This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Refactor func so caller handles anomalies
[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
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
538e84ed
KW
94#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
95 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
96#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
97 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 98#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 99#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 100
a687059c
LW
101#ifndef STATIC
102#define STATIC static
103#endif
104
b1603ef8 105
09b2b2e6 106struct RExC_state_t {
514a91f1
DM
107 U32 flags; /* RXf_* are we folding, multilining? */
108 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 109 char *precomp; /* uncompiled string. */
288b8c02 110 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 111 regexp *rx; /* perl core regexp structure */
538e84ed
KW
112 regexp_internal *rxi; /* internal data for regexp object
113 pprivate field */
fac92740 114 char *start; /* Start of input for compile */
830247a4
IZ
115 char *end; /* End of input for compile */
116 char *parse; /* Input-scan pointer. */
ea3daa5d 117 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 118 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
119 regnode *emit_bound; /* First regnode outside of the
120 allocated space */
f7c7e32a
DM
121 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
122 implies compiling, so don't emit */
9a81a976
KW
123 regnode_ssc emit_dummy; /* placeholder for emit to point to;
124 large enough for the largest
125 non-EXACTish node, so can use it as
126 scratch in pass1 */
830247a4
IZ
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
129 U32 seen;
ea3daa5d 130 SSize_t size; /* Code size. */
538e84ed
KW
131 I32 npar; /* Capture buffer count, (OPEN) plus
132 one. ("par" 0 is the whole
133 pattern)*/
134 I32 nestroot; /* root parens we are in - used by
135 accept */
830247a4
IZ
136 I32 extralen;
137 I32 seen_zerolen;
40d049e4
YO
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
02daf0ab
YO
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
e40e74fe
KW
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
147 utf8 */
81714fb9 148 HV *paren_names; /* Paren names */
538e84ed 149
40d049e4
YO
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
538e84ed
KW
152 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
153 through */
09a65838 154 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 155 I32 in_lookbehind;
4624b182 156 I32 contains_locale;
cfafade5 157 I32 contains_i;
bb3f3ed2 158 I32 override_recoding;
9d53c457 159 I32 in_multi_char_class;
3d2bd50a 160 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 161 within pattern */
b1603ef8
DM
162 int num_code_blocks; /* size of code_blocks[] */
163 int code_index; /* next code_blocks[] slot */
ee273784 164 SSize_t maxlen; /* mininum possible number of chars in string to match */
dc6d7f5c 165#ifdef ADD_TO_REGEXEC
830247a4
IZ
166 char *starttry; /* -Dr: where regtry was called. */
167#define RExC_starttry (pRExC_state->starttry)
168#endif
d24ca0c5 169 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 170#ifdef DEBUGGING
be8e71aa 171 const char *lastparse;
3dab1dad 172 I32 lastnum;
1f1031fe 173 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
174#define RExC_lastparse (pRExC_state->lastparse)
175#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 176#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 177#endif
09b2b2e6 178};
830247a4 179
e2509266 180#define RExC_flags (pRExC_state->flags)
514a91f1 181#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 182#define RExC_precomp (pRExC_state->precomp)
288b8c02 183#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 184#define RExC_rx (pRExC_state->rx)
f8fc2ecf 185#define RExC_rxi (pRExC_state->rxi)
fac92740 186#define RExC_start (pRExC_state->start)
830247a4
IZ
187#define RExC_end (pRExC_state->end)
188#define RExC_parse (pRExC_state->parse)
189#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237 190#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
191#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
192 others */
7122b237 193#endif
830247a4 194#define RExC_emit (pRExC_state->emit)
f7c7e32a 195#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 196#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 197#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
198#define RExC_naughty (pRExC_state->naughty)
199#define RExC_sawback (pRExC_state->sawback)
200#define RExC_seen (pRExC_state->seen)
201#define RExC_size (pRExC_state->size)
ee273784 202#define RExC_maxlen (pRExC_state->maxlen)
830247a4 203#define RExC_npar (pRExC_state->npar)
e2e6a0f1 204#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
205#define RExC_extralen (pRExC_state->extralen)
206#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 207#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 208#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 209#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
210#define RExC_open_parens (pRExC_state->open_parens)
211#define RExC_close_parens (pRExC_state->close_parens)
212#define RExC_opend (pRExC_state->opend)
81714fb9 213#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
214#define RExC_recurse (pRExC_state->recurse)
215#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 216#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
217#define RExC_study_chunk_recursed_bytes \
218 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 219#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 220#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 221#define RExC_contains_i (pRExC_state->contains_i)
9d53c457
KW
222#define RExC_override_recoding (pRExC_state->override_recoding)
223#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 224
cde0cee5 225
a687059c
LW
226#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
227#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 228 ((*s) == '{' && regcurly(s)))
a687059c
LW
229
230/*
231 * Flags to be passed up and down.
232 */
a687059c 233#define WORST 0 /* Worst case. */
a3b492c3 234#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 235
e64f369d 236/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
237 * character. (There needs to be a case: in the switch statement in regexec.c
238 * for any node marked SIMPLE.) Note that this is not the same thing as
239 * REGNODE_SIMPLE */
fda99bee 240#define SIMPLE 0x02
e64f369d 241#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
242#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
243#define TRYAGAIN 0x10 /* Weeded out a declaration. */
244#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 245
3dab1dad
YO
246#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
247
07be1b83
YO
248/* whether trie related optimizations are enabled */
249#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
250#define TRIE_STUDY_OPT
786e8c11 251#define FULL_TRIE_STUDY
07be1b83
YO
252#define TRIE_STCLASS
253#endif
1de06328
YO
254
255
40d049e4
YO
256
257#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
258#define PBITVAL(paren) (1 << ((paren) & 7))
259#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
260#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
261#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
262
bbd61b5f 263#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
264 if (!UTF) { \
265 *flagp = RESTART_UTF8; \
266 return NULL; \
267 } \
bbd61b5f 268 } STMT_END
40d049e4 269
f19b1a63
KW
270/* This converts the named class defined in regcomp.h to its equivalent class
271 * number defined in handy.h. */
272#define namedclass_to_classnum(class) ((int) ((class) / 2))
273#define classnum_to_namedclass(classnum) ((classnum) * 2)
274
de92f5e6
KW
275#define _invlist_union_complement_2nd(a, b, output) \
276 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
277#define _invlist_intersection_complement_2nd(a, b, output) \
278 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
279
1de06328
YO
280/* About scan_data_t.
281
282 During optimisation we recurse through the regexp program performing
283 various inplace (keyhole style) optimisations. In addition study_chunk
284 and scan_commit populate this data structure with information about
538e84ed 285 what strings MUST appear in the pattern. We look for the longest
3b753521 286 string that must appear at a fixed location, and we look for the
1de06328
YO
287 longest string that may appear at a floating location. So for instance
288 in the pattern:
538e84ed 289
1de06328 290 /FOO[xX]A.*B[xX]BAR/
538e84ed 291
1de06328
YO
292 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
293 strings (because they follow a .* construct). study_chunk will identify
294 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 295
1de06328 296 The strings can be composites, for instance
538e84ed 297
1de06328 298 /(f)(o)(o)/
538e84ed 299
1de06328 300 will result in a composite fixed substring 'foo'.
538e84ed 301
1de06328 302 For each string some basic information is maintained:
538e84ed 303
1de06328
YO
304 - offset or min_offset
305 This is the position the string must appear at, or not before.
306 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
307 characters must match before the string we are searching for.
308 Likewise when combined with minlenp and the length of the string it
538e84ed 309 tells us how many characters must appear after the string we have
1de06328 310 found.
538e84ed 311
1de06328
YO
312 - max_offset
313 Only used for floating strings. This is the rightmost point that
ea3daa5d 314 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 315 string can occur infinitely far to the right.
538e84ed 316
1de06328 317 - minlenp
2d608413
KW
318 A pointer to the minimum number of characters of the pattern that the
319 string was found inside. This is important as in the case of positive
538e84ed 320 lookahead or positive lookbehind we can have multiple patterns
1de06328 321 involved. Consider
538e84ed 322
1de06328 323 /(?=FOO).*F/
538e84ed 324
1de06328
YO
325 The minimum length of the pattern overall is 3, the minimum length
326 of the lookahead part is 3, but the minimum length of the part that
538e84ed 327 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 328 minimum length for the F is 1. This is important as the minimum length
538e84ed 329 is used to determine offsets in front of and behind the string being
1de06328 330 looked for. Since strings can be composites this is the length of the
486ec47a 331 pattern at the time it was committed with a scan_commit. Note that
1de06328 332 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 333 are not known until the full pattern has been compiled, thus the
1de06328 334 pointer to the value.
538e84ed 335
1de06328 336 - lookbehind
538e84ed 337
1de06328 338 In the case of lookbehind the string being searched for can be
538e84ed 339 offset past the start point of the final matching string.
1de06328
YO
340 If this value was just blithely removed from the min_offset it would
341 invalidate some of the calculations for how many chars must match
342 before or after (as they are derived from min_offset and minlen and
538e84ed 343 the length of the string being searched for).
1de06328
YO
344 When the final pattern is compiled and the data is moved from the
345 scan_data_t structure into the regexp structure the information
538e84ed
KW
346 about lookbehind is factored in, with the information that would
347 have been lost precalculated in the end_shift field for the
1de06328
YO
348 associated string.
349
350 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 351 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
352
353*/
2c2d71f5
JH
354
355typedef struct scan_data_t {
1de06328
YO
356 /*I32 len_min; unused */
357 /*I32 len_delta; unused */
49f55535 358 SSize_t pos_min;
ea3daa5d 359 SSize_t pos_delta;
2c2d71f5 360 SV *last_found;
ea3daa5d 361 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 362 SSize_t last_start_min;
ea3daa5d 363 SSize_t last_start_max;
1de06328
YO
364 SV **longest; /* Either &l_fixed, or &l_float. */
365 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 366 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 367 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
368 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
369 SV *longest_float; /* longest floating string found in pattern */
49f55535 370 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
371 SSize_t offset_float_max; /* latest point in string it can appear */
372 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 373 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
374 I32 flags;
375 I32 whilem_c;
49f55535 376 SSize_t *last_closep;
b8f7bb16 377 regnode_ssc *start_class;
2c2d71f5
JH
378} scan_data_t;
379
a687059c 380/*
e50aee73 381 * Forward declarations for pregcomp()'s friends.
a687059c 382 */
a0d0e21e 383
27da23d5 384static const scan_data_t zero_scan_data =
1de06328 385 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
386
387#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
388#define SF_BEFORE_SEOL 0x0001
389#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
390#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
391#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
392
44e3dfd2
BF
393#define SF_FIX_SHIFT_EOL (+2)
394#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
395
396#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
397#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
398
399#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
400#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
401#define SF_IS_INF 0x0040
402#define SF_HAS_PAR 0x0080
403#define SF_IN_PAR 0x0100
404#define SF_HAS_EVAL 0x0200
405#define SCF_DO_SUBSTR 0x0400
653099ff
GS
406#define SCF_DO_STCLASS_AND 0x0800
407#define SCF_DO_STCLASS_OR 0x1000
408#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 409#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 410
786e8c11 411#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 412#define SCF_SEEN_ACCEPT 0x8000
688e0391 413#define SCF_TRIE_DOING_RESTUDY 0x10000
07be1b83 414
43fead97 415#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
416
417/* The enums for all these are ordered so things work out correctly */
a62b1201 418#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
419#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
420 == REGEX_DEPENDS_CHARSET)
00b27cfc 421#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
422#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
423 >= REGEX_UNICODE_CHARSET)
424#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
425 == REGEX_ASCII_RESTRICTED_CHARSET)
426#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
427 >= REGEX_ASCII_RESTRICTED_CHARSET)
428#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
429 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 430
43fead97 431#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 432
f2c2a6ab
KW
433/* For programs that want to be strictly Unicode compatible by dying if any
434 * attempt is made to match a non-Unicode code point against a Unicode
435 * property. */
436#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
437
93733859 438#define OOB_NAMEDCLASS -1
b8c5462f 439
8e661ac5
KW
440/* There is no code point that is out-of-bounds, so this is problematic. But
441 * its only current use is to initialize a variable that is always set before
442 * looked at. */
443#define OOB_UNICODE 0xDEADBEEF
444
a0ed51b3
LW
445#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
446#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
447
8615cb43 448
b45f050a
JF
449/* length of regex to show in messages that don't mark a position within */
450#define RegexLengthToShowInErrorMessages 127
451
452/*
453 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
454 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
455 * op/pragma/warn/regcomp.
456 */
7253e4e3
RK
457#define MARKER1 "<-- HERE" /* marker as it appears in the description */
458#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 459
538e84ed
KW
460#define REPORT_LOCATION " in regex; marked by " MARKER1 \
461 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 462
c1d900c3
BF
463#define REPORT_LOCATION_ARGS(offset) \
464 UTF8fARG(UTF, offset, RExC_precomp), \
465 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
466
b45f050a
JF
467/*
468 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
469 * arg. Show regex, up to a maximum length. If it's too long, chop and add
470 * "...".
471 */
58e23c8d 472#define _FAIL(code) STMT_START { \
bfed75c6 473 const char *ellipses = ""; \
ccb2c380
MP
474 IV len = RExC_end - RExC_precomp; \
475 \
476 if (!SIZE_ONLY) \
a5e7bc51 477 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
478 if (len > RegexLengthToShowInErrorMessages) { \
479 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
480 len = RegexLengthToShowInErrorMessages - 10; \
481 ellipses = "..."; \
482 } \
58e23c8d 483 code; \
ccb2c380 484} STMT_END
8615cb43 485
58e23c8d 486#define FAIL(msg) _FAIL( \
c1d900c3
BF
487 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
488 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
489
490#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
491 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
492 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 493
b45f050a 494/*
b45f050a
JF
495 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
496 */
ccb2c380 497#define Simple_vFAIL(m) STMT_START { \
a28509cc 498 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380 499 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
c1d900c3 500 m, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 501} STMT_END
b45f050a
JF
502
503/*
504 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
505 */
ccb2c380
MP
506#define vFAIL(m) STMT_START { \
507 if (!SIZE_ONLY) \
a5e7bc51 508 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
509 Simple_vFAIL(m); \
510} STMT_END
b45f050a
JF
511
512/*
513 * Like Simple_vFAIL(), but accepts two arguments.
514 */
ccb2c380 515#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 516 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
517 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
518 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 519} STMT_END
b45f050a
JF
520
521/*
522 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
523 */
ccb2c380
MP
524#define vFAIL2(m,a1) STMT_START { \
525 if (!SIZE_ONLY) \
a5e7bc51 526 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
527 Simple_vFAIL2(m, a1); \
528} STMT_END
b45f050a
JF
529
530
531/*
532 * Like Simple_vFAIL(), but accepts three arguments.
533 */
ccb2c380 534#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 535 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
536 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
537 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 538} STMT_END
b45f050a
JF
539
540/*
541 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
542 */
ccb2c380
MP
543#define vFAIL3(m,a1,a2) STMT_START { \
544 if (!SIZE_ONLY) \
a5e7bc51 545 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
546 Simple_vFAIL3(m, a1, a2); \
547} STMT_END
b45f050a
JF
548
549/*
550 * Like Simple_vFAIL(), but accepts four arguments.
551 */
ccb2c380 552#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 553 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
554 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
555 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 556} STMT_END
b45f050a 557
95db3ffa
KW
558#define vFAIL4(m,a1,a2,a3) STMT_START { \
559 if (!SIZE_ONLY) \
560 SAVEFREESV(RExC_rx_sv); \
561 Simple_vFAIL4(m, a1, a2, a3); \
562} STMT_END
563
946095af
BF
564/* A specialized version of vFAIL2 that works with UTF8f */
565#define vFAIL2utf8f(m, a1) STMT_START { \
ef3f731d
BF
566 const IV offset = RExC_parse - RExC_precomp; \
567 if (!SIZE_ONLY) \
568 SAVEFREESV(RExC_rx_sv); \
569 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
570 REPORT_LOCATION_ARGS(offset)); \
946095af
BF
571} STMT_END
572
499333dc
KW
573/* These have asserts in them because of [perl #122671] Many warnings in
574 * regcomp.c can occur twice. If they get output in pass1 and later in that
575 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
576 * would get output again. So they should be output in pass2, and these
577 * asserts make sure new warnings follow that paradigm. */
946095af 578
5e0a247b
KW
579/* m is not necessarily a "literal string", in this macro */
580#define reg_warn_non_literal_string(loc, m) STMT_START { \
581 const IV offset = loc - RExC_precomp; \
499333dc 582 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
c1d900c3 583 m, REPORT_LOCATION_ARGS(offset)); \
5e0a247b
KW
584} STMT_END
585
668c081a 586#define ckWARNreg(loc,m) STMT_START { \
a28509cc 587 const IV offset = loc - RExC_precomp; \
499333dc 588 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 589 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
590} STMT_END
591
0d6106aa
KW
592#define vWARN_dep(loc, m) STMT_START { \
593 const IV offset = loc - RExC_precomp; \
499333dc 594 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
c1d900c3 595 REPORT_LOCATION_ARGS(offset)); \
0d6106aa
KW
596} STMT_END
597
147508a2
KW
598#define ckWARNdep(loc,m) STMT_START { \
599 const IV offset = loc - RExC_precomp; \
499333dc 600 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
147508a2 601 m REPORT_LOCATION, \
c1d900c3 602 REPORT_LOCATION_ARGS(offset)); \
147508a2
KW
603} STMT_END
604
668c081a 605#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 606 const IV offset = loc - RExC_precomp; \
499333dc 607 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18 608 m REPORT_LOCATION, \
c1d900c3 609 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
610} STMT_END
611
b23eb183 612#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 613 const IV offset = loc - RExC_precomp; \
499333dc 614 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3 615 m REPORT_LOCATION, \
c1d900c3 616 a1, REPORT_LOCATION_ARGS(offset)); \
2335b3d3
KW
617} STMT_END
618
668c081a 619#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 620 const IV offset = loc - RExC_precomp; \
499333dc 621 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 622 a1, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
623} STMT_END
624
625#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 626 const IV offset = loc - RExC_precomp; \
499333dc 627 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 628 a1, a2, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
629} STMT_END
630
668c081a
NC
631#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
632 const IV offset = loc - RExC_precomp; \
499333dc 633 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 634 a1, a2, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
635} STMT_END
636
ccb2c380 637#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 638 const IV offset = loc - RExC_precomp; \
499333dc 639 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 640 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
641} STMT_END
642
668c081a
NC
643#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
644 const IV offset = loc - RExC_precomp; \
499333dc 645 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 646 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
647} STMT_END
648
ccb2c380 649#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 650 const IV offset = loc - RExC_precomp; \
499333dc 651 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 652 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 653} STMT_END
9d1d55b5 654
8615cb43 655
cd439c50 656/* Allow for side effects in s */
ccb2c380
MP
657#define REGC(c,s) STMT_START { \
658 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
659} STMT_END
cd439c50 660
538e84ed 661/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
662 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
663 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
664 * Element 0 holds the number n.
07be1b83 665 * Position is 1 indexed.
fac92740 666 */
7122b237
YO
667#ifndef RE_TRACK_PATTERN_OFFSETS
668#define Set_Node_Offset_To_R(node,byte)
669#define Set_Node_Offset(node,byte)
670#define Set_Cur_Node_Offset
671#define Set_Node_Length_To_R(node,len)
672#define Set_Node_Length(node,len)
6a86c6ad 673#define Set_Node_Cur_Length(node,start)
538e84ed
KW
674#define Node_Offset(n)
675#define Node_Length(n)
7122b237
YO
676#define Set_Node_Offset_Length(node,offset,len)
677#define ProgLen(ri) ri->u.proglen
678#define SetProgLen(ri,x) ri->u.proglen = x
679#else
680#define ProgLen(ri) ri->u.offsets[0]
681#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
682#define Set_Node_Offset_To_R(node,byte) STMT_START { \
683 if (! SIZE_ONLY) { \
684 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 685 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 686 if((node) < 0) { \
538e84ed
KW
687 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
688 (int)(node)); \
ccb2c380
MP
689 } else { \
690 RExC_offsets[2*(node)-1] = (byte); \
691 } \
692 } \
693} STMT_END
694
695#define Set_Node_Offset(node,byte) \
696 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
697#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
698
699#define Set_Node_Length_To_R(node,len) STMT_START { \
700 if (! SIZE_ONLY) { \
701 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 702 __LINE__, (int)(node), (int)(len))); \
ccb2c380 703 if((node) < 0) { \
538e84ed
KW
704 Perl_croak(aTHX_ "value of node is %d in Length macro", \
705 (int)(node)); \
ccb2c380
MP
706 } else { \
707 RExC_offsets[2*(node)] = (len); \
708 } \
709 } \
710} STMT_END
711
712#define Set_Node_Length(node,len) \
713 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
714#define Set_Node_Cur_Length(node, start) \
715 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
716
717/* Get offsets and lengths */
718#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
719#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
720
07be1b83
YO
721#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
722 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
723 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
724} STMT_END
7122b237 725#endif
07be1b83
YO
726
727#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
728#define EXPERIMENTAL_INPLACESCAN
f427392e 729#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 730
9e9ecfdf 731#define DEBUG_RExC_seen() \
538e84ed
KW
732 DEBUG_OPTIMISE_MORE_r({ \
733 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
734 \
e384d5c1
YO
735 if (RExC_seen & REG_ZERO_LEN_SEEN) \
736 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
538e84ed 737 \
e384d5c1
YO
738 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
739 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
538e84ed 740 \
e384d5c1
YO
741 if (RExC_seen & REG_GPOS_SEEN) \
742 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
538e84ed 743 \
e384d5c1
YO
744 if (RExC_seen & REG_CANY_SEEN) \
745 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
538e84ed 746 \
e384d5c1
YO
747 if (RExC_seen & REG_RECURSE_SEEN) \
748 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
538e84ed 749 \
e384d5c1
YO
750 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
751 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 752 \
e384d5c1
YO
753 if (RExC_seen & REG_VERBARG_SEEN) \
754 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
538e84ed 755 \
e384d5c1
YO
756 if (RExC_seen & REG_CUTGROUP_SEEN) \
757 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
538e84ed 758 \
e384d5c1
YO
759 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
760 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 761 \
e384d5c1
YO
762 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
763 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 764 \
e384d5c1
YO
765 if (RExC_seen & REG_GOSTART_SEEN) \
766 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
538e84ed 767 \
ee273784
YO
768 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
769 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
770 \
538e84ed 771 PerlIO_printf(Perl_debug_log,"\n"); \
9e9ecfdf
YO
772 });
773
304ee84b
YO
774#define DEBUG_STUDYDATA(str,data,depth) \
775DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 776 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
777 "%*s" str "Pos:%"IVdf"/%"IVdf \
778 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
779 (int)(depth)*2, "", \
780 (IV)((data)->pos_min), \
781 (IV)((data)->pos_delta), \
304ee84b 782 (UV)((data)->flags), \
1de06328 783 (IV)((data)->whilem_c), \
304ee84b
YO
784 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
785 is_inf ? "INF " : "" \
1de06328
YO
786 ); \
787 if ((data)->last_found) \
788 PerlIO_printf(Perl_debug_log, \
789 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
790 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
791 SvPVX_const((data)->last_found), \
792 (IV)((data)->last_end), \
793 (IV)((data)->last_start_min), \
794 (IV)((data)->last_start_max), \
795 ((data)->longest && \
796 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
797 SvPVX_const((data)->longest_fixed), \
798 (IV)((data)->offset_fixed), \
799 ((data)->longest && \
800 (data)->longest==&((data)->longest_float)) ? "*" : "", \
801 SvPVX_const((data)->longest_float), \
802 (IV)((data)->offset_float_min), \
803 (IV)((data)->offset_float_max) \
804 ); \
805 PerlIO_printf(Perl_debug_log,"\n"); \
806});
807
549b4e78
KW
808#ifdef DEBUGGING
809
8e35b056
KW
810/* is c a control character for which we have a mnemonic? */
811#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
812
549b4e78
KW
813STATIC const char *
814S_cntrl_to_mnemonic(const U8 c)
815{
816 /* Returns the mnemonic string that represents character 'c', if one
817 * exists; NULL otherwise. The only ones that exist for the purposes of
818 * this routine are a few control characters */
819
820 switch (c) {
821 case '\a': return "\\a";
822 case '\b': return "\\b";
823 case ESC_NATIVE: return "\\e";
824 case '\f': return "\\f";
825 case '\n': return "\\n";
826 case '\r': return "\\r";
827 case '\t': return "\\t";
828 }
829
830 return NULL;
831}
832
833#endif
834
653099ff 835/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 836 Update the longest found anchored substring and the longest found
653099ff
GS
837 floating substrings if needed. */
838
4327152a 839STATIC void
ea3daa5d
FC
840S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
841 SSize_t *minlenp, int is_inf)
c277df42 842{
e1ec3a88
AL
843 const STRLEN l = CHR_SVLEN(data->last_found);
844 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 845 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 846
7918f24d
NC
847 PERL_ARGS_ASSERT_SCAN_COMMIT;
848
c277df42 849 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 850 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
851 if (*data->longest == data->longest_fixed) {
852 data->offset_fixed = l ? data->last_start_min : data->pos_min;
853 if (data->flags & SF_BEFORE_EOL)
b81d288d 854 data->flags
c277df42
IZ
855 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
856 else
857 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 858 data->minlen_fixed=minlenp;
1de06328 859 data->lookbehind_fixed=0;
a0ed51b3 860 }
304ee84b 861 else { /* *data->longest == data->longest_float */
c277df42 862 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
863 data->offset_float_max = (l
864 ? data->last_start_max
ea3daa5d
FC
865 : (data->pos_delta == SSize_t_MAX
866 ? SSize_t_MAX
867 : data->pos_min + data->pos_delta));
868 if (is_inf
869 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
870 data->offset_float_max = SSize_t_MAX;
c277df42 871 if (data->flags & SF_BEFORE_EOL)
b81d288d 872 data->flags
c277df42
IZ
873 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
874 else
875 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
876 data->minlen_float=minlenp;
877 data->lookbehind_float=0;
c277df42
IZ
878 }
879 }
880 SvCUR_set(data->last_found, 0);
0eda9292 881 {
a28509cc 882 SV * const sv = data->last_found;
097eb12c
AL
883 if (SvUTF8(sv) && SvMAGICAL(sv)) {
884 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
885 if (mg)
886 mg->mg_len = 0;
887 }
0eda9292 888 }
c277df42
IZ
889 data->last_end = -1;
890 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 891 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
892}
893
cdd87c1d
KW
894/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
895 * list that describes which code points it matches */
896
653099ff 897STATIC void
3420edd7 898S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 899{
cdd87c1d
KW
900 /* Set the SSC 'ssc' to match an empty string or any code point */
901
557bd3fb 902 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 903
71068078 904 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 905
cdd87c1d
KW
906 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
907 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
93e92956 908 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
909}
910
653099ff 911STATIC int
dc3bf405 912S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 913{
c144baaa
KW
914 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
915 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
916 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
917 * in any way, so there's no point in using it */
cdd87c1d
KW
918
919 UV start, end;
920 bool ret;
653099ff 921
557bd3fb 922 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 923
71068078 924 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 925
93e92956 926 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
927 return FALSE;
928 }
929
930 /* See if the list consists solely of the range 0 - Infinity */
931 invlist_iterinit(ssc->invlist);
932 ret = invlist_iternext(ssc->invlist, &start, &end)
933 && start == 0
934 && end == UV_MAX;
935
936 invlist_iterfinish(ssc->invlist);
937
938 if (ret) {
939 return TRUE;
940 }
941
942 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 943 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
944 int i;
945 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
946 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
947 return TRUE;
948 }
949 }
950 }
951
952 return FALSE;
653099ff
GS
953}
954
653099ff 955STATIC void
cdd87c1d 956S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 957{
cdd87c1d
KW
958 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
959 * string, any code point, or any posix class under locale */
960
557bd3fb 961 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 962
557bd3fb 963 Zero(ssc, 1, regnode_ssc);
71068078 964 set_ANYOF_SYNTHETIC(ssc);
93e92956 965 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 966 ssc_anything(ssc);
cdd87c1d
KW
967
968 /* If any portion of the regex is to operate under locale rules,
969 * initialization includes it. The reason this isn't done for all regexes
970 * is that the optimizer was written under the assumption that locale was
971 * all-or-nothing. Given the complexity and lack of documentation in the
972 * optimizer, and that there are inadequate test cases for locale, many
973 * parts of it may not work properly, it is safest to avoid locale unless
974 * necessary. */
975 if (RExC_contains_locale) {
976 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
977 }
978 else {
979 ANYOF_POSIXL_ZERO(ssc);
980 }
653099ff
GS
981}
982
b423522f 983STATIC int
dc3bf405
BF
984S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
985 const regnode_ssc *ssc)
b423522f
KW
986{
987 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
988 * to the list of code points matched, and locale posix classes; hence does
989 * not check its flags) */
990
991 UV start, end;
992 bool ret;
993
994 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
995
71068078 996 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
997
998 invlist_iterinit(ssc->invlist);
999 ret = invlist_iternext(ssc->invlist, &start, &end)
1000 && start == 0
1001 && end == UV_MAX;
1002
1003 invlist_iterfinish(ssc->invlist);
1004
1005 if (! ret) {
1006 return FALSE;
1007 }
1008
e0e1be5f 1009 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1010 return FALSE;
b423522f
KW
1011 }
1012
1013 return TRUE;
1014}
1015
1016STATIC SV*
1017S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1018 const regnode_charclass* const node)
b423522f
KW
1019{
1020 /* Returns a mortal inversion list defining which code points are matched
1021 * by 'node', which is of type ANYOF. Handles complementing the result if
1022 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1023 * returned list must, and will, contain every code point that is a
1024 * possibility. */
b423522f
KW
1025
1026 SV* invlist = sv_2mortal(_new_invlist(0));
1ee208c4 1027 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1028 unsigned int i;
1029 const U32 n = ARG(node);
31f05a37 1030 bool new_node_has_latin1 = FALSE;
b423522f
KW
1031
1032 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1033
1034 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1035 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1036 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1037 AV * const av = MUTABLE_AV(SvRV(rv));
1038 SV **const ary = AvARRAY(av);
1039 assert(RExC_rxi->data->what[n] == 's');
1040
1041 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1042 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1043 }
1044 else if (ary[0] && ary[0] != &PL_sv_undef) {
1045
1046 /* Here, no compile-time swash, and there are things that won't be
1047 * known until runtime -- we have to assume it could be anything */
1048 return _add_range_to_invlist(invlist, 0, UV_MAX);
1049 }
1ee208c4 1050 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1051
1052 /* Here no compile-time swash, and no run-time only data. Use the
1053 * node's inversion list */
1ee208c4
KW
1054 invlist = sv_2mortal(invlist_clone(ary[3]));
1055 }
1056
1057 /* Get the code points valid only under UTF-8 locales */
1058 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1059 && ary[2] && ary[2] != &PL_sv_undef)
1060 {
1061 only_utf8_locale_invlist = ary[2];
b423522f
KW
1062 }
1063 }
1064
dcb20b36
KW
1065 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1066 * code points, and an inversion list for the others, but if there are code
1067 * points that should match only conditionally on the target string being
1068 * UTF-8, those are placed in the inversion list, and not the bitmap.
1069 * Since there are circumstances under which they could match, they are
1070 * included in the SSC. But if the ANYOF node is to be inverted, we have
1071 * to exclude them here, so that when we invert below, the end result
1072 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1073 * have to do this here before we add the unconditionally matched code
1074 * points */
b423522f
KW
1075 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1076 _invlist_intersection_complement_2nd(invlist,
1077 PL_UpperLatin1,
1078 &invlist);
1079 }
1080
1081 /* Add in the points from the bit map */
dcb20b36 1082 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f
KW
1083 if (ANYOF_BITMAP_TEST(node, i)) {
1084 invlist = add_cp_to_invlist(invlist, i);
31f05a37 1085 new_node_has_latin1 = TRUE;
b423522f
KW
1086 }
1087 }
1088
1089 /* If this can match all upper Latin1 code points, have to add them
1090 * as well */
93e92956 1091 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
b423522f
KW
1092 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1093 }
1094
1095 /* Similarly for these */
93e92956 1096 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1097 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1098 }
1099
1100 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1101 _invlist_invert(invlist);
1102 }
31f05a37
KW
1103 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1104
1105 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1106 * locale. We can skip this if there are no 0-255 at all. */
1107 _invlist_union(invlist, PL_Latin1, &invlist);
1108 }
1109
1ee208c4
KW
1110 /* Similarly add the UTF-8 locale possible matches. These have to be
1111 * deferred until after the non-UTF-8 locale ones are taken care of just
1112 * above, or it leads to wrong results under ANYOF_INVERT */
1113 if (only_utf8_locale_invlist) {
31f05a37 1114 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1115 only_utf8_locale_invlist,
31f05a37
KW
1116 ANYOF_FLAGS(node) & ANYOF_INVERT,
1117 &invlist);
1118 }
b423522f
KW
1119
1120 return invlist;
1121}
1122
1051e1c4 1123/* These two functions currently do the exact same thing */
557bd3fb 1124#define ssc_init_zero ssc_init
653099ff 1125
cdd87c1d
KW
1126#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1127#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1128
557bd3fb 1129/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1130 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1131 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1132
653099ff 1133STATIC void
b423522f 1134S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1135 const regnode_charclass *and_with)
653099ff 1136{
cdd87c1d
KW
1137 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1138 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1139
a0dd4231
KW
1140 SV* anded_cp_list;
1141 U8 anded_flags;
1e6ade67 1142
cdd87c1d 1143 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1144
71068078 1145 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1146
1147 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1148 * the code point inversion list and just the relevant flags */
71068078 1149 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1150 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1151 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1152
1153 /* XXX This is a kludge around what appears to be deficiencies in the
1154 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1155 * there are paths through the optimizer where it doesn't get weeded
1156 * out when it should. And if we don't make some extra provision for
1157 * it like the code just below, it doesn't get added when it should.
1158 * This solution is to add it only when AND'ing, which is here, and
1159 * only when what is being AND'ed is the pristine, original node
1160 * matching anything. Thus it is like adding it to ssc_anything() but
1161 * only when the result is to be AND'ed. Probably the same solution
1162 * could be adopted for the same problem we have with /l matching,
1163 * which is solved differently in S_ssc_init(), and that would lead to
1164 * fewer false positives than that solution has. But if this solution
1165 * creates bugs, the consequences are only that a warning isn't raised
1166 * that should be; while the consequences for having /l bugs is
1167 * incorrect matches */
7dcac5f6 1168 if (ssc_is_anything((regnode_ssc *)and_with)) {
e9b08962
KW
1169 anded_flags |= ANYOF_WARN_SUPER;
1170 }
a0dd4231
KW
1171 }
1172 else {
5c0f85ef 1173 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
eff8b7dc 1174 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1175 }
1176
1177 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1178
1179 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1180 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1181 * 'and_with' may be inverted. When not inverted, we have the situation of
1182 * computing:
1183 * (C1 | P1) & (C2 | P2)
1184 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1185 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1186 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1187 * <= ((C1 & C2) | P1 | P2)
1188 * Alternatively, the last few steps could be:
1189 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1190 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1191 * <= (C1 | C2 | (P1 & P2))
1192 * We favor the second approach if either P1 or P2 is non-empty. This is
1193 * because these components are a barrier to doing optimizations, as what
1194 * they match cannot be known until the moment of matching as they are
1195 * dependent on the current locale, 'AND"ing them likely will reduce or
1196 * eliminate them.
1197 * But we can do better if we know that C1,P1 are in their initial state (a
1198 * frequent occurrence), each matching everything:
1199 * (<everything>) & (C2 | P2) = C2 | P2
1200 * Similarly, if C2,P2 are in their initial state (again a frequent
1201 * occurrence), the result is a no-op
1202 * (C1 | P1) & (<everything>) = C1 | P1
1203 *
1204 * Inverted, we have
1205 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1206 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1207 * <= (C1 & ~C2) | (P1 & ~P2)
1208 * */
1aa99e6b 1209
a0dd4231 1210 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1211 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1212 {
cdd87c1d 1213 unsigned int i;
8951c461 1214
cdd87c1d
KW
1215 ssc_intersection(ssc,
1216 anded_cp_list,
1217 FALSE /* Has already been inverted */
1218 );
c6b76537 1219
cdd87c1d
KW
1220 /* If either P1 or P2 is empty, the intersection will be also; can skip
1221 * the loop */
93e92956 1222 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1223 ANYOF_POSIXL_ZERO(ssc);
1224 }
e0e1be5f 1225 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1226
1227 /* Note that the Posix class component P from 'and_with' actually
1228 * looks like:
1229 * P = Pa | Pb | ... | Pn
1230 * where each component is one posix class, such as in [\w\s].
1231 * Thus
1232 * ~P = ~(Pa | Pb | ... | Pn)
1233 * = ~Pa & ~Pb & ... & ~Pn
1234 * <= ~Pa | ~Pb | ... | ~Pn
1235 * The last is something we can easily calculate, but unfortunately
1236 * is likely to have many false positives. We could do better
1237 * in some (but certainly not all) instances if two classes in
1238 * P have known relationships. For example
1239 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1240 * So
1241 * :lower: & :print: = :lower:
1242 * And similarly for classes that must be disjoint. For example,
1243 * since \s and \w can have no elements in common based on rules in
1244 * the POSIX standard,
1245 * \w & ^\S = nothing
1246 * Unfortunately, some vendor locales do not meet the Posix
1247 * standard, in particular almost everything by Microsoft.
1248 * The loop below just changes e.g., \w into \W and vice versa */
1249
1ee208c4 1250 regnode_charclass_posixl temp;
cdd87c1d
KW
1251 int add = 1; /* To calculate the index of the complement */
1252
1253 ANYOF_POSIXL_ZERO(&temp);
1254 for (i = 0; i < ANYOF_MAX; i++) {
1255 assert(i % 2 != 0
7dcac5f6
KW
1256 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1257 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1258
7dcac5f6 1259 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1260 ANYOF_POSIXL_SET(&temp, i + add);
1261 }
1262 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1263 }
1264 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1265
cdd87c1d
KW
1266 } /* else ssc already has no posixes */
1267 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1268 in its initial state */
71068078 1269 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1270 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1271 {
1272 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1273 * copy it over 'ssc' */
1274 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1275 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1276 StructCopy(and_with, ssc, regnode_ssc);
1277 }
1278 else {
1279 ssc->invlist = anded_cp_list;
1280 ANYOF_POSIXL_ZERO(ssc);
93e92956 1281 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1282 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1283 }
1284 }
1285 }
e0e1be5f 1286 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1287 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1288 {
1289 /* One or the other of P1, P2 is non-empty. */
93e92956 1290 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1291 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1292 }
cdd87c1d
KW
1293 ssc_union(ssc, anded_cp_list, FALSE);
1294 }
1295 else { /* P1 = P2 = empty */
1296 ssc_intersection(ssc, anded_cp_list, FALSE);
1297 }
137165a6 1298 }
653099ff
GS
1299}
1300
653099ff 1301STATIC void
cdd87c1d 1302S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1303 const regnode_charclass *or_with)
653099ff 1304{
cdd87c1d
KW
1305 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1306 * another SSC or a regular ANYOF class. Can create false positives if
1307 * 'or_with' is to be inverted. */
7918f24d 1308
a0dd4231
KW
1309 SV* ored_cp_list;
1310 U8 ored_flags;
c6b76537 1311
cdd87c1d 1312 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1313
71068078 1314 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1315
1316 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1317 * the code point inversion list and just the relevant flags */
71068078 1318 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1319 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1320 ored_flags = ANYOF_FLAGS(or_with);
1321 }
1322 else {
5c0f85ef 1323 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1324 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1325 }
1326
1327 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1328
1329 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1330 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1331 * 'or_with' may be inverted. When not inverted, we have the simple
1332 * situation of computing:
1333 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1334 * If P1|P2 yields a situation with both a class and its complement are
1335 * set, like having both \w and \W, this matches all code points, and we
1336 * can delete these from the P component of the ssc going forward. XXX We
1337 * might be able to delete all the P components, but I (khw) am not certain
1338 * about this, and it is better to be safe.
1339 *
1340 * Inverted, we have
1341 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1342 * <= (C1 | P1) | ~C2
1343 * <= (C1 | ~C2) | P1
1344 * (which results in actually simpler code than the non-inverted case)
1345 * */
9826f543 1346
a0dd4231 1347 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1348 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1349 {
cdd87c1d 1350 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1351 } /* else Not inverted */
93e92956 1352 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1353 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1354 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1355 unsigned int i;
1356 for (i = 0; i < ANYOF_MAX; i += 2) {
1357 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1358 {
1359 ssc_match_all_cp(ssc);
1360 ANYOF_POSIXL_CLEAR(ssc, i);
1361 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1362 }
1363 }
1364 }
1aa99e6b 1365 }
cdd87c1d
KW
1366
1367 ssc_union(ssc,
1368 ored_cp_list,
1369 FALSE /* Already has been inverted */
1370 );
653099ff
GS
1371}
1372
b423522f
KW
1373PERL_STATIC_INLINE void
1374S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1375{
1376 PERL_ARGS_ASSERT_SSC_UNION;
1377
71068078 1378 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1379
1380 _invlist_union_maybe_complement_2nd(ssc->invlist,
1381 invlist,
1382 invert2nd,
1383 &ssc->invlist);
1384}
1385
1386PERL_STATIC_INLINE void
1387S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1388 SV* const invlist,
1389 const bool invert2nd)
1390{
1391 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1392
71068078 1393 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1394
1395 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1396 invlist,
1397 invert2nd,
1398 &ssc->invlist);
1399}
1400
1401PERL_STATIC_INLINE void
1402S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1403{
1404 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1405
71068078 1406 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1407
1408 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1409}
1410
1411PERL_STATIC_INLINE void
1412S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1413{
1414 /* AND just the single code point 'cp' into the SSC 'ssc' */
1415
1416 SV* cp_list = _new_invlist(2);
1417
1418 PERL_ARGS_ASSERT_SSC_CP_AND;
1419
71068078 1420 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1421
1422 cp_list = add_cp_to_invlist(cp_list, cp);
1423 ssc_intersection(ssc, cp_list,
1424 FALSE /* Not inverted */
1425 );
1426 SvREFCNT_dec_NN(cp_list);
1427}
1428
1429PERL_STATIC_INLINE void
dc3bf405 1430S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1431{
1432 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1433 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1434
71068078 1435 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1436
1437 ANYOF_POSIXL_ZERO(ssc);
1438 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1439}
1440
1441STATIC void
1442S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1443{
1444 /* The inversion list in the SSC is marked mortal; now we need a more
1445 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1446 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1447 * map */
b423522f
KW
1448
1449 SV* invlist = invlist_clone(ssc->invlist);
1450
1451 PERL_ARGS_ASSERT_SSC_FINALIZE;
1452
71068078 1453 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1454
a0dd4231 1455 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1456 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1457 * by the time we reach here */
eff8b7dc 1458 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
a0dd4231 1459
b423522f
KW
1460 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1461
1ee208c4
KW
1462 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1463 NULL, NULL, NULL, FALSE);
b423522f 1464
85c8e306
KW
1465 /* Make sure is clone-safe */
1466 ssc->invlist = NULL;
1467
e0e1be5f 1468 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1469 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1470 }
1462525b
KW
1471
1472 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1473}
1474
a3621e74
YO
1475#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1476#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1477#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1478#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1479 ? (TRIE_LIST_CUR( idx ) - 1) \
1480 : 0 )
a3621e74 1481
3dab1dad
YO
1482
1483#ifdef DEBUGGING
07be1b83 1484/*
2b8b4781
NC
1485 dump_trie(trie,widecharmap,revcharmap)
1486 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1487 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1488
1489 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1490 The _interim_ variants are used for debugging the interim
1491 tables that are used to generate the final compressed
1492 representation which is what dump_trie expects.
1493
486ec47a 1494 Part of the reason for their existence is to provide a form
3dab1dad 1495 of documentation as to how the different representations function.
07be1b83
YO
1496
1497*/
3dab1dad
YO
1498
1499/*
3dab1dad
YO
1500 Dumps the final compressed table form of the trie to Perl_debug_log.
1501 Used for debugging make_trie().
1502*/
b9a59e08 1503
3dab1dad 1504STATIC void
2b8b4781
NC
1505S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1506 AV *revcharmap, U32 depth)
3dab1dad
YO
1507{
1508 U32 state;
ab3bbdeb 1509 SV *sv=sv_newmortal();
55eed653 1510 int colwidth= widecharmap ? 6 : 4;
2e64971a 1511 U16 word;
3dab1dad
YO
1512 GET_RE_DEBUG_FLAGS_DECL;
1513
7918f24d 1514 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1515
3dab1dad
YO
1516 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1517 (int)depth * 2 + 2,"",
1518 "Match","Base","Ofs" );
1519
1520 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1521 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1522 if ( tmp ) {
538e84ed 1523 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1524 colwidth,
538e84ed 1525 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1526 PL_colors[0], PL_colors[1],
1527 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1528 PERL_PV_ESCAPE_FIRSTCHAR
1529 )
ab3bbdeb 1530 );
3dab1dad
YO
1531 }
1532 }
1533 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1534 (int)depth * 2 + 2,"");
1535
1536 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1537 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1538 PerlIO_printf( Perl_debug_log, "\n");
1539
1e2e3d02 1540 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1541 const U32 base = trie->states[ state ].trans.base;
3dab1dad 1542
538e84ed
KW
1543 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1544 (int)depth * 2 + 2,"", (UV)state);
3dab1dad
YO
1545
1546 if ( trie->states[ state ].wordnum ) {
538e84ed
KW
1547 PerlIO_printf( Perl_debug_log, " W%4X",
1548 trie->states[ state ].wordnum );
3dab1dad
YO
1549 } else {
1550 PerlIO_printf( Perl_debug_log, "%6s", "" );
1551 }
1552
1553 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1554
1555 if ( base ) {
1556 U32 ofs = 0;
1557
1558 while( ( base + ofs < trie->uniquecharcount ) ||
1559 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
1560 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1561 != state))
3dab1dad
YO
1562 ofs++;
1563
1564 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1565
1566 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
1567 if ( ( base + ofs >= trie->uniquecharcount )
1568 && ( base + ofs - trie->uniquecharcount
1569 < trie->lasttrans )
1570 && trie->trans[ base + ofs
1571 - trie->uniquecharcount ].check == state )
3dab1dad 1572 {
ab3bbdeb
YO
1573 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1574 colwidth,
538e84ed
KW
1575 (UV)trie->trans[ base + ofs
1576 - trie->uniquecharcount ].next );
3dab1dad 1577 } else {
ab3bbdeb 1578 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1579 }
1580 }
1581
1582 PerlIO_printf( Perl_debug_log, "]");
1583
1584 }
1585 PerlIO_printf( Perl_debug_log, "\n" );
1586 }
538e84ed
KW
1587 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1588 (int)depth*2, "");
2e64971a
DM
1589 for (word=1; word <= trie->wordcount; word++) {
1590 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1591 (int)word, (int)(trie->wordinfo[word].prev),
1592 (int)(trie->wordinfo[word].len));
1593 }
1594 PerlIO_printf(Perl_debug_log, "\n" );
538e84ed 1595}
3dab1dad 1596/*
3dab1dad 1597 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 1598 List tries normally only are used for construction when the number of
3dab1dad
YO
1599 possible chars (trie->uniquecharcount) is very high.
1600 Used for debugging make_trie().
1601*/
1602STATIC void
55eed653 1603S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1604 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1605 U32 depth)
3dab1dad
YO
1606{
1607 U32 state;
ab3bbdeb 1608 SV *sv=sv_newmortal();
55eed653 1609 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1610 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1611
1612 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1613
3dab1dad 1614 /* print out the table precompression. */
ab3bbdeb
YO
1615 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1616 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1617 "------:-----+-----------------\n" );
538e84ed 1618
3dab1dad
YO
1619 for( state=1 ; state < next_alloc ; state ++ ) {
1620 U16 charid;
538e84ed 1621
ab3bbdeb 1622 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1623 (int)depth * 2 + 2,"", (UV)state );
1624 if ( ! trie->states[ state ].wordnum ) {
1625 PerlIO_printf( Perl_debug_log, "%5s| ","");
1626 } else {
1627 PerlIO_printf( Perl_debug_log, "W%4x| ",
1628 trie->states[ state ].wordnum
1629 );
1630 }
1631 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
1632 SV ** const tmp = av_fetch( revcharmap,
1633 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1634 if ( tmp ) {
1635 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1636 colwidth,
538e84ed
KW
1637 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1638 colwidth,
1639 PL_colors[0], PL_colors[1],
1640 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1641 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 1642 ) ,
1e2e3d02
YO
1643 TRIE_LIST_ITEM(state,charid).forid,
1644 (UV)TRIE_LIST_ITEM(state,charid).newstate
1645 );
538e84ed 1646 if (!(charid % 10))
664e119d
RGS
1647 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1648 (int)((depth * 2) + 14), "");
1e2e3d02 1649 }
ab3bbdeb
YO
1650 }
1651 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad 1652 }
538e84ed 1653}
3dab1dad
YO
1654
1655/*
3dab1dad 1656 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
1657 This is the normal DFA style state transition table, with a few
1658 twists to facilitate compression later.
3dab1dad
YO
1659 Used for debugging make_trie().
1660*/
1661STATIC void
55eed653 1662S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1663 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1664 U32 depth)
3dab1dad
YO
1665{
1666 U32 state;
1667 U16 charid;
ab3bbdeb 1668 SV *sv=sv_newmortal();
55eed653 1669 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1670 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1671
1672 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 1673
3dab1dad
YO
1674 /*
1675 print out the table precompression so that we can do a visual check
1676 that they are identical.
1677 */
538e84ed 1678
3dab1dad
YO
1679 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1680
1681 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1682 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1683 if ( tmp ) {
538e84ed 1684 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1685 colwidth,
538e84ed 1686 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1687 PL_colors[0], PL_colors[1],
1688 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1689 PERL_PV_ESCAPE_FIRSTCHAR
1690 )
ab3bbdeb 1691 );
3dab1dad
YO
1692 }
1693 }
1694
1695 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1696
1697 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1698 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1699 }
1700
1701 PerlIO_printf( Perl_debug_log, "\n" );
1702
1703 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1704
538e84ed 1705 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
3dab1dad
YO
1706 (int)depth * 2 + 2,"",
1707 (UV)TRIE_NODENUM( state ) );
1708
1709 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1710 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1711 if (v)
1712 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1713 else
1714 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1715 }
1716 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
538e84ed
KW
1717 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1718 (UV)trie->trans[ state ].check );
3dab1dad 1719 } else {
538e84ed
KW
1720 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1721 (UV)trie->trans[ state ].check,
3dab1dad
YO
1722 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1723 }
1724 }
07be1b83 1725}
3dab1dad
YO
1726
1727#endif
1728
2e64971a 1729
786e8c11
YO
1730/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1731 startbranch: the first branch in the whole branch sequence
1732 first : start branch of sequence of branch-exact nodes.
1733 May be the same as startbranch
1734 last : Thing following the last branch.
1735 May be the same as tail.
1736 tail : item following the branch sequence
1737 count : words in the sequence
cce29a1d 1738 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS)/
786e8c11 1739 depth : indent depth
3dab1dad 1740
786e8c11 1741Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1742
786e8c11
YO
1743A trie is an N'ary tree where the branches are determined by digital
1744decomposition of the key. IE, at the root node you look up the 1st character and
1745follow that branch repeat until you find the end of the branches. Nodes can be
1746marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1747
786e8c11 1748 /he|she|his|hers/
72f13be8 1749
786e8c11
YO
1750would convert into the following structure. Numbers represent states, letters
1751following numbers represent valid transitions on the letter from that state, if
1752the number is in square brackets it represents an accepting state, otherwise it
1753will be in parenthesis.
07be1b83 1754
786e8c11
YO
1755 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1756 | |
1757 | (2)
1758 | |
1759 (1) +-i->(6)-+-s->[7]
1760 |
1761 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1762
786e8c11
YO
1763 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1764
1765This shows that when matching against the string 'hers' we will begin at state 1
1766read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1767then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1768is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1769single traverse. We store a mapping from accepting to state to which word was
1770matched, and then when we have multiple possibilities we try to complete the
1771rest of the regex in the order in which they occured in the alternation.
1772
1773The only prior NFA like behaviour that would be changed by the TRIE support is
1774the silent ignoring of duplicate alternations which are of the form:
1775
1776 / (DUPE|DUPE) X? (?{ ... }) Y /x
1777
4b714af6 1778Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1779and without the optimisation. With the optimisations dupes will be silently
486ec47a 1780ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1781the following demonstrates:
1782
1783 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1784
1785which prints out 'word' three times, but
1786
1787 'words'=~/(word|word|word)(?{ print $1 })S/
1788
1789which doesnt print it out at all. This is due to other optimisations kicking in.
1790
1791Example of what happens on a structural level:
1792
486ec47a 1793The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1794
1795 1: CURLYM[1] {1,32767}(18)
1796 5: BRANCH(8)
1797 6: EXACT <ac>(16)
1798 8: BRANCH(11)
1799 9: EXACT <ad>(16)
1800 11: BRANCH(14)
1801 12: EXACT <ab>(16)
1802 16: SUCCEED(0)
1803 17: NOTHING(18)
1804 18: END(0)
1805
1806This would be optimizable with startbranch=5, first=5, last=16, tail=16
1807and should turn into:
1808
1809 1: CURLYM[1] {1,32767}(18)
1810 5: TRIE(16)
1811 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1812 <ac>
1813 <ad>
1814 <ab>
1815 16: SUCCEED(0)
1816 17: NOTHING(18)
1817 18: END(0)
1818
1819Cases where tail != last would be like /(?foo|bar)baz/:
1820
1821 1: BRANCH(4)
1822 2: EXACT <foo>(8)
1823 4: BRANCH(7)
1824 5: EXACT <bar>(8)
1825 7: TAIL(8)
1826 8: EXACT <baz>(10)
1827 10: END(0)
1828
1829which would be optimizable with startbranch=1, first=1, last=7, tail=8
1830and would end up looking like:
1831
1832 1: TRIE(8)
1833 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1834 <foo>
1835 <bar>
1836 7: TAIL(8)
1837 8: EXACT <baz>(10)
1838 10: END(0)
1839
c80e42f3 1840 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
1841
1842is the recommended Unicode-aware way of saying
1843
1844 *(d++) = uv;
1845*/
1846
fab2782b 1847#define TRIE_STORE_REVCHAR(val) \
786e8c11 1848 STMT_START { \
73031816 1849 if (UTF) { \
fab2782b 1850 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1851 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 1852 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
1853 SvCUR_set(zlopp, kapow - flrbbbbb); \
1854 SvPOK_on(zlopp); \
1855 SvUTF8_on(zlopp); \
1856 av_push(revcharmap, zlopp); \
1857 } else { \
fab2782b 1858 char ooooff = (char)val; \
73031816
NC
1859 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1860 } \
1861 } STMT_END
786e8c11 1862
914a25d5
KW
1863/* This gets the next character from the input, folding it if not already
1864 * folded. */
1865#define TRIE_READ_CHAR STMT_START { \
1866 wordlen++; \
1867 if ( UTF ) { \
1868 /* if it is UTF then it is either already folded, or does not need \
1869 * folding */ \
1c1d615a 1870 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
1871 } \
1872 else if (folder == PL_fold_latin1) { \
7d006b13
KW
1873 /* This folder implies Unicode rules, which in the range expressible \
1874 * by not UTF is the lower case, with the two exceptions, one of \
1875 * which should have been taken care of before calling this */ \
1876 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1877 uvc = toLOWER_L1(*uc); \
1878 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1879 len = 1; \
914a25d5
KW
1880 } else { \
1881 /* raw data, will be folded later if needed */ \
1882 uvc = (U32)*uc; \
1883 len = 1; \
1884 } \
786e8c11
YO
1885} STMT_END
1886
1887
1888
1889#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1890 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1891 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1892 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1893 } \
1894 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1895 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1896 TRIE_LIST_CUR( state )++; \
1897} STMT_END
07be1b83 1898
786e8c11
YO
1899#define TRIE_LIST_NEW(state) STMT_START { \
1900 Newxz( trie->states[ state ].trans.list, \
1901 4, reg_trie_trans_le ); \
1902 TRIE_LIST_CUR( state ) = 1; \
1903 TRIE_LIST_LEN( state ) = 4; \
1904} STMT_END
07be1b83 1905
786e8c11
YO
1906#define TRIE_HANDLE_WORD(state) STMT_START { \
1907 U16 dupe= trie->states[ state ].wordnum; \
1908 regnode * const noper_next = regnext( noper ); \
1909 \
786e8c11
YO
1910 DEBUG_r({ \
1911 /* store the word for dumping */ \
1912 SV* tmp; \
1913 if (OP(noper) != NOTHING) \
740cce10 1914 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1915 else \
740cce10 1916 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1917 av_push( trie_words, tmp ); \
786e8c11
YO
1918 }); \
1919 \
1920 curword++; \
2e64971a
DM
1921 trie->wordinfo[curword].prev = 0; \
1922 trie->wordinfo[curword].len = wordlen; \
1923 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1924 \
1925 if ( noper_next < tail ) { \
1926 if (!trie->jump) \
538e84ed
KW
1927 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1928 sizeof(U16) ); \
7f69552c 1929 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1930 if (!jumper) \
1931 jumper = noper_next; \
1932 if (!nextbranch) \
1933 nextbranch= regnext(cur); \
1934 } \
1935 \
1936 if ( dupe ) { \
2e64971a
DM
1937 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1938 /* chain, so that when the bits of chain are later */\
1939 /* linked together, the dups appear in the chain */\
1940 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1941 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1942 } else { \
1943 /* we haven't inserted this word yet. */ \
1944 trie->states[ state ].wordnum = curword; \
1945 } \
1946} STMT_END
07be1b83 1947
3dab1dad 1948
786e8c11
YO
1949#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1950 ( ( base + charid >= ucharcount \
1951 && base + charid < ubound \
1952 && state == trie->trans[ base - ucharcount + charid ].check \
1953 && trie->trans[ base - ucharcount + charid ].next ) \
1954 ? trie->trans[ base - ucharcount + charid ].next \
1955 : ( state==1 ? special : 0 ) \
1956 )
3dab1dad 1957
786e8c11
YO
1958#define MADE_TRIE 1
1959#define MADE_JUMP_TRIE 2
1960#define MADE_EXACT_TRIE 4
3dab1dad 1961
a3621e74 1962STATIC I32
538e84ed
KW
1963S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1964 regnode *first, regnode *last, regnode *tail,
1965 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
1966{
1967 /* first pass, loop through and scan words */
1968 reg_trie_data *trie;
55eed653 1969 HV *widecharmap = NULL;
2b8b4781 1970 AV *revcharmap = newAV();
a3621e74 1971 regnode *cur;
a3621e74
YO
1972 STRLEN len = 0;
1973 UV uvc = 0;
1974 U16 curword = 0;
1975 U32 next_alloc = 0;
786e8c11
YO
1976 regnode *jumper = NULL;
1977 regnode *nextbranch = NULL;
7f69552c 1978 regnode *convert = NULL;
2e64971a 1979 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1980 /* we just use folder as a flag in utf8 */
1e696034 1981 const U8 * folder = NULL;
a3621e74 1982
2b8b4781 1983#ifdef DEBUGGING
cf78de0b 1984 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
1985 AV *trie_words = NULL;
1986 /* along with revcharmap, this only used during construction but both are
1987 * useful during debugging so we store them in the struct when debugging.
8e11feef 1988 */
2b8b4781 1989#else
cf78de0b 1990 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 1991 STRLEN trie_charcount=0;
3dab1dad 1992#endif
2b8b4781 1993 SV *re_trie_maxbuff;
a3621e74 1994 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1995
1996 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1997#ifndef DEBUGGING
1998 PERL_UNUSED_ARG(depth);
1999#endif
a3621e74 2000
1e696034 2001 switch (flags) {
79a81a6e 2002 case EXACT: break;
2f7f8cb1 2003 case EXACTFA:
fab2782b 2004 case EXACTFU_SS:
1e696034
KW
2005 case EXACTFU: folder = PL_fold_latin1; break;
2006 case EXACTF: folder = PL_fold; break;
fab2782b 2007 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2008 }
2009
c944940b 2010 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2011 trie->refcount = 1;
3dab1dad 2012 trie->startstate = 1;
786e8c11 2013 trie->wordcount = word_count;
f8fc2ecf 2014 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2015 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 2016 if (flags == EXACT)
c944940b 2017 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2018 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2019 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2020
a3621e74 2021 DEBUG_r({
2b8b4781 2022 trie_words = newAV();
a3621e74 2023 });
a3621e74 2024
0111c4fd 2025 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2026 assert(re_trie_maxbuff);
a3621e74 2027 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2028 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2029 }
df826430 2030 DEBUG_TRIE_COMPILE_r({
538e84ed
KW
2031 PerlIO_printf( Perl_debug_log,
2032 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2033 (int)depth * 2 + 2, "",
2034 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2035 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2036 });
538e84ed 2037
7f69552c
YO
2038 /* Find the node we are going to overwrite */
2039 if ( first == startbranch && OP( last ) != BRANCH ) {
2040 /* whole branch chain */
2041 convert = first;
2042 } else {
2043 /* branch sub-chain */
2044 convert = NEXTOPER( first );
2045 }
538e84ed 2046
a3621e74
YO
2047 /* -- First loop and Setup --
2048
2049 We first traverse the branches and scan each word to determine if it
2050 contains widechars, and how many unique chars there are, this is
2051 important as we have to build a table with at least as many columns as we
2052 have unique chars.
2053
2054 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2055 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2056 the native representation of the character value as the key and IV's for
2057 the coded index.
a3621e74
YO
2058
2059 *TODO* If we keep track of how many times each character is used we can
2060 remap the columns so that the table compression later on is more
3b753521 2061 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2062 middle and the least common are on the outside. IMO this would be better
2063 than a most to least common mapping as theres a decent chance the most
2064 common letter will share a node with the least common, meaning the node
486ec47a 2065 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2066 case is when we have the least common nodes twice.
2067
2068 */
2069
a3621e74 2070 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2071 regnode *noper = NEXTOPER( cur );
e1ec3a88 2072 const U8 *uc = (U8*)STRING( noper );
df826430 2073 const U8 *e = uc + STR_LEN( noper );
bc031a7d 2074 int foldlen = 0;
07be1b83 2075 U32 wordlen = 0; /* required init */
bc031a7d
KW
2076 STRLEN minchars = 0;
2077 STRLEN maxchars = 0;
538e84ed
KW
2078 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2079 bitmap?*/
a3621e74 2080
3dab1dad 2081 if (OP(noper) == NOTHING) {
df826430
YO
2082 regnode *noper_next= regnext(noper);
2083 if (noper_next != tail && OP(noper_next) == flags) {
2084 noper = noper_next;
2085 uc= (U8*)STRING(noper);
2086 e= uc + STR_LEN(noper);
2087 trie->minlen= STR_LEN(noper);
2088 } else {
2089 trie->minlen= 0;
2090 continue;
2091 }
3dab1dad 2092 }
df826430 2093
fab2782b 2094 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2095 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2096 regardless of encoding */
fab2782b
YO
2097 if (OP( noper ) == EXACTFU_SS) {
2098 /* false positives are ok, so just set this */
0dc4a61d 2099 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2100 }
2101 }
bc031a7d
KW
2102 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2103 branch */
3dab1dad 2104 TRIE_CHARCOUNT(trie)++;
a3621e74 2105 TRIE_READ_CHAR;
645de4ce 2106
bc031a7d
KW
2107 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2108 * is in effect. Under /i, this character can match itself, or
2109 * anything that folds to it. If not under /i, it can match just
2110 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2111 * all fold to k, and all are single characters. But some folds
2112 * expand to more than one character, so for example LATIN SMALL
2113 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2114 * the string beginning at 'uc' is 'ffi', it could be matched by
2115 * three characters, or just by the one ligature character. (It
2116 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2117 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2118 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2119 * match.) The trie needs to know the minimum and maximum number
2120 * of characters that could match so that it can use size alone to
2121 * quickly reject many match attempts. The max is simple: it is
2122 * the number of folded characters in this branch (since a fold is
2123 * never shorter than what folds to it. */
2124
2125 maxchars++;
2126
2127 /* And the min is equal to the max if not under /i (indicated by
2128 * 'folder' being NULL), or there are no multi-character folds. If
2129 * there is a multi-character fold, the min is incremented just
2130 * once, for the character that folds to the sequence. Each
2131 * character in the sequence needs to be added to the list below of
2132 * characters in the trie, but we count only the first towards the
2133 * min number of characters needed. This is done through the
2134 * variable 'foldlen', which is returned by the macros that look
2135 * for these sequences as the number of bytes the sequence
2136 * occupies. Each time through the loop, we decrement 'foldlen' by
2137 * how many bytes the current char occupies. Only when it reaches
2138 * 0 do we increment 'minchars' or look for another multi-character
2139 * sequence. */
2140 if (folder == NULL) {
2141 minchars++;
2142 }
2143 else if (foldlen > 0) {
2144 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2145 }
2146 else {
bc031a7d
KW
2147 minchars++;
2148
2149 /* See if *uc is the beginning of a multi-character fold. If
2150 * so, we decrement the length remaining to look at, to account
2151 * for the current character this iteration. (We can use 'uc'
2152 * instead of the fold returned by TRIE_READ_CHAR because for
2153 * non-UTF, the latin1_safe macro is smart enough to account
2154 * for all the unfolded characters, and because for UTF, the
2155 * string will already have been folded earlier in the
2156 * compilation process */
2157 if (UTF) {
2158 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2159 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2160 }
2161 }
bc031a7d
KW
2162 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2163 foldlen--;
2164 }
645de4ce 2165 }
bc031a7d
KW
2166
2167 /* The current character (and any potential folds) should be added
2168 * to the possible matching characters for this position in this
2169 * branch */
a3621e74 2170 if ( uvc < 256 ) {
fab2782b
YO
2171 if ( folder ) {
2172 U8 folded= folder[ (U8) uvc ];
2173 if ( !trie->charmap[ folded ] ) {
2174 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2175 TRIE_STORE_REVCHAR( folded );
2176 }
2177 }
a3621e74
YO
2178 if ( !trie->charmap[ uvc ] ) {
2179 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2180 TRIE_STORE_REVCHAR( uvc );
a3621e74 2181 }
02daf0ab 2182 if ( set_bit ) {
62012aee
KW
2183 /* store the codepoint in the bitmap, and its folded
2184 * equivalent. */
fab2782b 2185 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2186
2187 /* store the folded codepoint */
fab2782b 2188 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2189
2190 if ( !UTF ) {
2191 /* store first byte of utf8 representation of
acdf4139 2192 variant codepoints */
6f2d5cbc 2193 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2194 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2195 }
2196 }
02daf0ab
YO
2197 set_bit = 0; /* We've done our bit :-) */
2198 }
a3621e74 2199 } else {
bc031a7d
KW
2200
2201 /* XXX We could come up with the list of code points that fold
2202 * to this using PL_utf8_foldclosures, except not for
2203 * multi-char folds, as there may be multiple combinations
2204 * there that could work, which needs to wait until runtime to
2205 * resolve (The comment about LIGATURE FFI above is such an
2206 * example */
2207
a3621e74 2208 SV** svpp;
55eed653
NC
2209 if ( !widecharmap )
2210 widecharmap = newHV();
a3621e74 2211
55eed653 2212 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2213
2214 if ( !svpp )
e4584336 2215 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2216
2217 if ( !SvTRUE( *svpp ) ) {
2218 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2219 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2220 }
2221 }
bc031a7d
KW
2222 } /* end loop through characters in this branch of the trie */
2223
2224 /* We take the min and max for this branch and combine to find the min
2225 * and max for all branches processed so far */
3dab1dad 2226 if( cur == first ) {
bc031a7d
KW
2227 trie->minlen = minchars;
2228 trie->maxlen = maxchars;
2229 } else if (minchars < trie->minlen) {
2230 trie->minlen = minchars;
2231 } else if (maxchars > trie->maxlen) {
2232 trie->maxlen = maxchars;
fab2782b 2233 }
a3621e74
YO
2234 } /* end first pass */
2235 DEBUG_TRIE_COMPILE_r(
538e84ed
KW
2236 PerlIO_printf( Perl_debug_log,
2237 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3dab1dad 2238 (int)depth * 2 + 2,"",
55eed653 2239 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2240 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2241 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2242 );
a3621e74
YO
2243
2244 /*
2245 We now know what we are dealing with in terms of unique chars and
2246 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2247 representation using a flat table will take. If it's over a reasonable
2248 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2249 conservative but potentially much slower representation using an array
2250 of lists.
2251
2252 At the end we convert both representations into the same compressed
2253 form that will be used in regexec.c for matching with. The latter
2254 is a form that cannot be used to construct with but has memory
2255 properties similar to the list form and access properties similar
2256 to the table form making it both suitable for fast searches and
2257 small enough that its feasable to store for the duration of a program.
2258
2259 See the comment in the code where the compressed table is produced
2260 inplace from the flat tabe representation for an explanation of how
2261 the compression works.
2262
2263 */
2264
2265
2e64971a
DM
2266 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2267 prev_states[1] = 0;
2268
538e84ed
KW
2269 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2270 > SvIV(re_trie_maxbuff) )
2271 {
a3621e74
YO
2272 /*
2273 Second Pass -- Array Of Lists Representation
2274
2275 Each state will be represented by a list of charid:state records
2276 (reg_trie_trans_le) the first such element holds the CUR and LEN
2277 points of the allocated array. (See defines above).
2278
2279 We build the initial structure using the lists, and then convert
2280 it into the compressed table form which allows faster lookups
2281 (but cant be modified once converted).
a3621e74
YO
2282 */
2283
a3621e74
YO
2284 STRLEN transcount = 1;
2285
538e84ed 2286 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2287 "%*sCompiling trie using list compiler\n",
2288 (int)depth * 2 + 2, ""));
686b73d4 2289
c944940b
JH
2290 trie->states = (reg_trie_state *)
2291 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2292 sizeof(reg_trie_state) );
a3621e74
YO
2293 TRIE_LIST_NEW(1);
2294 next_alloc = 2;
2295
2296 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2297
df826430 2298 regnode *noper = NEXTOPER( cur );
c445ea15 2299 U8 *uc = (U8*)STRING( noper );
df826430 2300 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
2301 U32 state = 1; /* required init */
2302 U16 charid = 0; /* sanity init */
07be1b83 2303 U32 wordlen = 0; /* required init */
c445ea15 2304
df826430
YO
2305 if (OP(noper) == NOTHING) {
2306 regnode *noper_next= regnext(noper);
2307 if (noper_next != tail && OP(noper_next) == flags) {
2308 noper = noper_next;
2309 uc= (U8*)STRING(noper);
2310 e= uc + STR_LEN(noper);
2311 }
2312 }
2313
3dab1dad 2314 if (OP(noper) != NOTHING) {
786e8c11 2315 for ( ; uc < e ; uc += len ) {
c445ea15 2316
786e8c11 2317 TRIE_READ_CHAR;
c445ea15 2318
786e8c11
YO
2319 if ( uvc < 256 ) {
2320 charid = trie->charmap[ uvc ];
c445ea15 2321 } else {
538e84ed
KW
2322 SV** const svpp = hv_fetch( widecharmap,
2323 (char*)&uvc,
2324 sizeof( UV ),
2325 0);
786e8c11
YO
2326 if ( !svpp ) {
2327 charid = 0;
2328 } else {
2329 charid=(U16)SvIV( *svpp );
2330 }
c445ea15 2331 }
538e84ed
KW
2332 /* charid is now 0 if we dont know the char read, or
2333 * nonzero if we do */
786e8c11 2334 if ( charid ) {
a3621e74 2335
786e8c11
YO
2336 U16 check;
2337 U32 newstate = 0;
a3621e74 2338
786e8c11
YO
2339 charid--;
2340 if ( !trie->states[ state ].trans.list ) {
2341 TRIE_LIST_NEW( state );
c445ea15 2342 }
538e84ed
KW
2343 for ( check = 1;
2344 check <= TRIE_LIST_USED( state );
2345 check++ )
2346 {
2347 if ( TRIE_LIST_ITEM( state, check ).forid
2348 == charid )
2349 {
786e8c11
YO
2350 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2351 break;
2352 }
2353 }
2354 if ( ! newstate ) {
2355 newstate = next_alloc++;
2e64971a 2356 prev_states[newstate] = state;
786e8c11
YO
2357 TRIE_LIST_PUSH( state, charid, newstate );
2358 transcount++;
2359 }
2360 state = newstate;
2361 } else {
2362 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2363 }
a28509cc 2364 }
c445ea15 2365 }
3dab1dad 2366 TRIE_HANDLE_WORD(state);
a3621e74
YO
2367
2368 } /* end second pass */
2369
1e2e3d02 2370 /* next alloc is the NEXT state to be allocated */
538e84ed 2371 trie->statecount = next_alloc;
c944940b
JH
2372 trie->states = (reg_trie_state *)
2373 PerlMemShared_realloc( trie->states,
2374 next_alloc
2375 * sizeof(reg_trie_state) );
a3621e74 2376
3dab1dad 2377 /* and now dump it out before we compress it */
2b8b4781
NC
2378 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2379 revcharmap, next_alloc,
2380 depth+1)
1e2e3d02 2381 );
a3621e74 2382
c944940b
JH
2383 trie->trans = (reg_trie_trans *)
2384 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2385 {
2386 U32 state;
a3621e74
YO
2387 U32 tp = 0;
2388 U32 zp = 0;
2389
2390
2391 for( state=1 ; state < next_alloc ; state ++ ) {
2392 U32 base=0;
2393
2394 /*
2395 DEBUG_TRIE_COMPILE_MORE_r(
2396 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2397 );
2398 */
2399
2400 if (trie->states[state].trans.list) {
2401 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2402 U16 maxid=minid;
a28509cc 2403 U16 idx;
a3621e74
YO
2404
2405 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2406 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2407 if ( forid < minid ) {
2408 minid=forid;
2409 } else if ( forid > maxid ) {
2410 maxid=forid;
2411 }
a3621e74
YO
2412 }
2413 if ( transcount < tp + maxid - minid + 1) {
2414 transcount *= 2;
c944940b
JH
2415 trie->trans = (reg_trie_trans *)
2416 PerlMemShared_realloc( trie->trans,
446bd890
NC
2417 transcount
2418 * sizeof(reg_trie_trans) );
538e84ed
KW
2419 Zero( trie->trans + (transcount / 2),
2420 transcount / 2,
2421 reg_trie_trans );
a3621e74
YO
2422 }
2423 base = trie->uniquecharcount + tp - minid;
2424 if ( maxid == minid ) {
2425 U32 set = 0;
2426 for ( ; zp < tp ; zp++ ) {
2427 if ( ! trie->trans[ zp ].next ) {
2428 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2429 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2430 1).newstate;
a3621e74
YO
2431 trie->trans[ zp ].check = state;
2432 set = 1;
2433 break;
2434 }
2435 }
2436 if ( !set ) {
538e84ed
KW
2437 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2438 1).newstate;
a3621e74
YO
2439 trie->trans[ tp ].check = state;
2440 tp++;
2441 zp = tp;
2442 }
2443 } else {
2444 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2445 const U32 tid = base
2446 - trie->uniquecharcount
2447 + TRIE_LIST_ITEM( state, idx ).forid;
2448 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2449 idx ).newstate;
a3621e74
YO
2450 trie->trans[ tid ].check = state;
2451 }
2452 tp += ( maxid - minid + 1 );
2453 }
2454 Safefree(trie->states[ state ].trans.list);
2455 }
2456 /*
2457 DEBUG_TRIE_COMPILE_MORE_r(
2458 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2459 );
2460 */
2461 trie->states[ state ].trans.base=base;
2462 }
cc601c31 2463 trie->lasttrans = tp + 1;
a3621e74
YO
2464 }
2465 } else {
2466 /*
2467 Second Pass -- Flat Table Representation.
2468
b423522f
KW
2469 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2470 each. We know that we will need Charcount+1 trans at most to store
2471 the data (one row per char at worst case) So we preallocate both
2472 structures assuming worst case.
a3621e74
YO
2473
2474 We then construct the trie using only the .next slots of the entry
2475 structs.
2476
b423522f
KW
2477 We use the .check field of the first entry of the node temporarily
2478 to make compression both faster and easier by keeping track of how
2479 many non zero fields are in the node.
a3621e74
YO
2480
2481 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2482 transition.
2483
b423522f
KW
2484 There are two terms at use here: state as a TRIE_NODEIDX() which is
2485 a number representing the first entry of the node, and state as a
2486 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2487 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2488 if there are 2 entrys per node. eg:
a3621e74
YO
2489
2490 A B A B
2491 1. 2 4 1. 3 7
2492 2. 0 3 3. 0 5
2493 3. 0 0 5. 0 0
2494 4. 0 0 7. 0 0
2495
b423522f
KW
2496 The table is internally in the right hand, idx form. However as we
2497 also have to deal with the states array which is indexed by nodenum
2498 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2499
2500 */
538e84ed 2501 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2502 "%*sCompiling trie using table compiler\n",
2503 (int)depth * 2 + 2, ""));
3dab1dad 2504
c944940b
JH
2505 trie->trans = (reg_trie_trans *)
2506 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2507 * trie->uniquecharcount + 1,
2508 sizeof(reg_trie_trans) );
2509 trie->states = (reg_trie_state *)
2510 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2511 sizeof(reg_trie_state) );
a3621e74
YO
2512 next_alloc = trie->uniquecharcount + 1;
2513
3dab1dad 2514
a3621e74
YO
2515 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2516
df826430 2517 regnode *noper = NEXTOPER( cur );
a28509cc 2518 const U8 *uc = (U8*)STRING( noper );
df826430 2519 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2520
2521 U32 state = 1; /* required init */
2522
2523 U16 charid = 0; /* sanity init */
2524 U32 accept_state = 0; /* sanity init */
a3621e74 2525
07be1b83 2526 U32 wordlen = 0; /* required init */
a3621e74 2527
df826430
YO
2528 if (OP(noper) == NOTHING) {
2529 regnode *noper_next= regnext(noper);
2530 if (noper_next != tail && OP(noper_next) == flags) {
2531 noper = noper_next;
2532 uc= (U8*)STRING(noper);
2533 e= uc + STR_LEN(noper);
2534 }
2535 }
fab2782b 2536
3dab1dad 2537 if ( OP(noper) != NOTHING ) {
786e8c11 2538 for ( ; uc < e ; uc += len ) {
a3621e74 2539
786e8c11 2540 TRIE_READ_CHAR;
a3621e74 2541
786e8c11
YO
2542 if ( uvc < 256 ) {
2543 charid = trie->charmap[ uvc ];
2544 } else {
538e84ed
KW
2545 SV* const * const svpp = hv_fetch( widecharmap,
2546 (char*)&uvc,
2547 sizeof( UV ),
2548 0);
786e8c11 2549 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2550 }
786e8c11
YO
2551 if ( charid ) {
2552 charid--;
2553 if ( !trie->trans[ state + charid ].next ) {
2554 trie->trans[ state + charid ].next = next_alloc;
2555 trie->trans[ state ].check++;
2e64971a
DM
2556 prev_states[TRIE_NODENUM(next_alloc)]
2557 = TRIE_NODENUM(state);
786e8c11
YO
2558 next_alloc += trie->uniquecharcount;
2559 }
2560 state = trie->trans[ state + charid ].next;
2561 } else {
2562 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2563 }
538e84ed
KW
2564 /* charid is now 0 if we dont know the char read, or
2565 * nonzero if we do */
a3621e74 2566 }
a3621e74 2567 }
3dab1dad
YO
2568 accept_state = TRIE_NODENUM( state );
2569 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2570
2571 } /* end second pass */
2572
3dab1dad 2573 /* and now dump it out before we compress it */
2b8b4781
NC
2574 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2575 revcharmap,
2576 next_alloc, depth+1));
a3621e74 2577
a3621e74
YO
2578 {
2579 /*
2580 * Inplace compress the table.*
2581
2582 For sparse data sets the table constructed by the trie algorithm will
2583 be mostly 0/FAIL transitions or to put it another way mostly empty.
2584 (Note that leaf nodes will not contain any transitions.)
2585
2586 This algorithm compresses the tables by eliminating most such
2587 transitions, at the cost of a modest bit of extra work during lookup:
2588
2589 - Each states[] entry contains a .base field which indicates the
2590 index in the state[] array wheres its transition data is stored.
2591
3b753521 2592 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2593
2594 - If .base is nonzero then charid is added to it to find an entry in
2595 the trans array.
2596
2597 -If trans[states[state].base+charid].check!=state then the
2598 transition is taken to be a 0/Fail transition. Thus if there are fail
2599 transitions at the front of the node then the .base offset will point
2600 somewhere inside the previous nodes data (or maybe even into a node
2601 even earlier), but the .check field determines if the transition is
2602 valid.
2603
786e8c11 2604 XXX - wrong maybe?
a3621e74 2605 The following process inplace converts the table to the compressed
3b753521 2606 table: We first do not compress the root node 1,and mark all its
a3621e74 2607 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2608 allows us to do a DFA construction from the compressed table later,
2609 and ensures that any .base pointers we calculate later are greater
2610 than 0.
a3621e74
YO
2611
2612 - We set 'pos' to indicate the first entry of the second node.
2613
2614 - We then iterate over the columns of the node, finding the first and
2615 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2616 and set the .check pointers accordingly, and advance pos
2617 appropriately and repreat for the next node. Note that when we copy
2618 the next pointers we have to convert them from the original
2619 NODEIDX form to NODENUM form as the former is not valid post
2620 compression.
2621
2622 - If a node has no transitions used we mark its base as 0 and do not
2623 advance the pos pointer.
2624
2625 - If a node only has one transition we use a second pointer into the
2626 structure to fill in allocated fail transitions from other states.
2627 This pointer is independent of the main pointer and scans forward
2628 looking for null transitions that are allocated to a state. When it
2629 finds one it writes the single transition into the "hole". If the
786e8c11 2630 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2631
2632 - Once compressed we can Renew/realloc the structures to release the
2633 excess space.
2634
2635 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2636 specifically Fig 3.47 and the associated pseudocode.
2637
2638 demq
2639 */
a3b680e6 2640 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2641 U32 state, charid;
a3621e74 2642 U32 pos = 0, zp=0;
1e2e3d02 2643 trie->statecount = laststate;
a3621e74
YO
2644
2645 for ( state = 1 ; state < laststate ; state++ ) {
2646 U8 flag = 0;
a28509cc
AL
2647 const U32 stateidx = TRIE_NODEIDX( state );
2648 const U32 o_used = trie->trans[ stateidx ].check;
2649 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2650 trie->trans[ stateidx ].check = 0;
2651
538e84ed
KW
2652 for ( charid = 0;
2653 used && charid < trie->uniquecharcount;
2654 charid++ )
2655 {
a3621e74
YO
2656 if ( flag || trie->trans[ stateidx + charid ].next ) {
2657 if ( trie->trans[ stateidx + charid ].next ) {
2658 if (o_used == 1) {
2659 for ( ; zp < pos ; zp++ ) {
2660 if ( ! trie->trans[ zp ].next ) {
2661 break;
2662 }
2663 }
538e84ed
KW
2664 trie->states[ state ].trans.base
2665 = zp
2666 + trie->uniquecharcount
2667 - charid ;
2668 trie->trans[ zp ].next
2669 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2670 + charid ].next );
a3621e74
YO
2671 trie->trans[ zp ].check = state;
2672 if ( ++zp > pos ) pos = zp;
2673 break;
2674 }
2675 used--;
2676 }
2677 if ( !flag ) {
2678 flag = 1;
538e84ed
KW
2679 trie->states[ state ].trans.base
2680 = pos + trie->uniquecharcount - charid ;
a3621e74 2681 }
538e84ed
KW
2682 trie->trans[ pos ].next
2683 = SAFE_TRIE_NODENUM(
2684 trie->trans[ stateidx + charid ].next );
a3621e74
YO
2685 trie->trans[ pos ].check = state;
2686 pos++;
2687 }
2688 }
2689 }
cc601c31 2690 trie->lasttrans = pos + 1;
c944940b
JH
2691 trie->states = (reg_trie_state *)
2692 PerlMemShared_realloc( trie->states, laststate
2693 * sizeof(reg_trie_state) );
a3621e74 2694 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2695 PerlIO_printf( Perl_debug_log,
2696 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2697 (int)depth * 2 + 2,"",
2698 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2699 + 1 ),
2700 (IV)next_alloc,
2701 (IV)pos,
2702 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
2703 );
2704
2705 } /* end table compress */
2706 }
1e2e3d02 2707 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2708 PerlIO_printf(Perl_debug_log,
2709 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1e2e3d02
YO
2710 (int)depth * 2 + 2, "",
2711 (UV)trie->statecount,
2712 (UV)trie->lasttrans)
2713 );
cc601c31 2714 /* resize the trans array to remove unused space */
c944940b
JH
2715 trie->trans = (reg_trie_trans *)
2716 PerlMemShared_realloc( trie->trans, trie->lasttrans
2717 * sizeof(reg_trie_trans) );
a3621e74 2718
538e84ed 2719 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2720 U8 nodetype =(U8)(flags & 0xFF);
2721 char *str=NULL;
538e84ed 2722
07be1b83 2723#ifdef DEBUGGING
e62cc96a 2724 regnode *optimize = NULL;
7122b237
YO
2725#ifdef RE_TRACK_PATTERN_OFFSETS
2726
b57a0404
JH
2727 U32 mjd_offset = 0;
2728 U32 mjd_nodelen = 0;
7122b237
YO
2729#endif /* RE_TRACK_PATTERN_OFFSETS */
2730#endif /* DEBUGGING */
a3621e74 2731 /*
3dab1dad
YO
2732 This means we convert either the first branch or the first Exact,
2733 depending on whether the thing following (in 'last') is a branch
2734 or not and whther first is the startbranch (ie is it a sub part of
2735 the alternation or is it the whole thing.)
3b753521 2736 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2737 the whole branch sequence, including the first.
a3621e74 2738 */
3dab1dad 2739 /* Find the node we are going to overwrite */
7f69552c 2740 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2741 /* branch sub-chain */
3dab1dad 2742 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2743#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2744 DEBUG_r({
2745 mjd_offset= Node_Offset((convert));
2746 mjd_nodelen= Node_Length((convert));
2747 });
7122b237 2748#endif
7f69552c 2749 /* whole branch chain */
7122b237
YO
2750 }
2751#ifdef RE_TRACK_PATTERN_OFFSETS
2752 else {
7f69552c
YO
2753 DEBUG_r({
2754 const regnode *nop = NEXTOPER( convert );
2755 mjd_offset= Node_Offset((nop));
2756 mjd_nodelen= Node_Length((nop));
2757 });
07be1b83
YO
2758 }
2759 DEBUG_OPTIMISE_r(
538e84ed
KW
2760 PerlIO_printf(Perl_debug_log,
2761 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
07be1b83 2762 (int)depth * 2 + 2, "",
786e8c11 2763 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2764 );
7122b237 2765#endif
538e84ed 2766 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
2767 split out as an EXACT and put in front of the TRIE node. */
2768 trie->startstate= 1;
55eed653 2769 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2770 U32 state;
1e2e3d02 2771 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2772 U32 ofs = 0;
8e11feef
RGS
2773 I32 idx = -1;
2774 U32 count = 0;
2775 const U32 base = trie->states[ state ].trans.base;
a3621e74 2776
3dab1dad 2777 if ( trie->states[state].wordnum )
8e11feef 2778 count = 1;
a3621e74 2779
8e11feef 2780 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2781 if ( ( base + ofs >= trie->uniquecharcount ) &&
2782 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2783 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2784 {
3dab1dad 2785 if ( ++count > 1 ) {
2b8b4781 2786 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2787 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2788 if ( state == 1 ) break;
3dab1dad
YO
2789 if ( count == 2 ) {
2790 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2791 DEBUG_OPTIMISE_r(
8e11feef
RGS
2792 PerlIO_printf(Perl_debug_log,
2793 "%*sNew Start State=%"UVuf" Class: [",
2794 (int)depth * 2 + 2, "",
786e8c11 2795 (UV)state));
be8e71aa 2796 if (idx >= 0) {
2b8b4781 2797 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2798 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2799
3dab1dad 2800 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2801 if ( folder )
2802 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2803 DEBUG_OPTIMISE_r(
f1f66076 2804 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2805 );
8e11feef
RGS
2806 }
2807 }
2808 TRIE_BITMAP_SET(trie,*ch);
2809 if ( folder )
2810 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2811 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2812 }
2813 idx = ofs;
2814 }
3dab1dad
YO
2815 }
2816 if ( count == 1 ) {
2b8b4781 2817 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2818 STRLEN len;
2819 char *ch = SvPV( *tmp, len );
de734bd5
A
2820 DEBUG_OPTIMISE_r({
2821 SV *sv=sv_newmortal();
8e11feef
RGS
2822 PerlIO_printf( Perl_debug_log,
2823 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2824 (int)depth * 2 + 2, "",
538e84ed
KW
2825 (UV)state, (UV)idx,
2826 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
2827 PL_colors[0], PL_colors[1],
2828 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 2829 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
2830 )
2831 );
2832 });
3dab1dad
YO
2833 if ( state==1 ) {
2834 OP( convert ) = nodetype;
2835 str=STRING(convert);
2836 STR_LEN(convert)=0;
2837 }
c490c714
YO
2838 STR_LEN(convert) += len;
2839 while (len--)
de734bd5 2840 *str++ = *ch++;
8e11feef 2841 } else {
538e84ed 2842#ifdef DEBUGGING
8e11feef
RGS
2843 if (state>1)
2844 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2845#endif
8e11feef
RGS
2846 break;
2847 }
2848 }
2e64971a 2849 trie->prefixlen = (state-1);
3dab1dad 2850 if (str) {
8e11feef 2851 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2852 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2853 trie->startstate = state;
07be1b83
YO
2854 trie->minlen -= (state - 1);
2855 trie->maxlen -= (state - 1);
33809eae
JH
2856#ifdef DEBUGGING
2857 /* At least the UNICOS C compiler choked on this
2858 * being argument to DEBUG_r(), so let's just have
2859 * it right here. */
2860 if (
2861#ifdef PERL_EXT_RE_BUILD
2862 1
2863#else
2864 DEBUG_r_TEST
2865#endif
2866 ) {
2867 regnode *fix = convert;
2868 U32 word = trie->wordcount;
2869 mjd_nodelen++;
2870 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2871 while( ++fix < n ) {
2872 Set_Node_Offset_Length(fix, 0, 0);
2873 }
2874 while (word--) {
2875 SV ** const tmp = av_fetch( trie_words, word, 0 );
2876 if (tmp) {
2877 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2878 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2879 else
2880 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2881 }
2882 }
2883 }
2884#endif
8e11feef
RGS
2885 if (trie->maxlen) {
2886 convert = n;
2887 } else {
3dab1dad 2888 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2889 DEBUG_r(optimize= n);
3dab1dad
YO
2890 }
2891 }
2892 }
538e84ed
KW
2893 if (!jumper)
2894 jumper = last;
3dab1dad 2895 if ( trie->maxlen ) {
8e11feef
RGS
2896 NEXT_OFF( convert ) = (U16)(tail - convert);
2897 ARG_SET( convert, data_slot );
538e84ed
KW
2898 /* Store the offset to the first unabsorbed branch in
2899 jump[0], which is otherwise unused by the jump logic.
786e8c11 2900 We use this when dumping a trie and during optimisation. */
538e84ed 2901 if (trie->jump)
7f69552c 2902 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 2903
6c48061a
YO
2904 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2905 * and there is a bitmap
2906 * and the first "jump target" node we found leaves enough room
2907 * then convert the TRIE node into a TRIEC node, with the bitmap
2908 * embedded inline in the opcode - this is hypothetically faster.
2909 */
2910 if ( !trie->states[trie->startstate].wordnum
2911 && trie->bitmap
2912 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2913 {
2914 OP( convert ) = TRIEC;
2915 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2916 PerlMemShared_free(trie->bitmap);
786e8c11 2917 trie->bitmap= NULL;
538e84ed 2918 } else
786e8c11 2919 OP( convert ) = TRIE;
a3621e74 2920
3dab1dad
YO
2921 /* store the type in the flags */
2922 convert->flags = nodetype;
a5ca303d 2923 DEBUG_r({
538e84ed
KW
2924 optimize = convert
2925 + NODE_STEP_REGNODE
a5ca303d
YO
2926 + regarglen[ OP( convert ) ];
2927 });
538e84ed 2928 /* XXX We really should free up the resource in trie now,
a5ca303d 2929 as we won't use them - (which resources?) dmq */
3dab1dad 2930 }
a3621e74 2931 /* needed for dumping*/
e62cc96a 2932 DEBUG_r(if (optimize) {
07be1b83 2933 regnode *opt = convert;
bcdf7404 2934
e62cc96a 2935 while ( ++opt < optimize) {
07be1b83
YO
2936 Set_Node_Offset_Length(opt,0,0);
2937 }
538e84ed
KW
2938 /*
2939 Try to clean up some of the debris left after the
786e8c11 2940 optimisation.
a3621e74 2941 */
786e8c11 2942 while( optimize < jumper ) {
07be1b83 2943 mjd_nodelen += Node_Length((optimize));
a3621e74 2944 OP( optimize ) = OPTIMIZED;
07be1b83 2945 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2946 optimize++;
2947 }
07be1b83 2948 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2949 });
2950 } /* end node insert */
2e64971a
DM
2951
2952 /* Finish populating the prev field of the wordinfo array. Walk back
2953 * from each accept state until we find another accept state, and if
2954 * so, point the first word's .prev field at the second word. If the
2955 * second already has a .prev field set, stop now. This will be the
2956 * case either if we've already processed that word's accept state,
3b753521
FN
2957 * or that state had multiple words, and the overspill words were
2958 * already linked up earlier.
2e64971a
DM
2959 */
2960 {
2961 U16 word;
2962 U32 state;
2963 U16 prev;
2964
2965 for (word=1; word <= trie->wordcount; word++) {
2966 prev = 0;
2967 if (trie->wordinfo[word].prev)
2968 continue;
2969 state = trie->wordinfo[word].accept;
2970 while (state) {
2971 state = prev_states[state];
2972 if (!state)
2973 break;
2974 prev = trie->states[state].wordnum;
2975 if (prev)
2976 break;
2977 }
2978 trie->wordinfo[word].prev = prev;
2979 }
2980 Safefree(prev_states);
2981 }
2982
2983
2984 /* and now dump out the compressed format */
2985 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2986
55eed653 2987 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2988#ifdef DEBUGGING
2989 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2990 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2991#else
03e70be4 2992 SvREFCNT_dec_NN(revcharmap);
07be1b83 2993#endif
538e84ed
KW
2994 return trie->jump
2995 ? MADE_JUMP_TRIE
2996 : trie->startstate>1
2997 ? MADE_EXACT_TRIE
786e8c11
YO
2998 : MADE_TRIE;
2999}
3000
615a2e7f
YO
3001STATIC regnode *
3002S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
786e8c11 3003{
b423522f
KW
3004/* The Trie is constructed and compressed now so we can build a fail array if
3005 * it's needed
786e8c11 3006
b423522f
KW
3007 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3008 3.32 in the
3009 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3010 Ullman 1985/88
786e8c11
YO
3011 ISBN 0-201-10088-6
3012
b423522f
KW
3013 We find the fail state for each state in the trie, this state is the longest
3014 proper suffix of the current state's 'word' that is also a proper prefix of
3015 another word in our trie. State 1 represents the word '' and is thus the
3016 default fail state. This allows the DFA not to have to restart after its
3017 tried and failed a word at a given point, it simply continues as though it
3018 had been matching the other word in the first place.
786e8c11
YO
3019 Consider
3020 'abcdgu'=~/abcdefg|cdgu/
b423522f
KW
3021 When we get to 'd' we are still matching the first word, we would encounter
3022 'g' which would fail, which would bring us to the state representing 'd' in
3023 the second word where we would try 'g' and succeed, proceeding to match
3024 'cdgu'.
786e8c11
YO
3025 */
3026 /* add a fail transition */
3251b653
NC
3027 const U32 trie_offset = ARG(source);
3028 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
3029 U32 *q;
3030 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 3031 const U32 numstates = trie->statecount;
786e8c11
YO
3032 const U32 ubound = trie->lasttrans + ucharcount;
3033 U32 q_read = 0;
3034 U32 q_write = 0;
3035 U32 charid;
3036 U32 base = trie->states[ 1 ].trans.base;
3037 U32 *fail;
3038 reg_ac_data *aho;
cf78de0b 3039 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
615a2e7f 3040 regnode *stclass;
786e8c11 3041 GET_RE_DEBUG_FLAGS_DECL;
7918f24d 3042
615a2e7f 3043 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
81611534 3044 PERL_UNUSED_CONTEXT;
786e8c11
YO
3045#ifndef DEBUGGING
3046 PERL_UNUSED_ARG(depth);
3047#endif
3048
615a2e7f
YO
3049 if ( OP(source) == TRIE ) {
3050 struct regnode_1 *op = (struct regnode_1 *)
3051 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3052 StructCopy(source,op,struct regnode_1);
3053 stclass = (regnode *)op;
3054 } else {
3055 struct regnode_charclass *op = (struct regnode_charclass *)
3056 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3057 StructCopy(source,op,struct regnode_charclass);
3058 stclass = (regnode *)op;
3059 }
3060 OP(stclass)+=2; /* covert the TRIE type to its AHO-CORASICK equivalent */
786e8c11
YO
3061
3062 ARG_SET( stclass, data_slot );
c944940b 3063 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 3064 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 3065 aho->trie=trie_offset;
446bd890
NC
3066 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3067 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 3068 Newxz( q, numstates, U32);
c944940b 3069 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
3070 aho->refcount = 1;
3071 fail = aho->fail;
3072 /* initialize fail[0..1] to be 1 so that we always have
3073 a valid final fail state */
3074 fail[ 0 ] = fail[ 1 ] = 1;
3075
3076 for ( charid = 0; charid < ucharcount ; charid++ ) {
3077 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3078 if ( newstate ) {
3079 q[ q_write ] = newstate;
3080 /* set to point at the root */
3081 fail[ q[ q_write++ ] ]=1;
3082 }
3083 }
3084 while ( q_read < q_write) {
3085 const U32 cur = q[ q_read++ % numstates ];
3086 base = trie->states[ cur ].trans.base;
3087
3088 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3089 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3090 if (ch_state) {
3091 U32 fail_state = cur;
3092 U32 fail_base;
3093 do {
3094 fail_state = fail[ fail_state ];
3095 fail_base = aho->states[ fail_state ].trans.base;
3096 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3097
3098 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3099 fail[ ch_state ] = fail_state;
3100 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3101 {
3102 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3103 }
3104 q[ q_write++ % numstates] = ch_state;
3105 }
3106 }
3107 }
3108 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3109 when we fail in state 1, this allows us to use the
3110 charclass scan to find a valid start char. This is based on the principle
3111 that theres a good chance the string being searched contains lots of stuff
3112 that cant be a start char.
3113 */
3114 fail[ 0 ] = fail[ 1 ] = 0;
3115 DEBUG_TRIE_COMPILE_r({
6d99fb9b 3116 PerlIO_printf(Perl_debug_log,
538e84ed 3117 "%*sStclass Failtable (%"UVuf" states): 0",
6d99fb9b 3118 (int)(depth * 2), "", (UV)numstates
1e2e3d02 3119 );
786e8c11
YO
3120 for( q_read=1; q_read<numstates; q_read++ ) {
3121 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3122 }
3123 PerlIO_printf(Perl_debug_log, "\n");
3124 });
3125 Safefree(q);
e384d5c1 3126 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
615a2e7f 3127 return stclass;
a3621e74
YO
3128}
3129
786e8c11 3130
07be1b83 3131#define DEBUG_PEEP(str,scan,depth) \
b515a41d 3132 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
3133 SV * const mysv=sv_newmortal(); \
3134 regnode *Next = regnext(scan); \
2395827c 3135 regprop(RExC_rx, mysv, scan, NULL); \
7f69552c 3136 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
3137 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3138 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 3139 }});
07be1b83 3140
1de06328 3141
bb914485 3142/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
3143 * one. The regop may be changed if the node(s) contain certain sequences that
3144 * require special handling. The joining is only done if:
bb914485
KW
3145 * 1) there is room in the current conglomerated node to entirely contain the
3146 * next one.
3147 * 2) they are the exact same node type
3148 *
87b8b349 3149 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
3150 * these get optimized out
3151 *
0a982f06
KW
3152 * If a node is to match under /i (folded), the number of characters it matches
3153 * can be different than its character length if it contains a multi-character
31f05a37
KW
3154 * fold. *min_subtract is set to the total delta number of characters of the
3155 * input nodes.
bb914485 3156 *
fb29cc72 3157 * And *unfolded_multi_char is set to indicate whether or not the node contains
31f05a37
KW
3158 * an unfolded multi-char fold. This happens when whether the fold is valid or
3159 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3160 * SMALL LETTER SHARP S, as only if the target string being matched against
3161 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3162 * folding rules depend on the locale in force at runtime. (Multi-char folds
3163 * whose components are all above the Latin1 range are not run-time locale
3164 * dependent, and have already been folded by the time this function is
3165 * called.)
f758bddf 3166 *
bb914485 3167 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
3168 * multi-character fold sequences. It's been wrong in Perl for a very long
3169 * time. There are three code points in Unicode whose multi-character folds
3170 * were long ago discovered to mess things up. The previous designs for
3171 * dealing with these involved assigning a special node for them. This
538e84ed 3172 * approach doesn't always work, as evidenced by this example:
a0c4c608 3173 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
538e84ed 3174 * Both sides fold to "sss", but if the pattern is parsed to create a node that
0a982f06 3175 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
3176 * successful match would have to cross the node's boundary. The new approach
3177 * that hopefully generally solves the problem generates an EXACTFU_SS node
538e84ed 3178 * that is "sss" in this case.
bb914485 3179 *
0a982f06 3180 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
3181 * just these three. Now the code is general, for all such cases. The
3182 * approach taken is:
0a982f06 3183 * 1) This routine examines each EXACTFish node that could contain multi-
538e84ed
KW
3184 * character folded sequences. Since a single character can fold into
3185 * such a sequence, the minimum match length for this node is less than
3186 * the number of characters in the node. This routine returns in
31f05a37
KW
3187 * *min_subtract how many characters to subtract from the the actual
3188 * length of the string to get a real minimum match length; it is 0 if
3189 * there are no multi-char foldeds. This delta is used by the caller to
3190 * adjust the min length of the match, and the delta between min and max,
3191 * so that the optimizer doesn't reject these possibilities based on size
3192 * constraints.
cb117658 3193 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
3194 * is used for an EXACTFU node that contains at least one "ss" sequence in
3195 * it. For non-UTF-8 patterns and strings, this is the only case where
3196 * there is a possible fold length change. That means that a regular
3197 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3198 * with length changes, and so can be processed faster. regexec.c takes
3199 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
31f05a37
KW
3200 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3201 * known until runtime). This saves effort in regex matching. However,
3202 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3203 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3204 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3205 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
0a982f06
KW
3206 * possibilities for the non-UTF8 patterns are quite simple, except for
3207 * the sharp s. All the ones that don't involve a UTF-8 target string are
3208 * members of a fold-pair, and arrays are set up for all of them so that
3209 * the other member of the pair can be found quickly. Code elsewhere in
3210 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3211 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3212 * described in the next item.
31f05a37
KW
3213 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3214 * validity of the fold won't be known until runtime, and so must remain
3215 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3216 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3217 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3218 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3219 * The reason this is a problem is that the optimizer part of regexec.c
3220 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3221 * that a character in the pattern corresponds to at most a single
3222 * character in the target string. (And I do mean character, and not byte
3223 * here, unlike other parts of the documentation that have never been
3224 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3225 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3226 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3227 * nodes, violate the assumption, and they are the only instances where it
3228 * is violated. I'm reluctant to try to change the assumption, as the
3229 * code involved is impenetrable to me (khw), so instead the code here
3230 * punts. This routine examines EXACTFL nodes, and (when the pattern
3231 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3232 * boolean indicating whether or not the node contains such a fold. When
3233 * it is true, the caller sets a flag that later causes the optimizer in
3234 * this file to not set values for the floating and fixed string lengths,
3235 * and thus avoids the optimizer code in regexec.c that makes the invalid
1ca267a5 3236 * assumption. Thus, there is no optimization based on string lengths for
31f05a37
KW
3237 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3238 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3239 * assumption is wrong only in these cases is that all other non-UTF-8
3240 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3241 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3242 * EXACTF nodes because we don't know at compile time if it actually
3243 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3244 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3245 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3246 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3247 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3248 * string would require the pattern to be forced into UTF-8, the overhead
3249 * of which we want to avoid. Similarly the unfolded multi-char folds in
3250 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3251 * locale.)
098b07d5
KW
3252 *
3253 * Similarly, the code that generates tries doesn't currently handle
3254 * not-already-folded multi-char folds, and it looks like a pain to change
3255 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3256 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3257 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3258 * using /iaa matching will be doing so almost entirely with ASCII
3259 * strings, so this should rarely be encountered in practice */
1de06328 3260
fb29cc72 3261#define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
07be1b83 3262 if (PL_regkind[OP(scan)] == EXACT) \
fb29cc72 3263 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
07be1b83 3264
be8e71aa 3265STATIC U32
538e84ed 3266S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
fb29cc72 3267 UV *min_subtract, bool *unfolded_multi_char,
538e84ed
KW
3268 U32 flags,regnode *val, U32 depth)
3269{
07be1b83
YO
3270 /* Merge several consecutive EXACTish nodes into one. */
3271 regnode *n = regnext(scan);
3272 U32 stringok = 1;
3273 regnode *next = scan + NODE_SZ_STR(scan);
3274 U32 merged = 0;
3275 U32 stopnow = 0;
3276#ifdef DEBUGGING
3277 regnode *stop = scan;
72f13be8 3278 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 3279#else
d47053eb
RGS
3280 PERL_UNUSED_ARG(depth);
3281#endif
7918f24d
NC
3282
3283 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 3284#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
3285 PERL_UNUSED_ARG(flags);
3286 PERL_UNUSED_ARG(val);
07be1b83 3287#endif
07be1b83 3288 DEBUG_PEEP("join",scan,depth);
bb914485 3289
3f410cf6
KW
3290 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3291 * EXACT ones that are mergeable to the current one. */
3292 while (n
3293 && (PL_regkind[OP(n)] == NOTHING
3294 || (stringok && OP(n) == OP(scan)))
07be1b83 3295 && NEXT_OFF(n)
3f410cf6
KW
3296 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3297 {
538e84ed 3298
07be1b83
YO
3299 if (OP(n) == TAIL || n > next)
3300 stringok = 0;
3301 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
3302 DEBUG_PEEP("skip:",n,depth);
3303 NEXT_OFF(scan) += NEXT_OFF(n);
3304 next = n + NODE_STEP_REGNODE;
3305#ifdef DEBUGGING
3306 if (stringok)
3307 stop = n;
3308#endif
3309 n = regnext(n);
3310 }
3311 else if (stringok) {
786e8c11 3312 const unsigned int oldl = STR_LEN(scan);
07be1b83 3313 regnode * const nnext = regnext(n);
b2230d39 3314
baa60164
KW
3315 /* XXX I (khw) kind of doubt that this works on platforms (should
3316 * Perl ever run on one) where U8_MAX is above 255 because of lots
3317 * of other assumptions */
79a81a6e 3318 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
3319 if (oldl + STR_LEN(n) > U8_MAX)
3320 break;
538e84ed 3321
07be1b83 3322 DEBUG_PEEP("merg",n,depth);
07be1b83 3323 merged++;
b2230d39 3324
07be1b83
YO
3325 NEXT_OFF(scan) += NEXT_OFF(n);
3326 STR_LEN(scan) += STR_LEN(n);
3327 next = n + NODE_SZ_STR(n);
3328 /* Now we can overwrite *n : */
3329 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3330#ifdef DEBUGGING
3331 stop = next - 1;
3332#endif
3333 n = nnext;
3334 if (stopnow) break;
3335 }
3336
d47053eb
RGS
3337#ifdef EXPERIMENTAL_INPLACESCAN
3338 if (flags && !NEXT_OFF(n)) {
3339 DEBUG_PEEP("atch", val, depth);
3340 if (reg_off_by_arg[OP(n)]) {
3341 ARG_SET(n, val - n);
3342 }
3343 else {
3344 NEXT_OFF(n) = val - n;
3345 }
3346 stopnow = 1;
3347 }
07be1b83
YO
3348#endif
3349 }
2c2b7f86 3350
9d071ca8 3351 *min_subtract = 0;
fb29cc72 3352 *unfolded_multi_char = FALSE;
f646642f 3353
3f410cf6
KW
3354 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3355 * can now analyze for sequences of problematic code points. (Prior to
3356 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
3357 * hence missed). The sequences only happen in folding, hence for any
3358 * non-EXACT EXACTish node */
86d6fcad 3359 if (OP(scan) != EXACT) {
31f05a37
KW
3360 U8* s0 = (U8*) STRING(scan);
3361 U8* s = s0;
3362 U8* s_end = s0 + STR_LEN(scan);
3363
3364 int total_count_delta = 0; /* Total delta number of characters that
3365 multi-char folds expand to */
f758bddf
KW
3366
3367 /* One pass is made over the node's string looking for all the
baa60164 3368 * possibilities. To avoid some tests in the loop, there are two main
f758bddf
KW
3369 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3370 * non-UTF-8 */
3371 if (UTF) {
31f05a37
KW
3372 U8* folded = NULL;
3373
3374 if (OP(scan) == EXACTFL) {
3375 U8 *d;
3376
3377 /* An EXACTFL node would already have been changed to another
3378 * node type unless there is at least one character in it that
3379 * is problematic; likely a character whose fold definition
3380 * won't be known until runtime, and so has yet to be folded.
3381 * For all but the UTF-8 locale, folds are 1-1 in length, but
3382 * to handle the UTF-8 case, we need to create a temporary
3383 * folded copy using UTF-8 locale rules in order to analyze it.
3384 * This is because our macros that look to see if a sequence is
3385 * a multi-char fold assume everything is folded (otherwise the
3386 * tests in those macros would be too complicated and slow).
3387 * Note that here, the non-problematic folds will have already
3388 * been done, so we can just copy such characters. We actually
3389 * don't completely fold the EXACTFL string. We skip the
3390 * unfolded multi-char folds, as that would just create work
3391 * below to figure out the size they already are */
3392
3393 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3394 d = folded;
3395 while (s < s_end) {
3396 STRLEN s_len = UTF8SKIP(s);
3397 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3398 Copy(s, d, s_len, U8);
3399 d += s_len;
3400 }
3401 else if (is_FOLDS_TO_MULTI_utf8(s)) {
fb29cc72 3402 *unfolded_multi_char = TRUE;
31f05a37
KW
3403 Copy(s, d, s_len, U8);
3404 d += s_len;
3405 }
3406 else if (isASCII(*s)) {
3407 *(d++) = toFOLD(*s);
3408 }
3409 else {
3410 STRLEN len;
3411 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3412 d += len;
3413 }
3414 s += s_len;
3415 }
3416
3417 /* Point the remainder of the routine to look at our temporary
3418 * folded copy */
3419 s = folded;
3420 s_end = d;
3421 } /* End of creating folded copy of EXACTFL string */
86d6fcad 3422
0a982f06
KW
3423 /* Examine the string for a multi-character fold sequence. UTF-8
3424 * patterns have all characters pre-folded by the time this code is
3425 * executed */
3426 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3427 length sequence we are looking for is 2 */
86d6fcad 3428 {
538e84ed 3429 int count = 0; /* How many characters in a multi-char fold */
251b239f 3430 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
0a982f06
KW
3431 if (! len) { /* Not a multi-char fold: get next char */
3432 s += UTF8SKIP(s);
3433 continue;
3434 }
bb914485 3435
31f05a37
KW
3436 /* Nodes with 'ss' require special handling, except for
3437 * EXACTFA-ish for which there is no multi-char fold to this */
0a982f06 3438 if (len == 2 && *s == 's' && *(s+1) == 's'
098b07d5
KW
3439 && OP(scan) != EXACTFA
3440 && OP(scan) != EXACTFA_NO_TRIE)
0a982f06
KW
3441 {
3442 count = 2;
31f05a37
KW
3443 if (OP(scan) != EXACTFL) {
3444 OP(scan) = EXACTFU_SS;
3445 }
0a982f06
KW
3446 s += 2;
3447 }
0a982f06 3448 else { /* Here is a generic multi-char fold. */
31f05a37
KW
3449 U8* multi_end = s + len;
3450
2fe3656f
KW
3451 /* Count how many characters are in it. In the case of
3452 * /aa, no folds which contain ASCII code points are
3453 * allowed, so check for those, and skip if found. */
31f05a37 3454 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
0a982f06
KW
3455 count = utf8_length(s, multi_end);
3456 s = multi_end;
3457 }
3458 else {
3459 while (s < multi_end) {