa5d
FC
969S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
970 SSize_t *minlenp, int is_inf)
c277df42 971{
e1ec3a88
AL
972 const STRLEN l = CHR_SVLEN(data->last_found);
973 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 974 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 975
7918f24d
NC
976 PERL_ARGS_ASSERT_SCAN_COMMIT;
977
c277df42 978 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 979 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
980 if (*data->longest == data->longest_fixed) {
981 data->offset_fixed = l ? data->last_start_min : data->pos_min;
982 if (data->flags & SF_BEFORE_EOL)
b81d288d 983 data->flags
c277df42
IZ
984 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
985 else
986 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 987 data->minlen_fixed=minlenp;
1de06328 988 data->lookbehind_fixed=0;
a0ed51b3 989 }
304ee84b 990 else { /* *data->longest == data->longest_float */
c277df42 991 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 992 data->offset_float_max = (l
646e8787
DM
993 ? data->last_start_max
994 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
995 ? SSize_t_MAX
996 : data->pos_min + data->pos_delta));
997 if (is_inf
998 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
999 data->offset_float_max = SSize_t_MAX;
c277df42 1000 if (data->flags & SF_BEFORE_EOL)
b81d288d 1001 data->flags
c277df42
IZ
1002 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1003 else
1004 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
1005 data->minlen_float=minlenp;
1006 data->lookbehind_float=0;
c277df42
IZ
1007 }
1008 }
1009 SvCUR_set(data->last_found, 0);
0eda9292 1010 {
a28509cc 1011 SV * const sv = data->last_found;
097eb12c
AL
1012 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1013 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1014 if (mg)
1015 mg->mg_len = 0;
1016 }
0eda9292 1017 }
c277df42
IZ
1018 data->last_end = -1;
1019 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1020 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1021}
1022
cdd87c1d
KW
1023/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1024 * list that describes which code points it matches */
1025
653099ff 1026STATIC void
3420edd7 1027S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1028{
cdd87c1d
KW
1029 /* Set the SSC 'ssc' to match an empty string or any code point */
1030
557bd3fb 1031 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1032
71068078 1033 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1034
cdd87c1d
KW
1035 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1036 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
93e92956 1037 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1038}
1039
653099ff 1040STATIC int
dc3bf405 1041S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1042{
c144baaa
KW
1043 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1044 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1045 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1046 * in any way, so there's no point in using it */
cdd87c1d
KW
1047
1048 UV start, end;
1049 bool ret;
653099ff 1050
557bd3fb 1051 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1052
71068078 1053 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1054
93e92956 1055 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1056 return FALSE;
1057 }
1058
1059 /* See if the list consists solely of the range 0 - Infinity */
1060 invlist_iterinit(ssc->invlist);
1061 ret = invlist_iternext(ssc->invlist, &start, &end)
1062 && start == 0
1063 && end == UV_MAX;
1064
1065 invlist_iterfinish(ssc->invlist);
1066
1067 if (ret) {
1068 return TRUE;
1069 }
1070
1071 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1072 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1073 int i;
1074 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1075 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1076 return TRUE;
1077 }
1078 }
1079 }
1080
1081 return FALSE;
653099ff
GS
1082}
1083
653099ff 1084STATIC void
cdd87c1d 1085S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1086{
cdd87c1d
KW
1087 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1088 * string, any code point, or any posix class under locale */
1089
557bd3fb 1090 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1091
557bd3fb 1092 Zero(ssc, 1, regnode_ssc);
71068078 1093 set_ANYOF_SYNTHETIC(ssc);
93e92956 1094 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1095 ssc_anything(ssc);
cdd87c1d 1096
2f306ab9
KW
1097 /* If any portion of the regex is to operate under locale rules that aren't
1098 * fully known at compile time, initialization includes it. The reason
1099 * this isn't done for all regexes is that the optimizer was written under
1100 * the assumption that locale was all-or-nothing. Given the complexity and
1101 * lack of documentation in the optimizer, and that there are inadequate
1102 * test cases for locale, many parts of it may not work properly, it is
1103 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1104 if (RExC_contains_locale) {
1105 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1106 }
1107 else {
1108 ANYOF_POSIXL_ZERO(ssc);
1109 }
653099ff
GS
1110}
1111
b423522f 1112STATIC int
dc3bf405
BF
1113S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1114 const regnode_ssc *ssc)
b423522f
KW
1115{
1116 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1117 * to the list of code points matched, and locale posix classes; hence does
1118 * not check its flags) */
1119
1120 UV start, end;
1121 bool ret;
1122
1123 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1124
71068078 1125 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1126
1127 invlist_iterinit(ssc->invlist);
1128 ret = invlist_iternext(ssc->invlist, &start, &end)
1129 && start == 0
1130 && end == UV_MAX;
1131
1132 invlist_iterfinish(ssc->invlist);
1133
1134 if (! ret) {
1135 return FALSE;
1136 }
1137
e0e1be5f 1138 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1139 return FALSE;
b423522f
KW
1140 }
1141
1142 return TRUE;
1143}
1144
1145STATIC SV*
1146S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1147 const regnode_charclass* const node)
b423522f
KW
1148{
1149 /* Returns a mortal inversion list defining which code points are matched
1150 * by 'node', which is of type ANYOF. Handles complementing the result if
1151 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1152 * returned list must, and will, contain every code point that is a
1153 * possibility. */
b423522f
KW
1154
1155 SV* invlist = sv_2mortal(_new_invlist(0));
1ee208c4 1156 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1157 unsigned int i;
1158 const U32 n = ARG(node);
31f05a37 1159 bool new_node_has_latin1 = FALSE;
b423522f
KW
1160
1161 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1162
1163 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1164 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1165 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1166 AV * const av = MUTABLE_AV(SvRV(rv));
1167 SV **const ary = AvARRAY(av);
1168 assert(RExC_rxi->data->what[n] == 's');
1169
1170 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1171 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1172 }
1173 else if (ary[0] && ary[0] != &PL_sv_undef) {
1174
1175 /* Here, no compile-time swash, and there are things that won't be
1176 * known until runtime -- we have to assume it could be anything */
1177 return _add_range_to_invlist(invlist, 0, UV_MAX);
1178 }
1ee208c4 1179 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1180
1181 /* Here no compile-time swash, and no run-time only data. Use the
1182 * node's inversion list */
1ee208c4
KW
1183 invlist = sv_2mortal(invlist_clone(ary[3]));
1184 }
1185
1186 /* Get the code points valid only under UTF-8 locales */
1187 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1188 && ary[2] && ary[2] != &PL_sv_undef)
1189 {
1190 only_utf8_locale_invlist = ary[2];
b423522f
KW
1191 }
1192 }
1193
dcb20b36
KW
1194 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1195 * code points, and an inversion list for the others, but if there are code
1196 * points that should match only conditionally on the target string being
1197 * UTF-8, those are placed in the inversion list, and not the bitmap.
1198 * Since there are circumstances under which they could match, they are
1199 * included in the SSC. But if the ANYOF node is to be inverted, we have
1200 * to exclude them here, so that when we invert below, the end result
1201 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1202 * have to do this here before we add the unconditionally matched code
1203 * points */
b423522f
KW
1204 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1205 _invlist_intersection_complement_2nd(invlist,
1206 PL_UpperLatin1,
1207 &invlist);
1208 }
1209
1210 /* Add in the points from the bit map */
dcb20b36 1211 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f
KW
1212 if (ANYOF_BITMAP_TEST(node, i)) {
1213 invlist = add_cp_to_invlist(invlist, i);
31f05a37 1214 new_node_has_latin1 = TRUE;
b423522f
KW
1215 }
1216 }
1217
1218 /* If this can match all upper Latin1 code points, have to add them
1219 * as well */
f240c685
KW
1220 if (OP(node) == ANYOFD
1221 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1222 {
b423522f
KW
1223 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1224 }
1225
1226 /* Similarly for these */
93e92956 1227 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1228 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1229 }
1230
1231 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1232 _invlist_invert(invlist);
1233 }
31f05a37
KW
1234 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1235
1236 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1237 * locale. We can skip this if there are no 0-255 at all. */
1238 _invlist_union(invlist, PL_Latin1, &invlist);
1239 }
1240
1ee208c4
KW
1241 /* Similarly add the UTF-8 locale possible matches. These have to be
1242 * deferred until after the non-UTF-8 locale ones are taken care of just
1243 * above, or it leads to wrong results under ANYOF_INVERT */
1244 if (only_utf8_locale_invlist) {
31f05a37 1245 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1246 only_utf8_locale_invlist,
31f05a37
KW
1247 ANYOF_FLAGS(node) & ANYOF_INVERT,
1248 &invlist);
1249 }
b423522f
KW
1250
1251 return invlist;
1252}
1253
1051e1c4 1254/* These two functions currently do the exact same thing */
557bd3fb 1255#define ssc_init_zero ssc_init
653099ff 1256
cdd87c1d
KW
1257#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1258#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1259
557bd3fb 1260/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1261 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1262 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1263
653099ff 1264STATIC void
b423522f 1265S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1266 const regnode_charclass *and_with)
653099ff 1267{
cdd87c1d
KW
1268 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1269 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1270
a0dd4231
KW
1271 SV* anded_cp_list;
1272 U8 anded_flags;
1e6ade67 1273
cdd87c1d 1274 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1275
71068078 1276 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1277
1278 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1279 * the code point inversion list and just the relevant flags */
71068078 1280 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1281 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1282 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1283
1284 /* XXX This is a kludge around what appears to be deficiencies in the
1285 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1286 * there are paths through the optimizer where it doesn't get weeded
1287 * out when it should. And if we don't make some extra provision for
1288 * it like the code just below, it doesn't get added when it should.
1289 * This solution is to add it only when AND'ing, which is here, and
1290 * only when what is being AND'ed is the pristine, original node
1291 * matching anything. Thus it is like adding it to ssc_anything() but
1292 * only when the result is to be AND'ed. Probably the same solution
1293 * could be adopted for the same problem we have with /l matching,
1294 * which is solved differently in S_ssc_init(), and that would lead to
1295 * fewer false positives than that solution has. But if this solution
1296 * creates bugs, the consequences are only that a warning isn't raised
1297 * that should be; while the consequences for having /l bugs is
1298 * incorrect matches */
7dcac5f6 1299 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1300 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1301 }
a0dd4231
KW
1302 }
1303 else {
5c0f85ef 1304 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1305 if (OP(and_with) == ANYOFD) {
1306 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1307 }
1308 else {
1309 anded_flags = ANYOF_FLAGS(and_with)
1310 &( ANYOF_COMMON_FLAGS
1311 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER);
1312 }
a0dd4231
KW
1313 }
1314
1315 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1316
1317 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1318 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1319 * 'and_with' may be inverted. When not inverted, we have the situation of
1320 * computing:
1321 * (C1 | P1) & (C2 | P2)
1322 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1323 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1324 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1325 * <= ((C1 & C2) | P1 | P2)
1326 * Alternatively, the last few steps could be:
1327 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1328 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1329 * <= (C1 | C2 | (P1 & P2))
1330 * We favor the second approach if either P1 or P2 is non-empty. This is
1331 * because these components are a barrier to doing optimizations, as what
1332 * they match cannot be known until the moment of matching as they are
1333 * dependent on the current locale, 'AND"ing them likely will reduce or
1334 * eliminate them.
1335 * But we can do better if we know that C1,P1 are in their initial state (a
1336 * frequent occurrence), each matching everything:
1337 * (<everything>) & (C2 | P2) = C2 | P2
1338 * Similarly, if C2,P2 are in their initial state (again a frequent
1339 * occurrence), the result is a no-op
1340 * (C1 | P1) & (<everything>) = C1 | P1
1341 *
1342 * Inverted, we have
1343 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1344 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1345 * <= (C1 & ~C2) | (P1 & ~P2)
1346 * */
1aa99e6b 1347
a0dd4231 1348 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1349 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1350 {
cdd87c1d 1351 unsigned int i;
8951c461 1352
cdd87c1d
KW
1353 ssc_intersection(ssc,
1354 anded_cp_list,
1355 FALSE /* Has already been inverted */
1356 );
c6b76537 1357
cdd87c1d
KW
1358 /* If either P1 or P2 is empty, the intersection will be also; can skip
1359 * the loop */
93e92956 1360 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1361 ANYOF_POSIXL_ZERO(ssc);
1362 }
e0e1be5f 1363 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1364
1365 /* Note that the Posix class component P from 'and_with' actually
1366 * looks like:
1367 * P = Pa | Pb | ... | Pn
1368 * where each component is one posix class, such as in [\w\s].
1369 * Thus
1370 * ~P = ~(Pa | Pb | ... | Pn)
1371 * = ~Pa & ~Pb & ... & ~Pn
1372 * <= ~Pa | ~Pb | ... | ~Pn
1373 * The last is something we can easily calculate, but unfortunately
1374 * is likely to have many false positives. We could do better
1375 * in some (but certainly not all) instances if two classes in
1376 * P have known relationships. For example
1377 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1378 * So
1379 * :lower: & :print: = :lower:
1380 * And similarly for classes that must be disjoint. For example,
1381 * since \s and \w can have no elements in common based on rules in
1382 * the POSIX standard,
1383 * \w & ^\S = nothing
1384 * Unfortunately, some vendor locales do not meet the Posix
1385 * standard, in particular almost everything by Microsoft.
1386 * The loop below just changes e.g., \w into \W and vice versa */
1387
1ee208c4 1388 regnode_charclass_posixl temp;
cdd87c1d
KW
1389 int add = 1; /* To calculate the index of the complement */
1390
1391 ANYOF_POSIXL_ZERO(&temp);
1392 for (i = 0; i < ANYOF_MAX; i++) {
1393 assert(i % 2 != 0
7dcac5f6
KW
1394 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1395 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1396
7dcac5f6 1397 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1398 ANYOF_POSIXL_SET(&temp, i + add);
1399 }
1400 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1401 }
1402 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1403
cdd87c1d
KW
1404 } /* else ssc already has no posixes */
1405 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1406 in its initial state */
71068078 1407 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1408 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1409 {
1410 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1411 * copy it over 'ssc' */
1412 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1413 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1414 StructCopy(and_with, ssc, regnode_ssc);
1415 }
1416 else {
1417 ssc->invlist = anded_cp_list;
1418 ANYOF_POSIXL_ZERO(ssc);
93e92956 1419 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1420 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1421 }
1422 }
1423 }
e0e1be5f 1424 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1425 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1426 {
1427 /* One or the other of P1, P2 is non-empty. */
93e92956 1428 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1429 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1430 }
cdd87c1d
KW
1431 ssc_union(ssc, anded_cp_list, FALSE);
1432 }
1433 else { /* P1 = P2 = empty */
1434 ssc_intersection(ssc, anded_cp_list, FALSE);
1435 }
137165a6 1436 }
653099ff
GS
1437}
1438
653099ff 1439STATIC void
cdd87c1d 1440S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1441 const regnode_charclass *or_with)
653099ff 1442{
cdd87c1d
KW
1443 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1444 * another SSC or a regular ANYOF class. Can create false positives if
1445 * 'or_with' is to be inverted. */
7918f24d 1446
a0dd4231
KW
1447 SV* ored_cp_list;
1448 U8 ored_flags;
c6b76537 1449
cdd87c1d 1450 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1451
71068078 1452 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1453
1454 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1455 * the code point inversion list and just the relevant flags */
71068078 1456 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1457 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1458 ored_flags = ANYOF_FLAGS(or_with);
1459 }
1460 else {
5c0f85ef 1461 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1462 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1463 if (OP(or_with) != ANYOFD) {
1464 ored_flags
1465 |= ANYOF_FLAGS(or_with)
1466 & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1467 }
a0dd4231
KW
1468 }
1469
1470 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1471
1472 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1473 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1474 * 'or_with' may be inverted. When not inverted, we have the simple
1475 * situation of computing:
1476 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1477 * If P1|P2 yields a situation with both a class and its complement are
1478 * set, like having both \w and \W, this matches all code points, and we
1479 * can delete these from the P component of the ssc going forward. XXX We
1480 * might be able to delete all the P components, but I (khw) am not certain
1481 * about this, and it is better to be safe.
1482 *
1483 * Inverted, we have
1484 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1485 * <= (C1 | P1) | ~C2
1486 * <= (C1 | ~C2) | P1
1487 * (which results in actually simpler code than the non-inverted case)
1488 * */
9826f543 1489
a0dd4231 1490 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1491 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1492 {
cdd87c1d 1493 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1494 } /* else Not inverted */
93e92956 1495 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1496 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1497 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1498 unsigned int i;
1499 for (i = 0; i < ANYOF_MAX; i += 2) {
1500 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1501 {
1502 ssc_match_all_cp(ssc);
1503 ANYOF_POSIXL_CLEAR(ssc, i);
1504 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1505 }
1506 }
1507 }
1aa99e6b 1508 }
cdd87c1d
KW
1509
1510 ssc_union(ssc,
1511 ored_cp_list,
1512 FALSE /* Already has been inverted */
1513 );
653099ff
GS
1514}
1515
b423522f
KW
1516PERL_STATIC_INLINE void
1517S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1518{
1519 PERL_ARGS_ASSERT_SSC_UNION;
1520
71068078 1521 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1522
1523 _invlist_union_maybe_complement_2nd(ssc->invlist,
1524 invlist,
1525 invert2nd,
1526 &ssc->invlist);
1527}
1528
1529PERL_STATIC_INLINE void
1530S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1531 SV* const invlist,
1532 const bool invert2nd)
1533{
1534 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1535
71068078 1536 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1537
1538 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1539 invlist,
1540 invert2nd,
1541 &ssc->invlist);
1542}
1543
1544PERL_STATIC_INLINE void
1545S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1546{
1547 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1548
71068078 1549 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1550
1551 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1552}
1553
1554PERL_STATIC_INLINE void
1555S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1556{
1557 /* AND just the single code point 'cp' into the SSC 'ssc' */
1558
1559 SV* cp_list = _new_invlist(2);
1560
1561 PERL_ARGS_ASSERT_SSC_CP_AND;
1562
71068078 1563 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1564
1565 cp_list = add_cp_to_invlist(cp_list, cp);
1566 ssc_intersection(ssc, cp_list,
1567 FALSE /* Not inverted */
1568 );
1569 SvREFCNT_dec_NN(cp_list);
1570}
1571
1572PERL_STATIC_INLINE void
dc3bf405 1573S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1574{
1575 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1576 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1577
71068078 1578 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1579
1580 ANYOF_POSIXL_ZERO(ssc);
1581 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1582}
1583
b35552de
KW
1584#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1585
1586STATIC bool
1587S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1588{
1589 /* The synthetic start class is used to hopefully quickly winnow down
1590 * places where a pattern could start a match in the target string. If it
1591 * doesn't really narrow things down that much, there isn't much point to
1592 * having the overhead of using it. This function uses some very crude
1593 * heuristics to decide if to use the ssc or not.
1594 *
1595 * It returns TRUE if 'ssc' rules out more than half what it considers to
1596 * be the "likely" possible matches, but of course it doesn't know what the
1597 * actual things being matched are going to be; these are only guesses
1598 *
1599 * For /l matches, it assumes that the only likely matches are going to be
1600 * in the 0-255 range, uniformly distributed, so half of that is 127
1601 * For /a and /d matches, it assumes that the likely matches will be just
1602 * the ASCII range, so half of that is 63
1603 * For /u and there isn't anything matching above the Latin1 range, it
1604 * assumes that that is the only range likely to be matched, and uses
1605 * half that as the cut-off: 127. If anything matches above Latin1,
1606 * it assumes that all of Unicode could match (uniformly), except for
1607 * non-Unicode code points and things in the General Category "Other"
1608 * (unassigned, private use, surrogates, controls and formats). This
1609 * is a much large number. */
1610
1611 const U32 max_match = (LOC)
1612 ? 127
1613 : (! UNI_SEMANTICS)
1614 ? 63
1615 : (invlist_highest(ssc->invlist) < 256)
1616 ? 127
1617 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1618 U32 count = 0; /* Running total of number of code points matched by
1619 'ssc' */
1620 UV start, end; /* Start and end points of current range in inversion
1621 list */
1622
1623 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1624
1625 invlist_iterinit(ssc->invlist);
1626 while (invlist_iternext(ssc->invlist, &start, &end)) {
1627
1628 /* /u is the only thing that we expect to match above 255; so if not /u
1629 * and even if there are matches above 255, ignore them. This catches
1630 * things like \d under /d which does match the digits above 255, but
1631 * since the pattern is /d, it is not likely to be expecting them */
1632 if (! UNI_SEMANTICS) {
1633 if (start > 255) {
1634 break;
1635 }
1636 end = MIN(end, 255);
1637 }
1638 count += end - start + 1;
1639 if (count > max_match) {
1640 invlist_iterfinish(ssc->invlist);
1641 return FALSE;
1642 }
1643 }
1644
1645 return TRUE;
1646}
1647
1648
b423522f
KW
1649STATIC void
1650S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1651{
1652 /* The inversion list in the SSC is marked mortal; now we need a more
1653 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1654 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1655 * map */
b423522f
KW
1656
1657 SV* invlist = invlist_clone(ssc->invlist);
1658
1659 PERL_ARGS_ASSERT_SSC_FINALIZE;
1660
71068078 1661 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1662
a0dd4231 1663 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1664 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1665 * by the time we reach here */
f240c685
KW
1666 assert(! (ANYOF_FLAGS(ssc)
1667 & ~( ANYOF_COMMON_FLAGS
1668 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)));
a0dd4231 1669
b423522f
KW
1670 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1671
1ee208c4
KW
1672 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1673 NULL, NULL, NULL, FALSE);
b423522f 1674
85c8e306
KW
1675 /* Make sure is clone-safe */
1676 ssc->invlist = NULL;
1677
e0e1be5f 1678 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1679 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1680 }
1462525b 1681
b2e90ddf
KW
1682 if (RExC_contains_locale) {
1683 OP(ssc) = ANYOFL;
1684 }
1685
1462525b 1686 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1687}
1688
a3621e74
YO
1689#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1690#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1691#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1692#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1693 ? (TRIE_LIST_CUR( idx ) - 1) \
1694 : 0 )
a3621e74 1695
3dab1dad
YO
1696
1697#ifdef DEBUGGING
07be1b83 1698/*
2b8b4781
NC
1699 dump_trie(trie,widecharmap,revcharmap)
1700 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1701 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1702
1703 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1704 The _interim_ variants are used for debugging the interim
1705 tables that are used to generate the final compressed
1706 representation which is what dump_trie expects.
1707
486ec47a 1708 Part of the reason for their existence is to provide a form
3dab1dad 1709 of documentation as to how the different representations function.
07be1b83
YO
1710
1711*/
3dab1dad
YO
1712
1713/*
3dab1dad
YO
1714 Dumps the final compressed table form of the trie to Perl_debug_log.
1715 Used for debugging make_trie().
1716*/
b9a59e08 1717
3dab1dad 1718STATIC void
2b8b4781
NC
1719S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1720 AV *revcharmap, U32 depth)
3dab1dad
YO
1721{
1722 U32 state;
ab3bbdeb 1723 SV *sv=sv_newmortal();
55eed653 1724 int colwidth= widecharmap ? 6 : 4;
2e64971a 1725 U16 word;
3dab1dad
YO
1726 GET_RE_DEBUG_FLAGS_DECL;
1727
7918f24d 1728 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1729
3dab1dad
YO
1730 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1731 (int)depth * 2 + 2,"",
1732 "Match","Base","Ofs" );
1733
1734 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1735 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1736 if ( tmp ) {
538e84ed 1737 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1738 colwidth,
538e84ed 1739 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1740 PL_colors[0], PL_colors[1],
1741 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1742 PERL_PV_ESCAPE_FIRSTCHAR
1743 )
ab3bbdeb 1744 );
3dab1dad
YO
1745 }
1746 }
1747 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1748 (int)depth * 2 + 2,"");
1749
1750 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1751 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1752 PerlIO_printf( Perl_debug_log, "\n");
1753
1e2e3d02 1754 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1755 const U32 base = trie->states[ state ].trans.base;
3dab1dad 1756
538e84ed
KW
1757 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1758 (int)depth * 2 + 2,"", (UV)state);
3dab1dad
YO
1759
1760 if ( trie->states[ state ].wordnum ) {
538e84ed
KW
1761 PerlIO_printf( Perl_debug_log, " W%4X",
1762 trie->states[ state ].wordnum );
3dab1dad
YO
1763 } else {
1764 PerlIO_printf( Perl_debug_log, "%6s", "" );
1765 }
1766
1767 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1768
1769 if ( base ) {
1770 U32 ofs = 0;
1771
1772 while( ( base + ofs < trie->uniquecharcount ) ||
1773 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
1774 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1775 != state))
3dab1dad
YO
1776 ofs++;
1777
1778 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1779
1780 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
1781 if ( ( base + ofs >= trie->uniquecharcount )
1782 && ( base + ofs - trie->uniquecharcount
1783 < trie->lasttrans )
1784 && trie->trans[ base + ofs
1785 - trie->uniquecharcount ].check == state )
3dab1dad 1786 {
ab3bbdeb
YO
1787 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1788 colwidth,
538e84ed
KW
1789 (UV)trie->trans[ base + ofs
1790 - trie->uniquecharcount ].next );
3dab1dad 1791 } else {
ab3bbdeb 1792 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1793 }
1794 }
1795
1796 PerlIO_printf( Perl_debug_log, "]");
1797
1798 }
1799 PerlIO_printf( Perl_debug_log, "\n" );
1800 }
538e84ed
KW
1801 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1802 (int)depth*2, "");
2e64971a
DM
1803 for (word=1; word <= trie->wordcount; word++) {
1804 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1805 (int)word, (int)(trie->wordinfo[word].prev),
1806 (int)(trie->wordinfo[word].len));
1807 }
1808 PerlIO_printf(Perl_debug_log, "\n" );
538e84ed 1809}
3dab1dad 1810/*
3dab1dad 1811 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 1812 List tries normally only are used for construction when the number of
3dab1dad
YO
1813 possible chars (trie->uniquecharcount) is very high.
1814 Used for debugging make_trie().
1815*/
1816STATIC void
55eed653 1817S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1818 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1819 U32 depth)
3dab1dad
YO
1820{
1821 U32 state;
ab3bbdeb 1822 SV *sv=sv_newmortal();
55eed653 1823 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1824 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1825
1826 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1827
3dab1dad 1828 /* print out the table precompression. */
ab3bbdeb
YO
1829 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1830 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1831 "------:-----+-----------------\n" );
538e84ed 1832
3dab1dad
YO
1833 for( state=1 ; state < next_alloc ; state ++ ) {
1834 U16 charid;
538e84ed 1835
ab3bbdeb 1836 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1837 (int)depth * 2 + 2,"", (UV)state );
1838 if ( ! trie->states[ state ].wordnum ) {
1839 PerlIO_printf( Perl_debug_log, "%5s| ","");
1840 } else {
1841 PerlIO_printf( Perl_debug_log, "W%4x| ",
1842 trie->states[ state ].wordnum
1843 );
1844 }
1845 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
1846 SV ** const tmp = av_fetch( revcharmap,
1847 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1848 if ( tmp ) {
1849 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1850 colwidth,
538e84ed
KW
1851 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1852 colwidth,
1853 PL_colors[0], PL_colors[1],
1854 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1855 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 1856 ) ,
1e2e3d02
YO
1857 TRIE_LIST_ITEM(state,charid).forid,
1858 (UV)TRIE_LIST_ITEM(state,charid).newstate
1859 );
538e84ed 1860 if (!(charid % 10))
664e119d
RGS
1861 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1862 (int)((depth * 2) + 14), "");
1e2e3d02 1863 }
ab3bbdeb
YO
1864 }
1865 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad 1866 }
538e84ed 1867}
3dab1dad
YO
1868
1869/*
3dab1dad 1870 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
1871 This is the normal DFA style state transition table, with a few
1872 twists to facilitate compression later.
3dab1dad
YO
1873 Used for debugging make_trie().
1874*/
1875STATIC void
55eed653 1876S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1877 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1878 U32 depth)
3dab1dad
YO
1879{
1880 U32 state;
1881 U16 charid;
ab3bbdeb 1882 SV *sv=sv_newmortal();
55eed653 1883 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1884 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1885
1886 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 1887
3dab1dad
YO
1888 /*
1889 print out the table precompression so that we can do a visual check
1890 that they are identical.
1891 */
538e84ed 1892
3dab1dad
YO
1893 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1894
1895 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1896 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1897 if ( tmp ) {
538e84ed 1898 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1899 colwidth,
538e84ed 1900 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1901 PL_colors[0], PL_colors[1],
1902 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1903 PERL_PV_ESCAPE_FIRSTCHAR
1904 )
ab3bbdeb 1905 );
3dab1dad
YO
1906 }
1907 }
1908
1909 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1910
1911 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1912 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1913 }
1914
1915 PerlIO_printf( Perl_debug_log, "\n" );
1916
1917 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1918
538e84ed 1919 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
3dab1dad
YO
1920 (int)depth * 2 + 2,"",
1921 (UV)TRIE_NODENUM( state ) );
1922
1923 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1924 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1925 if (v)
1926 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1927 else
1928 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1929 }
1930 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
538e84ed
KW
1931 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1932 (UV)trie->trans[ state ].check );
3dab1dad 1933 } else {
538e84ed
KW
1934 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1935 (UV)trie->trans[ state ].check,
3dab1dad
YO
1936 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1937 }
1938 }
07be1b83 1939}
3dab1dad
YO
1940
1941#endif
1942
2e64971a 1943
786e8c11
YO
1944/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1945 startbranch: the first branch in the whole branch sequence
1946 first : start branch of sequence of branch-exact nodes.
1947 May be the same as startbranch
1948 last : Thing following the last branch.
1949 May be the same as tail.
1950 tail : item following the branch sequence
1951 count : words in the sequence
a4525e78 1952 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 1953 depth : indent depth
3dab1dad 1954
786e8c11 1955Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1956
786e8c11
YO
1957A trie is an N'ary tree where the branches are determined by digital
1958decomposition of the key. IE, at the root node you look up the 1st character and
1959follow that branch repeat until you find the end of the branches. Nodes can be
1960marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1961
786e8c11 1962 /he|she|his|hers/
72f13be8 1963
786e8c11
YO
1964would convert into the following structure. Numbers represent states, letters
1965following numbers represent valid transitions on the letter from that state, if
1966the number is in square brackets it represents an accepting state, otherwise it
1967will be in parenthesis.
07be1b83 1968
786e8c11
YO
1969 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1970 | |
1971 | (2)
1972 | |
1973 (1) +-i->(6)-+-s->[7]
1974 |
1975 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1976
786e8c11
YO
1977 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1978
1979This shows that when matching against the string 'hers' we will begin at state 1
1980read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1981then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1982is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1983single traverse. We store a mapping from accepting to state to which word was
1984matched, and then when we have multiple possibilities we try to complete the
b8fda935 1985rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
1986
1987The only prior NFA like behaviour that would be changed by the TRIE support is
1988the silent ignoring of duplicate alternations which are of the form:
1989
1990 / (DUPE|DUPE) X? (?{ ... }) Y /x
1991
4b714af6 1992Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1993and without the optimisation. With the optimisations dupes will be silently
486ec47a 1994ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1995the following demonstrates:
1996
1997 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1998
1999which prints out 'word' three times, but
2000
2001 'words'=~/(word|word|word)(?{ print $1 })S/
2002
2003which doesnt print it out at all. This is due to other optimisations kicking in.
2004
2005Example of what happens on a structural level:
2006
486ec47a 2007The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2008
2009 1: CURLYM[1] {1,32767}(18)
2010 5: BRANCH(8)
2011 6: EXACT <ac>(16)
2012 8: BRANCH(11)
2013 9: EXACT <ad>(16)
2014 11: BRANCH(14)
2015 12: EXACT <ab>(16)
2016 16: SUCCEED(0)
2017 17: NOTHING(18)
2018 18: END(0)
2019
2020This would be optimizable with startbranch=5, first=5, last=16, tail=16
2021and should turn into:
2022
2023 1: CURLYM[1] {1,32767}(18)
2024 5: TRIE(16)
2025 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2026 <ac>
2027 <ad>
2028 <ab>
2029 16: SUCCEED(0)
2030 17: NOTHING(18)
2031 18: END(0)
2032
2033Cases where tail != last would be like /(?foo|bar)baz/:
2034
2035 1: BRANCH(4)
2036 2: EXACT <foo>(8)
2037 4: BRANCH(7)
2038 5: EXACT <bar>(8)
2039 7: TAIL(8)
2040 8: EXACT <baz>(10)
2041 10: END(0)
2042
2043which would be optimizable with startbranch=1, first=1, last=7, tail=8
2044and would end up looking like:
2045
2046 1: TRIE(8)
2047 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2048 <foo>
2049 <bar>
2050 7: TAIL(8)
2051 8: EXACT <baz>(10)
2052 10: END(0)
2053
c80e42f3 2054 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2055
2056is the recommended Unicode-aware way of saying
2057
2058 *(d++) = uv;
2059*/
2060
fab2782b 2061#define TRIE_STORE_REVCHAR(val) \
786e8c11 2062 STMT_START { \
73031816 2063 if (UTF) { \
668fcfea 2064 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2065 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2066 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2067 SvCUR_set(zlopp, kapow - flrbbbbb); \
2068 SvPOK_on(zlopp); \
2069 SvUTF8_on(zlopp); \
2070 av_push(revcharmap, zlopp); \
2071 } else { \
fab2782b 2072 char ooooff = (char)val; \
73031816
NC
2073 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2074 } \
2075 } STMT_END
786e8c11 2076
914a25d5
KW
2077/* This gets the next character from the input, folding it if not already
2078 * folded. */
2079#define TRIE_READ_CHAR STMT_START { \
2080 wordlen++; \
2081 if ( UTF ) { \
2082 /* if it is UTF then it is either already folded, or does not need \
2083 * folding */ \
1c1d615a 2084 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2085 } \
2086 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2087 /* This folder implies Unicode rules, which in the range expressible \
2088 * by not UTF is the lower case, with the two exceptions, one of \
2089 * which should have been taken care of before calling this */ \
2090 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2091 uvc = toLOWER_L1(*uc); \
2092 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2093 len = 1; \
914a25d5
KW
2094 } else { \
2095 /* raw data, will be folded later if needed */ \
2096 uvc = (U32)*uc; \
2097 len = 1; \
2098 } \
786e8c11
YO
2099} STMT_END
2100
2101
2102
2103#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2104 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
2105 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2106 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
2107 } \
2108 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2109 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2110 TRIE_LIST_CUR( state )++; \
2111} STMT_END
07be1b83 2112
786e8c11
YO
2113#define TRIE_LIST_NEW(state) STMT_START { \
2114 Newxz( trie->states[ state ].trans.list, \
2115 4, reg_trie_trans_le ); \
2116 TRIE_LIST_CUR( state ) = 1; \
2117 TRIE_LIST_LEN( state ) = 4; \
2118} STMT_END
07be1b83 2119
786e8c11
YO
2120#define TRIE_HANDLE_WORD(state) STMT_START { \
2121 U16 dupe= trie->states[ state ].wordnum; \
2122 regnode * const noper_next = regnext( noper ); \
2123 \
786e8c11
YO
2124 DEBUG_r({ \
2125 /* store the word for dumping */ \
2126 SV* tmp; \
2127 if (OP(noper) != NOTHING) \
740cce10 2128 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2129 else \
740cce10 2130 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2131 av_push( trie_words, tmp ); \
786e8c11
YO
2132 }); \
2133 \
2134 curword++; \
2e64971a
DM
2135 trie->wordinfo[curword].prev = 0; \
2136 trie->wordinfo[curword].len = wordlen; \
2137 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2138 \
2139 if ( noper_next < tail ) { \
2140 if (!trie->jump) \
538e84ed
KW
2141 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2142 sizeof(U16) ); \
7f69552c 2143 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2144 if (!jumper) \
2145 jumper = noper_next; \
2146 if (!nextbranch) \
2147 nextbranch= regnext(cur); \
2148 } \
2149 \
2150 if ( dupe ) { \
2e64971a
DM
2151 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2152 /* chain, so that when the bits of chain are later */\
2153 /* linked together, the dups appear in the chain */\
2154 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2155 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2156 } else { \
2157 /* we haven't inserted this word yet. */ \
2158 trie->states[ state ].wordnum = curword; \
2159 } \
2160} STMT_END
07be1b83 2161
3dab1dad 2162
786e8c11
YO
2163#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2164 ( ( base + charid >= ucharcount \
2165 && base + charid < ubound \
2166 && state == trie->trans[ base - ucharcount + charid ].check \
2167 && trie->trans[ base - ucharcount + charid ].next ) \
2168 ? trie->trans[ base - ucharcount + charid ].next \
2169 : ( state==1 ? special : 0 ) \
2170 )
3dab1dad 2171
786e8c11
YO
2172#define MADE_TRIE 1
2173#define MADE_JUMP_TRIE 2
2174#define MADE_EXACT_TRIE 4
3dab1dad 2175
a3621e74 2176STATIC I32
538e84ed
KW
2177S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2178 regnode *first, regnode *last, regnode *tail,
2179 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2180{
2181 /* first pass, loop through and scan words */
2182 reg_trie_data *trie;
55eed653 2183 HV *widecharmap = NULL;
2b8b4781 2184 AV *revcharmap = newAV();
a3621e74 2185 regnode *cur;
a3621e74
YO
2186 STRLEN len = 0;
2187 UV uvc = 0;
2188 U16 curword = 0;
2189 U32 next_alloc = 0;
786e8c11
YO
2190 regnode *jumper = NULL;
2191 regnode *nextbranch = NULL;
7f69552c 2192 regnode *convert = NULL;
2e64971a 2193 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2194 /* we just use folder as a flag in utf8 */
1e696034 2195 const U8 * folder = NULL;
a3621e74 2196
2b8b4781 2197#ifdef DEBUGGING
cf78de0b 2198 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2199 AV *trie_words = NULL;
2200 /* along with revcharmap, this only used during construction but both are
2201 * useful during debugging so we store them in the struct when debugging.
8e11feef 2202 */
2b8b4781 2203#else
cf78de0b 2204 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2205 STRLEN trie_charcount=0;
3dab1dad 2206#endif
2b8b4781 2207 SV *re_trie_maxbuff;
a3621e74 2208 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2209
2210 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2211#ifndef DEBUGGING
2212 PERL_UNUSED_ARG(depth);
2213#endif
a3621e74 2214
1e696034 2215 switch (flags) {
a4525e78 2216 case EXACT: case EXACTL: break;
2f7f8cb1 2217 case EXACTFA:
fab2782b 2218 case EXACTFU_SS:
a4525e78
KW
2219 case EXACTFU:
2220 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2221 case EXACTF: folder = PL_fold; break;
fab2782b 2222 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2223 }
2224
c944940b 2225 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2226 trie->refcount = 1;
3dab1dad 2227 trie->startstate = 1;
786e8c11 2228 trie->wordcount = word_count;
f8fc2ecf 2229 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2230 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2231 if (flags == EXACT || flags == EXACTL)
c944940b 2232 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2233 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2234 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2235
a3621e74 2236 DEBUG_r({
2b8b4781 2237 trie_words = newAV();
a3621e74 2238 });
a3621e74 2239
0111c4fd 2240 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2241 assert(re_trie_maxbuff);
a3621e74 2242 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2243 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2244 }
df826430 2245 DEBUG_TRIE_COMPILE_r({
538e84ed
KW
2246 PerlIO_printf( Perl_debug_log,
2247 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2248 (int)depth * 2 + 2, "",
2249 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2250 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2251 });
538e84ed 2252
7f69552c
YO
2253 /* Find the node we are going to overwrite */
2254 if ( first == startbranch && OP( last ) != BRANCH ) {
2255 /* whole branch chain */
2256 convert = first;
2257 } else {
2258 /* branch sub-chain */
2259 convert = NEXTOPER( first );
2260 }
538e84ed 2261
a3621e74
YO
2262 /* -- First loop and Setup --
2263
2264 We first traverse the branches and scan each word to determine if it
2265 contains widechars, and how many unique chars there are, this is
2266 important as we have to build a table with at least as many columns as we
2267 have unique chars.
2268
2269 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2270 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2271 the native representation of the character value as the key and IV's for
2272 the coded index.
a3621e74
YO
2273
2274 *TODO* If we keep track of how many times each character is used we can
2275 remap the columns so that the table compression later on is more
3b753521 2276 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2277 middle and the least common are on the outside. IMO this would be better
2278 than a most to least common mapping as theres a decent chance the most
2279 common letter will share a node with the least common, meaning the node
486ec47a 2280 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2281 case is when we have the least common nodes twice.
2282
2283 */
2284
a3621e74 2285 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2286 regnode *noper = NEXTOPER( cur );
e1ec3a88 2287 const U8 *uc = (U8*)STRING( noper );
df826430 2288 const U8 *e = uc + STR_LEN( noper );
bc031a7d 2289 int foldlen = 0;
07be1b83 2290 U32 wordlen = 0; /* required init */
bc031a7d
KW
2291 STRLEN minchars = 0;
2292 STRLEN maxchars = 0;
538e84ed
KW
2293 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2294 bitmap?*/
a3621e74 2295
3dab1dad 2296 if (OP(noper) == NOTHING) {
df826430
YO
2297 regnode *noper_next= regnext(noper);
2298 if (noper_next != tail && OP(noper_next) == flags) {
2299 noper = noper_next;
2300 uc= (U8*)STRING(noper);
2301 e= uc + STR_LEN(noper);
2302 trie->minlen= STR_LEN(noper);
2303 } else {
2304 trie->minlen= 0;
2305 continue;
2306 }
3dab1dad 2307 }
df826430 2308
fab2782b 2309 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2310 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2311 regardless of encoding */
fab2782b
YO
2312 if (OP( noper ) == EXACTFU_SS) {
2313 /* false positives are ok, so just set this */
0dc4a61d 2314 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2315 }
2316 }
bc031a7d
KW
2317 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2318 branch */
3dab1dad 2319 TRIE_CHARCOUNT(trie)++;
a3621e74 2320 TRIE_READ_CHAR;
645de4ce 2321
bc031a7d
KW
2322 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2323 * is in effect. Under /i, this character can match itself, or
2324 * anything that folds to it. If not under /i, it can match just
2325 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2326 * all fold to k, and all are single characters. But some folds
2327 * expand to more than one character, so for example LATIN SMALL
2328 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2329 * the string beginning at 'uc' is 'ffi', it could be matched by
2330 * three characters, or just by the one ligature character. (It
2331 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2332 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2333 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2334 * match.) The trie needs to know the minimum and maximum number
2335 * of characters that could match so that it can use size alone to
2336 * quickly reject many match attempts. The max is simple: it is
2337 * the number of folded characters in this branch (since a fold is
2338 * never shorter than what folds to it. */
2339
2340 maxchars++;
2341
2342 /* And the min is equal to the max if not under /i (indicated by
2343 * 'folder' being NULL), or there are no multi-character folds. If
2344 * there is a multi-character fold, the min is incremented just
2345 * once, for the character that folds to the sequence. Each
2346 * character in the sequence needs to be added to the list below of
2347 * characters in the trie, but we count only the first towards the
2348 * min number of characters needed. This is done through the
2349 * variable 'foldlen', which is returned by the macros that look
2350 * for these sequences as the number of bytes the sequence
2351 * occupies. Each time through the loop, we decrement 'foldlen' by
2352 * how many bytes the current char occupies. Only when it reaches
2353 * 0 do we increment 'minchars' or look for another multi-character
2354 * sequence. */
2355 if (folder == NULL) {
2356 minchars++;
2357 }
2358 else if (foldlen > 0) {
2359 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2360 }
2361 else {
bc031a7d
KW
2362 minchars++;
2363
2364 /* See if *uc is the beginning of a multi-character fold. If
2365 * so, we decrement the length remaining to look at, to account
2366 * for the current character this iteration. (We can use 'uc'
2367 * instead of the fold returned by TRIE_READ_CHAR because for
2368 * non-UTF, the latin1_safe macro is smart enough to account
2369 * for all the unfolded characters, and because for UTF, the
2370 * string will already have been folded earlier in the
2371 * compilation process */
2372 if (UTF) {
2373 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2374 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2375 }
2376 }
bc031a7d
KW
2377 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2378 foldlen--;
2379 }
645de4ce 2380 }
bc031a7d
KW
2381
2382 /* The current character (and any potential folds) should be added
2383 * to the possible matching characters for this position in this
2384 * branch */
a3621e74 2385 if ( uvc < 256 ) {
fab2782b
YO
2386 if ( folder ) {
2387 U8 folded= folder[ (U8) uvc ];
2388 if ( !trie->charmap[ folded ] ) {
2389 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2390 TRIE_STORE_REVCHAR( folded );
2391 }
2392 }
a3621e74
YO
2393 if ( !trie->charmap[ uvc ] ) {
2394 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2395 TRIE_STORE_REVCHAR( uvc );
a3621e74 2396 }
02daf0ab 2397 if ( set_bit ) {
62012aee
KW
2398 /* store the codepoint in the bitmap, and its folded
2399 * equivalent. */
fab2782b 2400 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2401
2402 /* store the folded codepoint */
fab2782b 2403 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2404
2405 if ( !UTF ) {
2406 /* store first byte of utf8 representation of
acdf4139 2407 variant codepoints */
6f2d5cbc 2408 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2409 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2410 }
2411 }
02daf0ab
YO
2412 set_bit = 0; /* We've done our bit :-) */
2413 }
a3621e74 2414 } else {
bc031a7d
KW
2415
2416 /* XXX We could come up with the list of code points that fold
2417 * to this using PL_utf8_foldclosures, except not for
2418 * multi-char folds, as there may be multiple combinations
2419 * there that could work, which needs to wait until runtime to
2420 * resolve (The comment about LIGATURE FFI above is such an
2421 * example */
2422
a3621e74 2423 SV** svpp;
55eed653
NC
2424 if ( !widecharmap )
2425 widecharmap = newHV();
a3621e74 2426
55eed653 2427 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2428
2429 if ( !svpp )
e4584336 2430 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2431
2432 if ( !SvTRUE( *svpp ) ) {
2433 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2434 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2435 }
2436 }
bc031a7d
KW
2437 } /* end loop through characters in this branch of the trie */
2438
2439 /* We take the min and max for this branch and combine to find the min
2440 * and max for all branches processed so far */
3dab1dad 2441 if( cur == first ) {
bc031a7d
KW
2442 trie->minlen = minchars;
2443 trie->maxlen = maxchars;
2444 } else if (minchars < trie->minlen) {
2445 trie->minlen = minchars;
2446 } else if (maxchars > trie->maxlen) {
2447 trie->maxlen = maxchars;
fab2782b 2448 }
a3621e74
YO
2449 } /* end first pass */
2450 DEBUG_TRIE_COMPILE_r(
538e84ed
KW
2451 PerlIO_printf( Perl_debug_log,
2452 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3dab1dad 2453 (int)depth * 2 + 2,"",
55eed653 2454 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2455 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2456 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2457 );
a3621e74
YO
2458
2459 /*
2460 We now know what we are dealing with in terms of unique chars and
2461 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2462 representation using a flat table will take. If it's over a reasonable
2463 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2464 conservative but potentially much slower representation using an array
2465 of lists.
2466
2467 At the end we convert both representations into the same compressed
2468 form that will be used in regexec.c for matching with. The latter
2469 is a form that cannot be used to construct with but has memory
2470 properties similar to the list form and access properties similar
2471 to the table form making it both suitable for fast searches and
2472 small enough that its feasable to store for the duration of a program.
2473
2474 See the comment in the code where the compressed table is produced
2475 inplace from the flat tabe representation for an explanation of how
2476 the compression works.
2477
2478 */
2479
2480
2e64971a
DM
2481 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2482 prev_states[1] = 0;
2483
538e84ed
KW
2484 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2485 > SvIV(re_trie_maxbuff) )
2486 {
a3621e74
YO
2487 /*
2488 Second Pass -- Array Of Lists Representation
2489
2490 Each state will be represented by a list of charid:state records
2491 (reg_trie_trans_le) the first such element holds the CUR and LEN
2492 points of the allocated array. (See defines above).
2493
2494 We build the initial structure using the lists, and then convert
2495 it into the compressed table form which allows faster lookups
2496 (but cant be modified once converted).
a3621e74
YO
2497 */
2498
a3621e74
YO
2499 STRLEN transcount = 1;
2500
538e84ed 2501 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2502 "%*sCompiling trie using list compiler\n",
2503 (int)depth * 2 + 2, ""));
686b73d4 2504
c944940b
JH
2505 trie->states = (reg_trie_state *)
2506 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2507 sizeof(reg_trie_state) );
a3621e74
YO
2508 TRIE_LIST_NEW(1);
2509 next_alloc = 2;
2510
2511 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2512
df826430 2513 regnode *noper = NEXTOPER( cur );
c445ea15 2514 U8 *uc = (U8*)STRING( noper );
df826430 2515 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
2516 U32 state = 1; /* required init */
2517 U16 charid = 0; /* sanity init */
07be1b83 2518 U32 wordlen = 0; /* required init */
c445ea15 2519
df826430
YO
2520 if (OP(noper) == NOTHING) {
2521 regnode *noper_next= regnext(noper);
2522 if (noper_next != tail && OP(noper_next) == flags) {
2523 noper = noper_next;
2524 uc= (U8*)STRING(noper);
2525 e= uc + STR_LEN(noper);
2526 }
2527 }
2528
3dab1dad 2529 if (OP(noper) != NOTHING) {
786e8c11 2530 for ( ; uc < e ; uc += len ) {
c445ea15 2531
786e8c11 2532 TRIE_READ_CHAR;
c445ea15 2533
786e8c11
YO
2534 if ( uvc < 256 ) {
2535 charid = trie->charmap[ uvc ];
c445ea15 2536 } else {
538e84ed
KW
2537 SV** const svpp = hv_fetch( widecharmap,
2538 (char*)&uvc,
2539 sizeof( UV ),
2540 0);
786e8c11
YO
2541 if ( !svpp ) {
2542 charid = 0;
2543 } else {
2544 charid=(U16)SvIV( *svpp );
2545 }
c445ea15 2546 }
538e84ed
KW
2547 /* charid is now 0 if we dont know the char read, or
2548 * nonzero if we do */
786e8c11 2549 if ( charid ) {
a3621e74 2550
786e8c11
YO
2551 U16 check;
2552 U32 newstate = 0;
a3621e74 2553
786e8c11
YO
2554 charid--;
2555 if ( !trie->states[ state ].trans.list ) {
2556 TRIE_LIST_NEW( state );
c445ea15 2557 }
538e84ed
KW
2558 for ( check = 1;
2559 check <= TRIE_LIST_USED( state );
2560 check++ )
2561 {
2562 if ( TRIE_LIST_ITEM( state, check ).forid
2563 == charid )
2564 {
786e8c11
YO
2565 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2566 break;
2567 }
2568 }
2569 if ( ! newstate ) {
2570 newstate = next_alloc++;
2e64971a 2571 prev_states[newstate] = state;
786e8c11
YO
2572 TRIE_LIST_PUSH( state, charid, newstate );
2573 transcount++;
2574 }
2575 state = newstate;
2576 } else {
2577 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2578 }
a28509cc 2579 }
c445ea15 2580 }
3dab1dad 2581 TRIE_HANDLE_WORD(state);
a3621e74
YO
2582
2583 } /* end second pass */
2584
1e2e3d02 2585 /* next alloc is the NEXT state to be allocated */
538e84ed 2586 trie->statecount = next_alloc;
c944940b
JH
2587 trie->states = (reg_trie_state *)
2588 PerlMemShared_realloc( trie->states,
2589 next_alloc
2590 * sizeof(reg_trie_state) );
a3621e74 2591
3dab1dad 2592 /* and now dump it out before we compress it */
2b8b4781
NC
2593 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2594 revcharmap, next_alloc,
2595 depth+1)
1e2e3d02 2596 );
a3621e74 2597
c944940b
JH
2598 trie->trans = (reg_trie_trans *)
2599 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2600 {
2601 U32 state;
a3621e74
YO
2602 U32 tp = 0;
2603 U32 zp = 0;
2604
2605
2606 for( state=1 ; state < next_alloc ; state ++ ) {
2607 U32 base=0;
2608
2609 /*
2610 DEBUG_TRIE_COMPILE_MORE_r(
2611 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2612 );
2613 */
2614
2615 if (trie->states[state].trans.list) {
2616 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2617 U16 maxid=minid;
a28509cc 2618 U16 idx;
a3621e74
YO
2619
2620 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2621 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2622 if ( forid < minid ) {
2623 minid=forid;
2624 } else if ( forid > maxid ) {
2625 maxid=forid;
2626 }
a3621e74
YO
2627 }
2628 if ( transcount < tp + maxid - minid + 1) {
2629 transcount *= 2;
c944940b
JH
2630 trie->trans = (reg_trie_trans *)
2631 PerlMemShared_realloc( trie->trans,
446bd890
NC
2632 transcount
2633 * sizeof(reg_trie_trans) );
538e84ed
KW
2634 Zero( trie->trans + (transcount / 2),
2635 transcount / 2,
2636 reg_trie_trans );
a3621e74
YO
2637 }
2638 base = trie->uniquecharcount + tp - minid;
2639 if ( maxid == minid ) {
2640 U32 set = 0;
2641 for ( ; zp < tp ; zp++ ) {
2642 if ( ! trie->trans[ zp ].next ) {
2643 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2644 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2645 1).newstate;
a3621e74
YO
2646 trie->trans[ zp ].check = state;
2647 set = 1;
2648 break;
2649 }
2650 }
2651 if ( !set ) {
538e84ed
KW
2652 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2653 1).newstate;
a3621e74
YO
2654 trie->trans[ tp ].check = state;
2655 tp++;
2656 zp = tp;
2657 }
2658 } else {
2659 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2660 const U32 tid = base
2661 - trie->uniquecharcount
2662 + TRIE_LIST_ITEM( state, idx ).forid;
2663 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2664 idx ).newstate;
a3621e74
YO
2665 trie->trans[ tid ].check = state;
2666 }
2667 tp += ( maxid - minid + 1 );
2668 }
2669 Safefree(trie->states[ state ].trans.list);
2670 }
2671 /*
2672 DEBUG_TRIE_COMPILE_MORE_r(
2673 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2674 );
2675 */
2676 trie->states[ state ].trans.base=base;
2677 }
cc601c31 2678 trie->lasttrans = tp + 1;
a3621e74
YO
2679 }
2680 } else {
2681 /*
2682 Second Pass -- Flat Table Representation.
2683
b423522f
KW
2684 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2685 each. We know that we will need Charcount+1 trans at most to store
2686 the data (one row per char at worst case) So we preallocate both
2687 structures assuming worst case.
a3621e74
YO
2688
2689 We then construct the trie using only the .next slots of the entry
2690 structs.
2691
b423522f
KW
2692 We use the .check field of the first entry of the node temporarily
2693 to make compression both faster and easier by keeping track of how
2694 many non zero fields are in the node.
a3621e74
YO
2695
2696 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2697 transition.
2698
b423522f
KW
2699 There are two terms at use here: state as a TRIE_NODEIDX() which is
2700 a number representing the first entry of the node, and state as a
2701 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2702 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2703 if there are 2 entrys per node. eg:
a3621e74
YO
2704
2705 A B A B
2706 1. 2 4 1. 3 7
2707 2. 0 3 3. 0 5
2708 3. 0 0 5. 0 0
2709 4. 0 0 7. 0 0
2710
b423522f
KW
2711 The table is internally in the right hand, idx form. However as we
2712 also have to deal with the states array which is indexed by nodenum
2713 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2714
2715 */
538e84ed 2716 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2717 "%*sCompiling trie using table compiler\n",
2718 (int)depth * 2 + 2, ""));
3dab1dad 2719
c944940b
JH
2720 trie->trans = (reg_trie_trans *)
2721 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2722 * trie->uniquecharcount + 1,
2723 sizeof(reg_trie_trans) );
2724 trie->states = (reg_trie_state *)
2725 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2726 sizeof(reg_trie_state) );
a3621e74
YO
2727 next_alloc = trie->uniquecharcount + 1;
2728
3dab1dad 2729
a3621e74
YO
2730 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2731
df826430 2732 regnode *noper = NEXTOPER( cur );
a28509cc 2733 const U8 *uc = (U8*)STRING( noper );
df826430 2734 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2735
2736 U32 state = 1; /* required init */
2737
2738 U16 charid = 0; /* sanity init */
2739 U32 accept_state = 0; /* sanity init */
a3621e74 2740
07be1b83 2741 U32 wordlen = 0; /* required init */
a3621e74 2742
df826430
YO
2743 if (OP(noper) == NOTHING) {
2744 regnode *noper_next= regnext(noper);
2745 if (noper_next != tail && OP(noper_next) == flags) {
2746 noper = noper_next;
2747 uc= (U8*)STRING(noper);
2748 e= uc + STR_LEN(noper);
2749 }
2750 }
fab2782b 2751
3dab1dad 2752 if ( OP(noper) != NOTHING ) {
786e8c11 2753 for ( ; uc < e ; uc += len ) {
a3621e74 2754
786e8c11 2755 TRIE_READ_CHAR;
a3621e74 2756
786e8c11
YO
2757 if ( uvc < 256 ) {
2758 charid = trie->charmap[ uvc ];
2759 } else {
538e84ed
KW
2760 SV* const * const svpp = hv_fetch( widecharmap,
2761 (char*)&uvc,
2762 sizeof( UV ),
2763 0);
786e8c11 2764 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2765 }
786e8c11
YO
2766 if ( charid ) {
2767 charid--;
2768 if ( !trie->trans[ state + charid ].next ) {
2769 trie->trans[ state + charid ].next = next_alloc;
2770 trie->trans[ state ].check++;
2e64971a
DM
2771 prev_states[TRIE_NODENUM(next_alloc)]
2772 = TRIE_NODENUM(state);
786e8c11
YO
2773 next_alloc += trie->uniquecharcount;
2774 }
2775 state = trie->trans[ state + charid ].next;
2776 } else {
2777 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2778 }
538e84ed
KW
2779 /* charid is now 0 if we dont know the char read, or
2780 * nonzero if we do */
a3621e74 2781 }
a3621e74 2782 }
3dab1dad
YO
2783 accept_state = TRIE_NODENUM( state );
2784 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2785
2786 } /* end second pass */
2787
3dab1dad 2788 /* and now dump it out before we compress it */
2b8b4781
NC
2789 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2790 revcharmap,
2791 next_alloc, depth+1));
a3621e74 2792
a3621e74
YO
2793 {
2794 /*
2795 * Inplace compress the table.*
2796
2797 For sparse data sets the table constructed by the trie algorithm will
2798 be mostly 0/FAIL transitions or to put it another way mostly empty.
2799 (Note that leaf nodes will not contain any transitions.)
2800
2801 This algorithm compresses the tables by eliminating most such
2802 transitions, at the cost of a modest bit of extra work during lookup:
2803
2804 - Each states[] entry contains a .base field which indicates the
2805 index in the state[] array wheres its transition data is stored.
2806
3b753521 2807 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2808
2809 - If .base is nonzero then charid is added to it to find an entry in
2810 the trans array.
2811
2812 -If trans[states[state].base+charid].check!=state then the
2813 transition is taken to be a 0/Fail transition. Thus if there are fail
2814 transitions at the front of the node then the .base offset will point
2815 somewhere inside the previous nodes data (or maybe even into a node
2816 even earlier), but the .check field determines if the transition is
2817 valid.
2818
786e8c11 2819 XXX - wrong maybe?
a3621e74 2820 The following process inplace converts the table to the compressed
3b753521 2821 table: We first do not compress the root node 1,and mark all its
a3621e74 2822 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2823 allows us to do a DFA construction from the compressed table later,
2824 and ensures that any .base pointers we calculate later are greater
2825 than 0.
a3621e74
YO
2826
2827 - We set 'pos' to indicate the first entry of the second node.
2828
2829 - We then iterate over the columns of the node, finding the first and
2830 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2831 and set the .check pointers accordingly, and advance pos
2832 appropriately and repreat for the next node. Note that when we copy
2833 the next pointers we have to convert them from the original
2834 NODEIDX form to NODENUM form as the former is not valid post
2835 compression.
2836
2837 - If a node has no transitions used we mark its base as 0 and do not
2838 advance the pos pointer.
2839
2840 - If a node only has one transition we use a second pointer into the
2841 structure to fill in allocated fail transitions from other states.
2842 This pointer is independent of the main pointer and scans forward
2843 looking for null transitions that are allocated to a state. When it
2844 finds one it writes the single transition into the "hole". If the
786e8c11 2845 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2846
2847 - Once compressed we can Renew/realloc the structures to release the
2848 excess space.
2849
2850 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2851 specifically Fig 3.47 and the associated pseudocode.
2852
2853 demq
2854 */
a3b680e6 2855 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2856 U32 state, charid;
a3621e74 2857 U32 pos = 0, zp=0;
1e2e3d02 2858 trie->statecount = laststate;
a3621e74
YO
2859
2860 for ( state = 1 ; state < laststate ; state++ ) {
2861 U8 flag = 0;
a28509cc
AL
2862 const U32 stateidx = TRIE_NODEIDX( state );
2863 const U32 o_used = trie->trans[ stateidx ].check;
2864 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2865 trie->trans[ stateidx ].check = 0;
2866
538e84ed
KW
2867 for ( charid = 0;
2868 used && charid < trie->uniquecharcount;
2869 charid++ )
2870 {
a3621e74
YO
2871 if ( flag || trie->trans[ stateidx + charid ].next ) {
2872 if ( trie->trans[ stateidx + charid ].next ) {
2873 if (o_used == 1) {
2874 for ( ; zp < pos ; zp++ ) {
2875 if ( ! trie->trans[ zp ].next ) {
2876 break;
2877 }
2878 }
538e84ed
KW
2879 trie->states[ state ].trans.base
2880 = zp
2881 + trie->uniquecharcount
2882 - charid ;
2883 trie->trans[ zp ].next
2884 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2885 + charid ].next );
a3621e74
YO
2886 trie->trans[ zp ].check = state;
2887 if ( ++zp > pos ) pos = zp;
2888 break;
2889 }
2890 used--;
2891 }
2892 if ( !flag ) {
2893 flag = 1;
538e84ed
KW
2894 trie->states[ state ].trans.base
2895 = pos + trie->uniquecharcount - charid ;
a3621e74 2896 }
538e84ed
KW
2897 trie->trans[ pos ].next
2898 = SAFE_TRIE_NODENUM(
2899 trie->trans[ stateidx + charid ].next );
a3621e74
YO
2900 trie->trans[ pos ].check = state;
2901 pos++;
2902 }
2903 }
2904 }
cc601c31 2905 trie->lasttrans = pos + 1;
c944940b
JH
2906 trie->states = (reg_trie_state *)
2907 PerlMemShared_realloc( trie->states, laststate
2908 * sizeof(reg_trie_state) );
a3621e74 2909 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2910 PerlIO_printf( Perl_debug_log,
2911 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2912 (int)depth * 2 + 2,"",
2913 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2914 + 1 ),
2915 (IV)next_alloc,
2916 (IV)pos,
2917 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
2918 );
2919
2920 } /* end table compress */
2921 }
1e2e3d02 2922 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2923 PerlIO_printf(Perl_debug_log,
2924 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1e2e3d02
YO
2925 (int)depth * 2 + 2, "",
2926 (UV)trie->statecount,
2927 (UV)trie->lasttrans)
2928 );
cc601c31 2929 /* resize the trans array to remove unused space */
c944940b
JH
2930 trie->trans = (reg_trie_trans *)
2931 PerlMemShared_realloc( trie->trans, trie->lasttrans
2932 * sizeof(reg_trie_trans) );
a3621e74 2933
538e84ed 2934 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2935 U8 nodetype =(U8)(flags & 0xFF);
2936 char *str=NULL;
538e84ed 2937
07be1b83 2938#ifdef DEBUGGING
e62cc96a 2939 regnode *optimize = NULL;
7122b237
YO
2940#ifdef RE_TRACK_PATTERN_OFFSETS
2941
b57a0404
JH
2942 U32 mjd_offset = 0;
2943 U32 mjd_nodelen = 0;
7122b237
YO
2944#endif /* RE_TRACK_PATTERN_OFFSETS */
2945#endif /* DEBUGGING */
a3621e74 2946 /*
3dab1dad
YO
2947 This means we convert either the first branch or the first Exact,
2948 depending on whether the thing following (in 'last') is a branch
2949 or not and whther first is the startbranch (ie is it a sub part of
2950 the alternation or is it the whole thing.)
3b753521 2951 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2952 the whole branch sequence, including the first.
a3621e74 2953 */
3dab1dad 2954 /* Find the node we are going to overwrite */
7f69552c 2955 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2956 /* branch sub-chain */
3dab1dad 2957 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2958#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2959 DEBUG_r({
2960 mjd_offset= Node_Offset((convert));
2961 mjd_nodelen= Node_Length((convert));
2962 });
7122b237 2963#endif
7f69552c 2964 /* whole branch chain */
7122b237
YO
2965 }
2966#ifdef RE_TRACK_PATTERN_OFFSETS
2967 else {
7f69552c
YO
2968 DEBUG_r({
2969 const regnode *nop = NEXTOPER( convert );
2970 mjd_offset= Node_Offset((nop));
2971 mjd_nodelen= Node_Length((nop));
2972 });
07be1b83
YO
2973 }
2974 DEBUG_OPTIMISE_r(
538e84ed
KW
2975 PerlIO_printf(Perl_debug_log,
2976 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
07be1b83 2977 (int)depth * 2 + 2, "",
786e8c11 2978 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2979 );
7122b237 2980#endif
538e84ed 2981 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
2982 split out as an EXACT and put in front of the TRIE node. */
2983 trie->startstate= 1;
55eed653 2984 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2985 U32 state;
1e2e3d02 2986 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2987 U32 ofs = 0;
8e11feef
RGS
2988 I32 idx = -1;
2989 U32 count = 0;
2990 const U32 base = trie->states[ state ].trans.base;
a3621e74 2991
3dab1dad 2992 if ( trie->states[state].wordnum )
8e11feef 2993 count = 1;
a3621e74 2994
8e11feef 2995 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2996 if ( ( base + ofs >= trie->uniquecharcount ) &&
2997 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2998 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2999 {
3dab1dad 3000 if ( ++count > 1 ) {
2b8b4781 3001 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 3002 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3003 if ( state == 1 ) break;
3dab1dad
YO
3004 if ( count == 2 ) {
3005 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3006 DEBUG_OPTIMISE_r(
8e11feef
RGS
3007 PerlIO_printf(Perl_debug_log,
3008 "%*sNew Start State=%"UVuf" Class: [",
3009 (int)depth * 2 + 2, "",
786e8c11 3010 (UV)state));
be8e71aa 3011 if (idx >= 0) {
2b8b4781 3012 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 3013 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3014
3dab1dad 3015 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
3016 if ( folder )
3017 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 3018 DEBUG_OPTIMISE_r(
f1f66076 3019 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 3020 );
8e11feef
RGS
3021 }
3022 }
3023 TRIE_BITMAP_SET(trie,*ch);
3024 if ( folder )
3025 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3026 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3027 }
3028 idx = ofs;
3029 }
3dab1dad
YO
3030 }
3031 if ( count == 1 ) {
2b8b4781 3032 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
3033 STRLEN len;
3034 char *ch = SvPV( *tmp, len );
de734bd5
A
3035 DEBUG_OPTIMISE_r({
3036 SV *sv=sv_newmortal();
8e11feef
RGS
3037 PerlIO_printf( Perl_debug_log,
3038 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3039 (int)depth * 2 + 2, "",
538e84ed
KW
3040 (UV)state, (UV)idx,
3041 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3042 PL_colors[0], PL_colors[1],
3043 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3044 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3045 )
3046 );
3047 });
3dab1dad
YO
3048 if ( state==1 ) {
3049 OP( convert ) = nodetype;
3050 str=STRING(convert);
3051 STR_LEN(convert)=0;
3052 }
c490c714
YO
3053 STR_LEN(convert) += len;
3054 while (len--)
de734bd5 3055 *str++ = *ch++;
8e11feef 3056 } else {
538e84ed 3057#ifdef DEBUGGING
8e11feef
RGS
3058 if (state>1)
3059 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 3060#endif
8e11feef
RGS
3061 break;
3062 }
3063 }
2e64971a 3064 trie->prefixlen = (state-1);
3dab1dad 3065 if (str) {
8e11feef 3066 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3067 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3068 trie->startstate = state;
07be1b83
YO
3069 trie->minlen -= (state - 1);
3070 trie->maxlen -= (state - 1);
33809eae
JH
3071#ifdef DEBUGGING
3072 /* At least the UNICOS C compiler choked on this
3073 * being argument to DEBUG_r(), so let's just have
3074 * it right here. */
3075 if (
3076#ifdef PERL_EXT_RE_BUILD
3077 1
3078#else
3079 DEBUG_r_TEST
3080#endif
3081 ) {
3082 regnode *fix = convert;
3083 U32 word = trie->wordcount;
3084 mjd_nodelen++;
3085 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3086 while( ++fix < n ) {
3087 Set_Node_Offset_Length(fix, 0, 0);
3088 }
3089 while (word--) {
3090 SV ** const tmp = av_fetch( trie_words, word, 0 );
3091 if (tmp) {
3092 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3093 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3094 else
3095 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3096 }
3097 }
3098 }
3099#endif
8e11feef
RGS
3100 if (trie->maxlen) {
3101 convert = n;
3102 } else {
3dab1dad 3103 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3104 DEBUG_r(optimize= n);
3dab1dad
YO
3105 }
3106 }
3107 }
538e84ed
KW
3108 if (!jumper)
3109 jumper = last;
3dab1dad 3110 if ( trie->maxlen ) {
8e11feef
RGS
3111 NEXT_OFF( convert ) = (U16)(tail - convert);
3112 ARG_SET( convert, data_slot );
538e84ed
KW
3113 /* Store the offset to the first unabsorbed branch in
3114 jump[0], which is otherwise unused by the jump logic.
786e8c11 3115 We use this when dumping a trie and during optimisation. */
538e84ed 3116 if (trie->jump)
7f69552c 3117 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3118
6c48061a
YO
3119 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3120 * and there is a bitmap
3121 * and the first "jump target" node we found leaves enough room
3122 * then convert the TRIE node into a TRIEC node, with the bitmap
3123 * embedded inline in the opcode - this is hypothetically faster.
3124 */
3125 if ( !trie->states[trie->startstate].wordnum
3126 && trie->bitmap
3127 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3128 {
3129 OP( convert ) = TRIEC;
3130 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3131 PerlMemShared_free(trie->bitmap);
786e8c11 3132 trie->bitmap= NULL;
538e84ed 3133 } else
786e8c11 3134 OP( convert ) = TRIE;
a3621e74 3135
3dab1dad
YO
3136 /* store the type in the flags */
3137 convert->flags = nodetype;
a5ca303d 3138 DEBUG_r({
538e84ed
KW
3139 optimize = convert
3140 + NODE_STEP_REGNODE
a5ca303d
YO
3141 + regarglen[ OP( convert ) ];
3142 });
538e84ed 3143 /* XXX We really should free up the resource in trie now,
a5ca303d 3144 as we won't use them - (which resources?) dmq */
3dab1dad 3145 }
a3621e74 3146 /* needed for dumping*/
e62cc96a 3147 DEBUG_r(if (optimize) {
07be1b83 3148 regnode *opt = convert;
bcdf7404 3149
e62cc96a 3150 while ( ++opt < optimize) {
07be1b83
YO
3151 Set_Node_Offset_Length(opt,0,0);
3152 }
538e84ed
KW
3153 /*
3154 Try to clean up some of the debris left after the
786e8c11 3155 optimisation.
a3621e74 3156 */
786e8c11 3157 while( optimize < jumper ) {
07be1b83 3158 mjd_nodelen += Node_Length((optimize));
a3621e74 3159 OP( optimize ) = OPTIMIZED;
07be1b83 3160 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3161 optimize++;
3162 }
07be1b83 3163 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3164 });
3165 } /* end node insert */
2e64971a
DM
3166
3167 /* Finish populating the prev field of the wordinfo array. Walk back
3168 * from each accept state until we find another accept state, and if
3169 * so, point the first word's .prev field at the second word. If the
3170 * second already has a .prev field set, stop now. This will be the
3171 * case either if we've already processed that word's accept state,
3b753521
FN
3172 * or that state had multiple words, and the overspill words were
3173 * already linked up earlier.
2e64971a
DM
3174 */
3175 {
3176 U16 word;
3177 U32 state;
3178 U16 prev;
3179
3180 for (word=1; word <= trie->wordcount; word++) {
3181 prev = 0;
3182 if (trie->wordinfo[word].prev)
3183 continue;
3184 state = trie->wordinfo[word].accept;
3185 while (state) {
3186 state = prev_states[state];
3187 if (!state)
3188 break;
3189 prev = trie->states[state].wordnum;
3190 if (prev)
3191 break;
3192 }
3193 trie->wordinfo[word].prev = prev;
3194 }
3195 Safefree(prev_states);
3196 }
3197
3198
3199 /* and now dump out the compressed format */
3200 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3201
55eed653 3202 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
3203#ifdef DEBUGGING
3204 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3205 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3206#else
03e70be4 3207 SvREFCNT_dec_NN(revcharmap);
07be1b83 3208#endif
538e84ed
KW
3209 return trie->jump
3210 ? MADE_JUMP_TRIE
3211 : trie->startstate>1
3212 ? MADE_EXACT_TRIE
786e8c11
YO
3213 : MADE_TRIE;
3214}
3215
615a2e7f
YO
3216STATIC regnode *
3217S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
786e8c11 3218{
b423522f
KW
3219/* The Trie is constructed and compressed now so we can build a fail array if
3220 * it's needed
786e8c11 3221
b423522f
KW
3222 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3223 3.32 in the
3224 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3225 Ullman 1985/88
786e8c11
YO
3226 ISBN 0-201-10088-6
3227
b423522f
KW
3228 We find the fail state for each state in the trie, this state is the longest
3229 proper suffix of the current state's 'word' that is also a proper prefix of
3230 another word in our trie. State 1 represents the word '' and is thus the
3231 default fail state. This allows the DFA not to have to restart after its
3232 tried and failed a word at a given point, it simply continues as though it
3233 had been matching the other word in the first place.
786e8c11
YO
3234 Consider
3235 'abcdgu'=~/abcdefg|cdgu/
b423522f
KW
3236 When we get to 'd' we are still matching the first word, we would encounter
3237 'g' which would fail, which would bring us to the state representing 'd' in
3238 the second word where we would try 'g' and succeed, proceeding to match
3239 'cdgu'.
786e8c11
YO
3240 */
3241 /* add a fail transition */
3251b653
NC
3242 const U32 trie_offset = ARG(source);
3243 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
3244 U32 *q;
3245 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 3246 const U32 numstates = trie->statecount;
786e8c11
YO
3247 const U32 ubound = trie->lasttrans + ucharcount;
3248 U32 q_read = 0;
3249 U32 q_write = 0;
3250 U32 charid;
3251 U32 base = trie->states[ 1 ].trans.base;
3252 U32 *fail;
3253 reg_ac_data *aho;
cf78de0b 3254 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
615a2e7f 3255 regnode *stclass;
786e8c11 3256 GET_RE_DEBUG_FLAGS_DECL;
7918f24d 3257
615a2e7f 3258 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
81611534 3259 PERL_UNUSED_CONTEXT;
786e8c11
YO
3260#ifndef DEBUGGING
3261 PERL_UNUSED_ARG(depth);
3262#endif
3263
615a2e7f
YO
3264 if ( OP(source) == TRIE ) {
3265 struct regnode_1 *op = (struct regnode_1 *)
3266 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3267 StructCopy(source,op,struct regnode_1);
3268 stclass = (regnode *)op;
3269 } else {
3270 struct regnode_charclass *op = (struct regnode_charclass *)
3271 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3272 StructCopy(source,op,struct regnode_charclass);
3273 stclass = (regnode *)op;
3274 }
2f306ab9 3275 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
786e8c11
YO
3276
3277 ARG_SET( stclass, data_slot );
c944940b 3278 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 3279 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 3280 aho->trie=trie_offset;
446bd890
NC
3281 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3282 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 3283 Newxz( q, numstates, U32);
c944940b 3284 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
3285 aho->refcount = 1;
3286 fail = aho->fail;
3287 /* initialize fail[0..1] to be 1 so that we always have
3288 a valid final fail state */
3289 fail[ 0 ] = fail[ 1 ] = 1;
3290
3291 for ( charid = 0; charid < ucharcount ; charid++ ) {
3292 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3293 if ( newstate ) {
3294 q[ q_write ] = newstate;
3295 /* set to point at the root */
3296 fail[ q[ q_write++ ] ]=1;
3297 }
3298 }
3299 while ( q_read < q_write) {
3300 const U32 cur = q[ q_read++ % numstates ];
3301 base = trie->states[ cur ].trans.base;
3302
3303 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3304 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3305 if (ch_state) {
3306 U32 fail_state = cur;
3307 U32 fail_base;
3308 do {
3309 fail_state = fail[ fail_state ];
3310 fail_base = aho->states[ fail_state ].trans.base;
3311 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3312
3313 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3314 fail[ ch_state ] = fail_state;
3315 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3316 {
3317 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3318 }
3319 q[ q_write++ % numstates] = ch_state;
3320 }
3321 }
3322 }
3323 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3324 when we fail in state 1, this allows us to use the
3325 charclass scan to find a valid start char. This is based on the principle
3326 that theres a good chance the string being searched contains lots of stuff
3327 that cant be a start char.
3328 */
3329 fail[ 0 ] = fail[ 1 ] = 0;
3330 DEBUG_TRIE_COMPILE_r({
6d99fb9b 3331 PerlIO_printf(Perl_debug_log,
538e84ed 3332 "%*sStclass Failtable (%"UVuf" states): 0",
6d99fb9b 3333 (int)(depth * 2), "", (UV)numstates
1e2e3d02 3334 );
786e8c11
YO
3335 for( q_read=1; q_read<numstates; q_read++ ) {
3336 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3337 }
3338 PerlIO_printf(Perl_debug_log, "\n");
3339 });
3340 Safefree(q);
e384d5c1 3341 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
615a2e7f 3342 return stclass;
a3621e74
YO
3343}
3344
786e8c11 3345
07be1b83 3346#define DEBUG_PEEP(str,scan,depth) \
b515a41d 3347 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83 3348 regnode *Next = regnext(scan); \
8b9781c9 3349 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
fdfb4f21 3350 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
c9f0d54c
YO
3351 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3352 Next ? (REG_NODE_NUM(Next)) : 0 ); \
fdfb4f21
YO
3353 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3354 PerlIO_printf(Perl_debug_log, "\n"); \
b515a41d 3355 }});
07be1b83 3356
bb914485 3357/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
3358 * one. The regop may be changed if the node(s) contain certain sequences that
3359 * require special handling. The joining is only done if:
bb914485
KW
3360 * 1) there is room in the current conglomerated node to entirely contain the
3361 * next one.
3362 * 2) they are the exact same node type
3363 *
87b8b349 3364 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
3365 * these get optimized out
3366 *
0a982f06
KW
3367 * If a node is to match under /i (folded), the number of characters it matches
3368 * can be different than its character length if it contains a multi-character
31f05a37
KW
3369 * fold. *min_subtract is set to the total delta number of characters of the
3370 * input nodes.
bb914485 3371 *
fb29cc72 3372 * And *unfolded_multi_char is set to indicate whether or not the node contains
31f05a37
KW
3373 * an unfolded multi-char fold. This happens when whether the fold is valid or
3374 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3375 * SMALL LETTER SHARP S, as only if the target string being matched against
3376 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3377 * folding rules depend on the locale in force at runtime. (Multi-char folds
3378 * whose components are all above the Latin1 range are not run-time locale
3379 * dependent, and have already been folded by the time this function is
3380 * called.)
f758bddf 3381 *
bb914485 3382 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
3383 * multi-character fold sequences. It's been wrong in Perl for a very long
3384 * time. There are three code points in Unicode whose multi-character folds
3385 * were long ago discovered to mess things up. The previous designs for
3386 * dealing with these involved assigning a special node for them. This
538e84ed 3387 * approach doesn't always work, as evidenced by this example:
a0c4c608 3388 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
538e84ed 3389 * Both sides fold to "sss", but if the pattern is parsed to create a node that
0a982f06 3390 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
3391 * successful match would have to cross the node's boundary. The new approach
3392 * that hopefully generally solves the problem generates an EXACTFU_SS node
538e84ed 3393 * that is "sss" in this case.
bb914485 3394 *
0a982f06 3395 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
3396 * just these three. Now the code is general, for all such cases. The
3397 * approach taken is:
0a982f06 3398 * 1) This routine examines each EXACTFish node that could contain multi-
538e84ed
KW
3399 * character folded sequences. Since a single character can fold into
3400 * such a sequence, the minimum match length for this node is less than
3401 * the number of characters in the node. This routine returns in
31f05a37
KW
3402 * *min_subtract how many characters to subtract from the the actual
3403 * length of the string to get a real minimum match length; it is 0 if
3404 * there are no multi-char foldeds. This delta is used by the caller to
3405 * adjust the min length of the match, and the delta between min and max,
3406 * so that the optimizer doesn't reject these possibilities based on size
3407 * constraints.
cb117658 3408 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
3409 * is used for an EXACTFU node that contains at least one "ss" sequence in
3410 * it. For non-UTF-8 patterns and strings, this is the only case where
3411 * there is a possible fold length change. That means that a regular
3412 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3413 * with length changes, and so can be processed faster. regexec.c takes
3414 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
31f05a37
KW
3415 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3416 * known until runtime). This saves effort in regex matching. However,
3417 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3418 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3419 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3420 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
0a982f06
KW
3421 * possibilities for the non-UTF8 patterns are quite simple, except for
3422 * the sharp s. All the ones that don't involve a UTF-8 target string are
3423 * members of a fold-pair, and arrays are set up for all of them so that
3424 * the other member of the pair can be found quickly. Code elsewhere in
3425 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3426 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3427 * described in the next item.
31f05a37
KW
3428 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3429 * validity of the fold won't be known until runtime, and so must remain
3430 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3431 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3432 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3433 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3434 * The reason this is a problem is that the optimizer part of regexec.c
3435 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3436 * that a character in the pattern corresponds to at most a single
3437 * character in the target string. (And I do mean character, and not byte
3438 * here, unlike other parts of the documentation that have never been
3439 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3440 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3441 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3442 * nodes, violate the assumption, and they are the only instances where it
3443 * is violated. I'm reluctant to try to change the assumption, as the
3444 * code involved is impenetrable to me (khw), so instead the code here
3445 * punts. This routine examines EXACTFL nodes, and (when the pattern
3446 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3447 * boolean indicating whether or not the node contains such a fold. When
3448 * it is true, the caller sets a flag that later causes the optimizer in
3449 * this file to not set values for the floating and fixed string lengths,
3450 * and thus avoids the optimizer code in regexec.c that makes the invalid
1ca267a5 3451 * assumption. Thus, there is no optimization based on string lengths for
31f05a37
KW
3452 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3453 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3454 * assumption is wrong only in these cases is that all other non-UTF-8
3455 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3456 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3457 * EXACTF nodes because we don't know at compile time if it actually
3458 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3459 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3460 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3461 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3462 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3463 * string would require the pattern to be forced into UTF-8, the overhead
3464 * of which we want to avoid. Similarly the unfolded multi-char folds in
3465 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3466 * locale.)
098b07d5
KW
3467 *
3468 * Similarly, the code that generates tries doesn't currently handle
3469 * not-already-folded multi-char folds, and it looks like a pain to change
3470 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3471 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3472 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3473 * using /iaa matching will be doing so almost entirely with ASCII
3474 * strings, so this should rarely be encountered in practice */
1de06328 3475
fb29cc72 3476#define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \