This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
[perl #121854] use re 'taint' regression
[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)
26faadbd 96#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 97#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 98
a687059c
LW
99#ifndef STATIC
100#define STATIC static
101#endif
102
b1603ef8 103
09b2b2e6 104struct RExC_state_t {
514a91f1
DM
105 U32 flags; /* RXf_* are we folding, multilining? */
106 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 107 char *precomp; /* uncompiled string. */
288b8c02 108 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 109 regexp *rx; /* perl core regexp structure */
538e84ed
KW
110 regexp_internal *rxi; /* internal data for regexp object
111 pprivate field */
fac92740 112 char *start; /* Start of input for compile */
830247a4
IZ
113 char *end; /* End of input for compile */
114 char *parse; /* Input-scan pointer. */
ea3daa5d 115 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 116 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
117 regnode *emit_bound; /* First regnode outside of the
118 allocated space */
f7c7e32a
DM
119 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
120 implies compiling, so don't emit */
9a81a976
KW
121 regnode_ssc emit_dummy; /* placeholder for emit to point to;
122 large enough for the largest
123 non-EXACTish node, so can use it as
124 scratch in pass1 */
830247a4
IZ
125 I32 naughty; /* How bad is this pattern? */
126 I32 sawback; /* Did we see \1, ...? */
127 U32 seen;
ea3daa5d 128 SSize_t size; /* Code size. */
538e84ed
KW
129 I32 npar; /* Capture buffer count, (OPEN) plus
130 one. ("par" 0 is the whole
131 pattern)*/
132 I32 nestroot; /* root parens we are in - used by
133 accept */
830247a4
IZ
134 I32 extralen;
135 I32 seen_zerolen;
40d049e4
YO
136 regnode **open_parens; /* pointers to open parens */
137 regnode **close_parens; /* pointers to close parens */
138 regnode *opend; /* END node in program */
02daf0ab
YO
139 I32 utf8; /* whether the pattern is utf8 or not */
140 I32 orig_utf8; /* whether the pattern was originally in utf8 */
141 /* XXX use this for future optimisation of case
142 * where pattern must be upgraded to utf8. */
e40e74fe
KW
143 I32 uni_semantics; /* If a d charset modifier should use unicode
144 rules, even if the pattern is not in
145 utf8 */
81714fb9 146 HV *paren_names; /* Paren names */
538e84ed 147
40d049e4
YO
148 regnode **recurse; /* Recurse regops */
149 I32 recurse_count; /* Number of recurse regops */
538e84ed
KW
150 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
151 through */
09a65838 152 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 153 I32 in_lookbehind;
4624b182 154 I32 contains_locale;
cfafade5 155 I32 contains_i;
bb3f3ed2 156 I32 override_recoding;
9d53c457 157 I32 in_multi_char_class;
3d2bd50a 158 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 159 within pattern */
b1603ef8
DM
160 int num_code_blocks; /* size of code_blocks[] */
161 int code_index; /* next code_blocks[] slot */
ee273784 162 SSize_t maxlen; /* mininum possible number of chars in string to match */
dc6d7f5c 163#ifdef ADD_TO_REGEXEC
830247a4
IZ
164 char *starttry; /* -Dr: where regtry was called. */
165#define RExC_starttry (pRExC_state->starttry)
166#endif
d24ca0c5 167 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 168#ifdef DEBUGGING
be8e71aa 169 const char *lastparse;
3dab1dad 170 I32 lastnum;
1f1031fe 171 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
172#define RExC_lastparse (pRExC_state->lastparse)
173#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 174#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 175#endif
09b2b2e6 176};
830247a4 177
e2509266 178#define RExC_flags (pRExC_state->flags)
514a91f1 179#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 180#define RExC_precomp (pRExC_state->precomp)
288b8c02 181#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 182#define RExC_rx (pRExC_state->rx)
f8fc2ecf 183#define RExC_rxi (pRExC_state->rxi)
fac92740 184#define RExC_start (pRExC_state->start)
830247a4
IZ
185#define RExC_end (pRExC_state->end)
186#define RExC_parse (pRExC_state->parse)
187#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237 188#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
189#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
190 others */
7122b237 191#endif
830247a4 192#define RExC_emit (pRExC_state->emit)
f7c7e32a 193#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 194#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 195#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
196#define RExC_naughty (pRExC_state->naughty)
197#define RExC_sawback (pRExC_state->sawback)
198#define RExC_seen (pRExC_state->seen)
199#define RExC_size (pRExC_state->size)
ee273784 200#define RExC_maxlen (pRExC_state->maxlen)
830247a4 201#define RExC_npar (pRExC_state->npar)
e2e6a0f1 202#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
203#define RExC_extralen (pRExC_state->extralen)
204#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 205#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 206#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 207#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
208#define RExC_open_parens (pRExC_state->open_parens)
209#define RExC_close_parens (pRExC_state->close_parens)
210#define RExC_opend (pRExC_state->opend)
81714fb9 211#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
212#define RExC_recurse (pRExC_state->recurse)
213#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 214#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
215#define RExC_study_chunk_recursed_bytes \
216 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 217#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 218#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 219#define RExC_contains_i (pRExC_state->contains_i)
9d53c457
KW
220#define RExC_override_recoding (pRExC_state->override_recoding)
221#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 222
cde0cee5 223
a687059c
LW
224#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
225#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
4d68ffa0 226 ((*s) == '{' && regcurly(s, FALSE)))
a687059c
LW
227
228/*
229 * Flags to be passed up and down.
230 */
a687059c 231#define WORST 0 /* Worst case. */
a3b492c3 232#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 233
e64f369d 234/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
237 * REGNODE_SIMPLE */
fda99bee 238#define SIMPLE 0x02
e64f369d 239#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
240#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
241#define TRYAGAIN 0x10 /* Weeded out a declaration. */
242#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 243
3dab1dad
YO
244#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
245
07be1b83
YO
246/* whether trie related optimizations are enabled */
247#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
248#define TRIE_STUDY_OPT
786e8c11 249#define FULL_TRIE_STUDY
07be1b83
YO
250#define TRIE_STCLASS
251#endif
1de06328
YO
252
253
40d049e4
YO
254
255#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
256#define PBITVAL(paren) (1 << ((paren) & 7))
257#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
258#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
259#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
260
bbd61b5f 261#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
262 if (!UTF) { \
263 *flagp = RESTART_UTF8; \
264 return NULL; \
265 } \
bbd61b5f 266 } STMT_END
40d049e4 267
f19b1a63
KW
268/* This converts the named class defined in regcomp.h to its equivalent class
269 * number defined in handy.h. */
270#define namedclass_to_classnum(class) ((int) ((class) / 2))
271#define classnum_to_namedclass(classnum) ((classnum) * 2)
272
de92f5e6
KW
273#define _invlist_union_complement_2nd(a, b, output) \
274 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
275#define _invlist_intersection_complement_2nd(a, b, output) \
276 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
277
1de06328
YO
278/* About scan_data_t.
279
280 During optimisation we recurse through the regexp program performing
281 various inplace (keyhole style) optimisations. In addition study_chunk
282 and scan_commit populate this data structure with information about
538e84ed 283 what strings MUST appear in the pattern. We look for the longest
3b753521 284 string that must appear at a fixed location, and we look for the
1de06328
YO
285 longest string that may appear at a floating location. So for instance
286 in the pattern:
538e84ed 287
1de06328 288 /FOO[xX]A.*B[xX]BAR/
538e84ed 289
1de06328
YO
290 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
291 strings (because they follow a .* construct). study_chunk will identify
292 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 293
1de06328 294 The strings can be composites, for instance
538e84ed 295
1de06328 296 /(f)(o)(o)/
538e84ed 297
1de06328 298 will result in a composite fixed substring 'foo'.
538e84ed 299
1de06328 300 For each string some basic information is maintained:
538e84ed 301
1de06328
YO
302 - offset or min_offset
303 This is the position the string must appear at, or not before.
304 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
305 characters must match before the string we are searching for.
306 Likewise when combined with minlenp and the length of the string it
538e84ed 307 tells us how many characters must appear after the string we have
1de06328 308 found.
538e84ed 309
1de06328
YO
310 - max_offset
311 Only used for floating strings. This is the rightmost point that
ea3daa5d 312 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 313 string can occur infinitely far to the right.
538e84ed 314
1de06328 315 - minlenp
2d608413
KW
316 A pointer to the minimum number of characters of the pattern that the
317 string was found inside. This is important as in the case of positive
538e84ed 318 lookahead or positive lookbehind we can have multiple patterns
1de06328 319 involved. Consider
538e84ed 320
1de06328 321 /(?=FOO).*F/
538e84ed 322
1de06328
YO
323 The minimum length of the pattern overall is 3, the minimum length
324 of the lookahead part is 3, but the minimum length of the part that
538e84ed 325 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 326 minimum length for the F is 1. This is important as the minimum length
538e84ed 327 is used to determine offsets in front of and behind the string being
1de06328 328 looked for. Since strings can be composites this is the length of the
486ec47a 329 pattern at the time it was committed with a scan_commit. Note that
1de06328 330 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 331 are not known until the full pattern has been compiled, thus the
1de06328 332 pointer to the value.
538e84ed 333
1de06328 334 - lookbehind
538e84ed 335
1de06328 336 In the case of lookbehind the string being searched for can be
538e84ed 337 offset past the start point of the final matching string.
1de06328
YO
338 If this value was just blithely removed from the min_offset it would
339 invalidate some of the calculations for how many chars must match
340 before or after (as they are derived from min_offset and minlen and
538e84ed 341 the length of the string being searched for).
1de06328
YO
342 When the final pattern is compiled and the data is moved from the
343 scan_data_t structure into the regexp structure the information
538e84ed
KW
344 about lookbehind is factored in, with the information that would
345 have been lost precalculated in the end_shift field for the
1de06328
YO
346 associated string.
347
348 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 349 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
350
351*/
2c2d71f5
JH
352
353typedef struct scan_data_t {
1de06328
YO
354 /*I32 len_min; unused */
355 /*I32 len_delta; unused */
49f55535 356 SSize_t pos_min;
ea3daa5d 357 SSize_t pos_delta;
2c2d71f5 358 SV *last_found;
ea3daa5d 359 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 360 SSize_t last_start_min;
ea3daa5d 361 SSize_t last_start_max;
1de06328
YO
362 SV **longest; /* Either &l_fixed, or &l_float. */
363 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 364 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 365 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
366 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
367 SV *longest_float; /* longest floating string found in pattern */
49f55535 368 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
369 SSize_t offset_float_max; /* latest point in string it can appear */
370 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 371 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
372 I32 flags;
373 I32 whilem_c;
49f55535 374 SSize_t *last_closep;
b8f7bb16 375 regnode_ssc *start_class;
2c2d71f5
JH
376} scan_data_t;
377
c02c3054
KW
378/* The below is perhaps overboard, but this allows us to save a test at the
379 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
380 * and 'a' differ by a single bit; the same with the upper and lower case of
381 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
382 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
383 * then inverts it to form a mask, with just a single 0, in the bit position
384 * where the upper- and lowercase differ. XXX There are about 40 other
385 * instances in the Perl core where this micro-optimization could be used.
386 * Should decide if maintenance cost is worse, before changing those
387 *
388 * Returns a boolean as to whether or not 'v' is either a lowercase or
389 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
390 * compile-time constant, the generated code is better than some optimizing
391 * compilers figure out, amounting to a mask and test. The results are
392 * meaningless if 'c' is not one of [A-Za-z] */
393#define isARG2_lower_or_UPPER_ARG1(c, v) \
394 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
395
a687059c 396/*
e50aee73 397 * Forward declarations for pregcomp()'s friends.
a687059c 398 */
a0d0e21e 399
27da23d5 400static const scan_data_t zero_scan_data =
1de06328 401 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
402
403#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
404#define SF_BEFORE_SEOL 0x0001
405#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
406#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
407#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
408
44e3dfd2
BF
409#define SF_FIX_SHIFT_EOL (+2)
410#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
411
412#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
413#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
414
415#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
416#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
417#define SF_IS_INF 0x0040
418#define SF_HAS_PAR 0x0080
419#define SF_IN_PAR 0x0100
420#define SF_HAS_EVAL 0x0200
421#define SCF_DO_SUBSTR 0x0400
653099ff
GS
422#define SCF_DO_STCLASS_AND 0x0800
423#define SCF_DO_STCLASS_OR 0x1000
424#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 425#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 426
786e8c11 427#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 428#define SCF_SEEN_ACCEPT 0x8000
688e0391 429#define SCF_TRIE_DOING_RESTUDY 0x10000
07be1b83 430
43fead97 431#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
432
433/* The enums for all these are ordered so things work out correctly */
a62b1201 434#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
435#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
436 == REGEX_DEPENDS_CHARSET)
00b27cfc 437#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
438#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
439 >= REGEX_UNICODE_CHARSET)
440#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
441 == REGEX_ASCII_RESTRICTED_CHARSET)
442#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
443 >= REGEX_ASCII_RESTRICTED_CHARSET)
444#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
445 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 446
43fead97 447#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 448
f2c2a6ab
KW
449/* For programs that want to be strictly Unicode compatible by dying if any
450 * attempt is made to match a non-Unicode code point against a Unicode
451 * property. */
452#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
453
93733859 454#define OOB_NAMEDCLASS -1
b8c5462f 455
8e661ac5
KW
456/* There is no code point that is out-of-bounds, so this is problematic. But
457 * its only current use is to initialize a variable that is always set before
458 * looked at. */
459#define OOB_UNICODE 0xDEADBEEF
460
a0ed51b3
LW
461#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
462#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
463
8615cb43 464
b45f050a
JF
465/* length of regex to show in messages that don't mark a position within */
466#define RegexLengthToShowInErrorMessages 127
467
468/*
469 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
470 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
471 * op/pragma/warn/regcomp.
472 */
7253e4e3
RK
473#define MARKER1 "<-- HERE" /* marker as it appears in the description */
474#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 475
538e84ed
KW
476#define REPORT_LOCATION " in regex; marked by " MARKER1 \
477 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 478
c1d900c3
BF
479#define REPORT_LOCATION_ARGS(offset) \
480 UTF8fARG(UTF, offset, RExC_precomp), \
481 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
482
b45f050a
JF
483/*
484 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
485 * arg. Show regex, up to a maximum length. If it's too long, chop and add
486 * "...".
487 */
58e23c8d 488#define _FAIL(code) STMT_START { \
bfed75c6 489 const char *ellipses = ""; \
ccb2c380
MP
490 IV len = RExC_end - RExC_precomp; \
491 \
492 if (!SIZE_ONLY) \
a5e7bc51 493 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
494 if (len > RegexLengthToShowInErrorMessages) { \
495 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
496 len = RegexLengthToShowInErrorMessages - 10; \
497 ellipses = "..."; \
498 } \
58e23c8d 499 code; \
ccb2c380 500} STMT_END
8615cb43 501
58e23c8d 502#define FAIL(msg) _FAIL( \
c1d900c3
BF
503 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
504 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
505
506#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
507 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
508 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 509
b45f050a 510/*
b45f050a
JF
511 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
512 */
ccb2c380 513#define Simple_vFAIL(m) STMT_START { \
a28509cc 514 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380 515 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
c1d900c3 516 m, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 517} STMT_END
b45f050a
JF
518
519/*
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
521 */
ccb2c380
MP
522#define vFAIL(m) STMT_START { \
523 if (!SIZE_ONLY) \
a5e7bc51 524 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
525 Simple_vFAIL(m); \
526} STMT_END
b45f050a
JF
527
528/*
529 * Like Simple_vFAIL(), but accepts two arguments.
530 */
ccb2c380 531#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 532 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
533 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
534 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 535} STMT_END
b45f050a
JF
536
537/*
538 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
539 */
ccb2c380
MP
540#define vFAIL2(m,a1) STMT_START { \
541 if (!SIZE_ONLY) \
a5e7bc51 542 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
543 Simple_vFAIL2(m, a1); \
544} STMT_END
b45f050a
JF
545
546
547/*
548 * Like Simple_vFAIL(), but accepts three arguments.
549 */
ccb2c380 550#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 551 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
552 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
553 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 554} STMT_END
b45f050a
JF
555
556/*
557 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
558 */
ccb2c380
MP
559#define vFAIL3(m,a1,a2) STMT_START { \
560 if (!SIZE_ONLY) \
a5e7bc51 561 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
562 Simple_vFAIL3(m, a1, a2); \
563} STMT_END
b45f050a
JF
564
565/*
566 * Like Simple_vFAIL(), but accepts four arguments.
567 */
ccb2c380 568#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 569 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
570 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
571 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 572} STMT_END
b45f050a 573
95db3ffa
KW
574#define vFAIL4(m,a1,a2,a3) STMT_START { \
575 if (!SIZE_ONLY) \
576 SAVEFREESV(RExC_rx_sv); \
577 Simple_vFAIL4(m, a1, a2, a3); \
578} STMT_END
579
946095af
BF
580/* A specialized version of vFAIL2 that works with UTF8f */
581#define vFAIL2utf8f(m, a1) STMT_START { \
ef3f731d
BF
582 const IV offset = RExC_parse - RExC_precomp; \
583 if (!SIZE_ONLY) \
584 SAVEFREESV(RExC_rx_sv); \
585 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
586 REPORT_LOCATION_ARGS(offset)); \
946095af
BF
587} STMT_END
588
589
5e0a247b
KW
590/* m is not necessarily a "literal string", in this macro */
591#define reg_warn_non_literal_string(loc, m) STMT_START { \
592 const IV offset = loc - RExC_precomp; \
593 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
c1d900c3 594 m, REPORT_LOCATION_ARGS(offset)); \
5e0a247b
KW
595} STMT_END
596
668c081a 597#define ckWARNreg(loc,m) STMT_START { \
a28509cc 598 const IV offset = loc - RExC_precomp; \
f10f4c18 599 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 600 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
601} STMT_END
602
0d6106aa
KW
603#define vWARN_dep(loc, m) STMT_START { \
604 const IV offset = loc - RExC_precomp; \
605 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
c1d900c3 606 REPORT_LOCATION_ARGS(offset)); \
0d6106aa
KW
607} STMT_END
608
147508a2
KW
609#define ckWARNdep(loc,m) STMT_START { \
610 const IV offset = loc - RExC_precomp; \
611 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
612 m REPORT_LOCATION, \
c1d900c3 613 REPORT_LOCATION_ARGS(offset)); \
147508a2
KW
614} STMT_END
615
668c081a 616#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 617 const IV offset = loc - RExC_precomp; \
d1d15184 618 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18 619 m REPORT_LOCATION, \
c1d900c3 620 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
621} STMT_END
622
b23eb183 623#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 624 const IV offset = loc - RExC_precomp; \
b23eb183 625 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3 626 m REPORT_LOCATION, \
c1d900c3 627 a1, REPORT_LOCATION_ARGS(offset)); \
2335b3d3
KW
628} STMT_END
629
668c081a 630#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 631 const IV offset = loc - RExC_precomp; \
668c081a 632 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 633 a1, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
634} STMT_END
635
636#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 637 const IV offset = loc - RExC_precomp; \
ccb2c380 638 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 639 a1, a2, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
640} STMT_END
641
668c081a
NC
642#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
643 const IV offset = loc - RExC_precomp; \
644 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 645 a1, a2, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
646} STMT_END
647
ccb2c380 648#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 649 const IV offset = loc - RExC_precomp; \
ccb2c380 650 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 651 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
652} STMT_END
653
668c081a
NC
654#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
655 const IV offset = loc - RExC_precomp; \
656 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 657 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
658} STMT_END
659
ccb2c380 660#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 661 const IV offset = loc - RExC_precomp; \
ccb2c380 662 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 663 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 664} STMT_END
9d1d55b5 665
8615cb43 666
cd439c50 667/* Allow for side effects in s */
ccb2c380
MP
668#define REGC(c,s) STMT_START { \
669 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
670} STMT_END
cd439c50 671
538e84ed 672/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
673 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
674 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
675 * Element 0 holds the number n.
07be1b83 676 * Position is 1 indexed.
fac92740 677 */
7122b237
YO
678#ifndef RE_TRACK_PATTERN_OFFSETS
679#define Set_Node_Offset_To_R(node,byte)
680#define Set_Node_Offset(node,byte)
681#define Set_Cur_Node_Offset
682#define Set_Node_Length_To_R(node,len)
683#define Set_Node_Length(node,len)
6a86c6ad 684#define Set_Node_Cur_Length(node,start)
538e84ed
KW
685#define Node_Offset(n)
686#define Node_Length(n)
7122b237
YO
687#define Set_Node_Offset_Length(node,offset,len)
688#define ProgLen(ri) ri->u.proglen
689#define SetProgLen(ri,x) ri->u.proglen = x
690#else
691#define ProgLen(ri) ri->u.offsets[0]
692#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
693#define Set_Node_Offset_To_R(node,byte) STMT_START { \
694 if (! SIZE_ONLY) { \
695 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 696 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 697 if((node) < 0) { \
538e84ed
KW
698 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
699 (int)(node)); \
ccb2c380
MP
700 } else { \
701 RExC_offsets[2*(node)-1] = (byte); \
702 } \
703 } \
704} STMT_END
705
706#define Set_Node_Offset(node,byte) \
707 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
708#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
709
710#define Set_Node_Length_To_R(node,len) STMT_START { \
711 if (! SIZE_ONLY) { \
712 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 713 __LINE__, (int)(node), (int)(len))); \
ccb2c380 714 if((node) < 0) { \
538e84ed
KW
715 Perl_croak(aTHX_ "value of node is %d in Length macro", \
716 (int)(node)); \
ccb2c380
MP
717 } else { \
718 RExC_offsets[2*(node)] = (len); \
719 } \
720 } \
721} STMT_END
722
723#define Set_Node_Length(node,len) \
724 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
725#define Set_Node_Cur_Length(node, start) \
726 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
727
728/* Get offsets and lengths */
729#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
730#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
731
07be1b83
YO
732#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
733 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
734 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
735} STMT_END
7122b237 736#endif
07be1b83
YO
737
738#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
739#define EXPERIMENTAL_INPLACESCAN
f427392e 740#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 741
9e9ecfdf 742#define DEBUG_RExC_seen() \
538e84ed
KW
743 DEBUG_OPTIMISE_MORE_r({ \
744 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
745 \
e384d5c1
YO
746 if (RExC_seen & REG_ZERO_LEN_SEEN) \
747 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
538e84ed 748 \
e384d5c1
YO
749 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
750 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
538e84ed 751 \
e384d5c1
YO
752 if (RExC_seen & REG_GPOS_SEEN) \
753 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
538e84ed 754 \
e384d5c1
YO
755 if (RExC_seen & REG_CANY_SEEN) \
756 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
538e84ed 757 \
e384d5c1
YO
758 if (RExC_seen & REG_RECURSE_SEEN) \
759 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
538e84ed 760 \
e384d5c1
YO
761 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
762 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 763 \
e384d5c1
YO
764 if (RExC_seen & REG_VERBARG_SEEN) \
765 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
538e84ed 766 \
e384d5c1
YO
767 if (RExC_seen & REG_CUTGROUP_SEEN) \
768 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
538e84ed 769 \
e384d5c1
YO
770 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
771 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 772 \
e384d5c1
YO
773 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
774 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 775 \
e384d5c1
YO
776 if (RExC_seen & REG_GOSTART_SEEN) \
777 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
538e84ed 778 \
ee273784
YO
779 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
780 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
781 \
538e84ed 782 PerlIO_printf(Perl_debug_log,"\n"); \
9e9ecfdf
YO
783 });
784
304ee84b
YO
785#define DEBUG_STUDYDATA(str,data,depth) \
786DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 787 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
788 "%*s" str "Pos:%"IVdf"/%"IVdf \
789 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
790 (int)(depth)*2, "", \
791 (IV)((data)->pos_min), \
792 (IV)((data)->pos_delta), \
304ee84b 793 (UV)((data)->flags), \
1de06328 794 (IV)((data)->whilem_c), \
304ee84b
YO
795 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
796 is_inf ? "INF " : "" \
1de06328
YO
797 ); \
798 if ((data)->last_found) \
799 PerlIO_printf(Perl_debug_log, \
800 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
801 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
802 SvPVX_const((data)->last_found), \
803 (IV)((data)->last_end), \
804 (IV)((data)->last_start_min), \
805 (IV)((data)->last_start_max), \
806 ((data)->longest && \
807 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
808 SvPVX_const((data)->longest_fixed), \
809 (IV)((data)->offset_fixed), \
810 ((data)->longest && \
811 (data)->longest==&((data)->longest_float)) ? "*" : "", \
812 SvPVX_const((data)->longest_float), \
813 (IV)((data)->offset_float_min), \
814 (IV)((data)->offset_float_max) \
815 ); \
816 PerlIO_printf(Perl_debug_log,"\n"); \
817});
818
653099ff 819/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 820 Update the longest found anchored substring and the longest found
653099ff
GS
821 floating substrings if needed. */
822
4327152a 823STATIC void
ea3daa5d
FC
824S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
825 SSize_t *minlenp, int is_inf)
c277df42 826{
e1ec3a88
AL
827 const STRLEN l = CHR_SVLEN(data->last_found);
828 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 829 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 830
7918f24d
NC
831 PERL_ARGS_ASSERT_SCAN_COMMIT;
832
c277df42 833 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 834 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
835 if (*data->longest == data->longest_fixed) {
836 data->offset_fixed = l ? data->last_start_min : data->pos_min;
837 if (data->flags & SF_BEFORE_EOL)
b81d288d 838 data->flags
c277df42
IZ
839 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
840 else
841 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 842 data->minlen_fixed=minlenp;
1de06328 843 data->lookbehind_fixed=0;
a0ed51b3 844 }
304ee84b 845 else { /* *data->longest == data->longest_float */
c277df42 846 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
847 data->offset_float_max = (l
848 ? data->last_start_max
ea3daa5d
FC
849 : (data->pos_delta == SSize_t_MAX
850 ? SSize_t_MAX
851 : data->pos_min + data->pos_delta));
852 if (is_inf
853 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
854 data->offset_float_max = SSize_t_MAX;
c277df42 855 if (data->flags & SF_BEFORE_EOL)
b81d288d 856 data->flags
c277df42
IZ
857 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
858 else
859 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
860 data->minlen_float=minlenp;
861 data->lookbehind_float=0;
c277df42
IZ
862 }
863 }
864 SvCUR_set(data->last_found, 0);
0eda9292 865 {
a28509cc 866 SV * const sv = data->last_found;
097eb12c
AL
867 if (SvUTF8(sv) && SvMAGICAL(sv)) {
868 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
869 if (mg)
870 mg->mg_len = 0;
871 }
0eda9292 872 }
c277df42
IZ
873 data->last_end = -1;
874 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 875 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
876}
877
cdd87c1d
KW
878/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
879 * list that describes which code points it matches */
880
653099ff 881STATIC void
3420edd7 882S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 883{
cdd87c1d
KW
884 /* Set the SSC 'ssc' to match an empty string or any code point */
885
557bd3fb 886 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 887
71068078 888 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 889
cdd87c1d
KW
890 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
891 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
a0dd4231 892 ANYOF_FLAGS(ssc) |= ANYOF_EMPTY_STRING; /* Plus match empty string */
653099ff
GS
893}
894
653099ff 895STATIC int
cdd87c1d 896S_ssc_is_anything(pTHX_ const regnode_ssc *ssc)
653099ff 897{
c144baaa
KW
898 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
899 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
900 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
901 * in any way, so there's no point in using it */
cdd87c1d
KW
902
903 UV start, end;
904 bool ret;
653099ff 905
557bd3fb 906 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 907
71068078 908 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 909
77ebeeba 910 if (! (ANYOF_FLAGS(ssc) & ANYOF_EMPTY_STRING)) {
cdd87c1d
KW
911 return FALSE;
912 }
913
914 /* See if the list consists solely of the range 0 - Infinity */
915 invlist_iterinit(ssc->invlist);
916 ret = invlist_iternext(ssc->invlist, &start, &end)
917 && start == 0
918 && end == UV_MAX;
919
920 invlist_iterfinish(ssc->invlist);
921
922 if (ret) {
923 return TRUE;
924 }
925
926 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 927 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
928 int i;
929 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
930 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
931 return TRUE;
932 }
933 }
934 }
935
936 return FALSE;
653099ff
GS
937}
938
653099ff 939STATIC void
cdd87c1d 940S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 941{
cdd87c1d
KW
942 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
943 * string, any code point, or any posix class under locale */
944
557bd3fb 945 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 946
557bd3fb 947 Zero(ssc, 1, regnode_ssc);
71068078 948 set_ANYOF_SYNTHETIC(ssc);
cdd87c1d 949 ARG_SET(ssc, ANYOF_NONBITMAP_EMPTY);
3420edd7 950 ssc_anything(ssc);
cdd87c1d
KW
951
952 /* If any portion of the regex is to operate under locale rules,
953 * initialization includes it. The reason this isn't done for all regexes
954 * is that the optimizer was written under the assumption that locale was
955 * all-or-nothing. Given the complexity and lack of documentation in the
956 * optimizer, and that there are inadequate test cases for locale, many
957 * parts of it may not work properly, it is safest to avoid locale unless
958 * necessary. */
959 if (RExC_contains_locale) {
960 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
961 }
962 else {
963 ANYOF_POSIXL_ZERO(ssc);
964 }
653099ff
GS
965}
966
b423522f
KW
967STATIC int
968S_ssc_is_cp_posixl_init(pTHX_ const RExC_state_t *pRExC_state,
969 const regnode_ssc *ssc)
970{
971 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
972 * to the list of code points matched, and locale posix classes; hence does
973 * not check its flags) */
974
975 UV start, end;
976 bool ret;
977
978 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
979
71068078 980 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
981
982 invlist_iterinit(ssc->invlist);
983 ret = invlist_iternext(ssc->invlist, &start, &end)
984 && start == 0
985 && end == UV_MAX;
986
987 invlist_iterfinish(ssc->invlist);
988
989 if (! ret) {
990 return FALSE;
991 }
992
e0e1be5f 993 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 994 return FALSE;
b423522f
KW
995 }
996
997 return TRUE;
998}
999
1000STATIC SV*
1001S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1002 const regnode_charclass* const node)
b423522f
KW
1003{
1004 /* Returns a mortal inversion list defining which code points are matched
1005 * by 'node', which is of type ANYOF. Handles complementing the result if
1006 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1007 * returned list must, and will, contain every code point that is a
1008 * possibility. */
b423522f
KW
1009
1010 SV* invlist = sv_2mortal(_new_invlist(0));
1ee208c4 1011 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1012 unsigned int i;
1013 const U32 n = ARG(node);
31f05a37 1014 bool new_node_has_latin1 = FALSE;
b423522f
KW
1015
1016 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1017
1018 /* Look at the data structure created by S_set_ANYOF_arg() */
1019 if (n != ANYOF_NONBITMAP_EMPTY) {
1020 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1021 AV * const av = MUTABLE_AV(SvRV(rv));
1022 SV **const ary = AvARRAY(av);
1023 assert(RExC_rxi->data->what[n] == 's');
1024
1025 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1026 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1027 }
1028 else if (ary[0] && ary[0] != &PL_sv_undef) {
1029
1030 /* Here, no compile-time swash, and there are things that won't be
1031 * known until runtime -- we have to assume it could be anything */
1032 return _add_range_to_invlist(invlist, 0, UV_MAX);
1033 }
1ee208c4 1034 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1035
1036 /* Here no compile-time swash, and no run-time only data. Use the
1037 * node's inversion list */
1ee208c4
KW
1038 invlist = sv_2mortal(invlist_clone(ary[3]));
1039 }
1040
1041 /* Get the code points valid only under UTF-8 locales */
1042 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1043 && ary[2] && ary[2] != &PL_sv_undef)
1044 {
1045 only_utf8_locale_invlist = ary[2];
b423522f
KW
1046 }
1047 }
1048
1049 /* An ANYOF node contains a bitmap for the first 256 code points, and an
1050 * inversion list for the others, but if there are code points that should
1051 * match only conditionally on the target string being UTF-8, those are
1052 * placed in the inversion list, and not the bitmap. Since there are
1053 * circumstances under which they could match, they are included in the
1054 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1055 * here, so that when we invert below, the end result actually does include
1056 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1057 * before we add the unconditionally matched code points */
1058 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1059 _invlist_intersection_complement_2nd(invlist,
1060 PL_UpperLatin1,
1061 &invlist);
1062 }
1063
1064 /* Add in the points from the bit map */
1065 for (i = 0; i < 256; i++) {
1066 if (ANYOF_BITMAP_TEST(node, i)) {
1067 invlist = add_cp_to_invlist(invlist, i);
31f05a37 1068 new_node_has_latin1 = TRUE;
b423522f
KW
1069 }
1070 }
1071
1072 /* If this can match all upper Latin1 code points, have to add them
1073 * as well */
7bc66b18 1074 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
b423522f
KW
1075 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1076 }
1077
1078 /* Similarly for these */
1079 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1080 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1081 }
1082
1083 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1084 _invlist_invert(invlist);
1085 }
31f05a37
KW
1086 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1087
1088 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1089 * locale. We can skip this if there are no 0-255 at all. */
1090 _invlist_union(invlist, PL_Latin1, &invlist);
1091 }
1092
1ee208c4
KW
1093 /* Similarly add the UTF-8 locale possible matches. These have to be
1094 * deferred until after the non-UTF-8 locale ones are taken care of just
1095 * above, or it leads to wrong results under ANYOF_INVERT */
1096 if (only_utf8_locale_invlist) {
31f05a37 1097 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1098 only_utf8_locale_invlist,
31f05a37
KW
1099 ANYOF_FLAGS(node) & ANYOF_INVERT,
1100 &invlist);
1101 }
b423522f
KW
1102
1103 return invlist;
1104}
1105
1051e1c4 1106/* These two functions currently do the exact same thing */
557bd3fb 1107#define ssc_init_zero ssc_init
653099ff 1108
cdd87c1d
KW
1109#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1110#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1111
557bd3fb 1112/* 'AND' a given class with another one. Can create false positives. 'ssc'
8efd3f97 1113 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
b8f7bb16 1114 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1115
653099ff 1116STATIC void
b423522f 1117S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1118 const regnode_charclass *and_with)
653099ff 1119{
cdd87c1d
KW
1120 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1121 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1122
a0dd4231
KW
1123 SV* anded_cp_list;
1124 U8 anded_flags;
1e6ade67 1125
cdd87c1d 1126 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1127
71068078 1128 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1129
1130 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1131 * the code point inversion list and just the relevant flags */
71068078 1132 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1133 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1134 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1135
1136 /* XXX This is a kludge around what appears to be deficiencies in the
1137 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1138 * there are paths through the optimizer where it doesn't get weeded
1139 * out when it should. And if we don't make some extra provision for
1140 * it like the code just below, it doesn't get added when it should.
1141 * This solution is to add it only when AND'ing, which is here, and
1142 * only when what is being AND'ed is the pristine, original node
1143 * matching anything. Thus it is like adding it to ssc_anything() but
1144 * only when the result is to be AND'ed. Probably the same solution
1145 * could be adopted for the same problem we have with /l matching,
1146 * which is solved differently in S_ssc_init(), and that would lead to
1147 * fewer false positives than that solution has. But if this solution
1148 * creates bugs, the consequences are only that a warning isn't raised
1149 * that should be; while the consequences for having /l bugs is
1150 * incorrect matches */
7dcac5f6 1151 if (ssc_is_anything((regnode_ssc *)and_with)) {
e9b08962
KW
1152 anded_flags |= ANYOF_WARN_SUPER;
1153 }
a0dd4231
KW
1154 }
1155 else {
5c0f85ef 1156 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
eff8b7dc 1157 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1158 }
1159
1160 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1161
1162 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1163 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1164 * 'and_with' may be inverted. When not inverted, we have the situation of
1165 * computing:
1166 * (C1 | P1) & (C2 | P2)
1167 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1168 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1169 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1170 * <= ((C1 & C2) | P1 | P2)
1171 * Alternatively, the last few steps could be:
1172 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1173 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1174 * <= (C1 | C2 | (P1 & P2))
1175 * We favor the second approach if either P1 or P2 is non-empty. This is
1176 * because these components are a barrier to doing optimizations, as what
1177 * they match cannot be known until the moment of matching as they are
1178 * dependent on the current locale, 'AND"ing them likely will reduce or
1179 * eliminate them.
1180 * But we can do better if we know that C1,P1 are in their initial state (a
1181 * frequent occurrence), each matching everything:
1182 * (<everything>) & (C2 | P2) = C2 | P2
1183 * Similarly, if C2,P2 are in their initial state (again a frequent
1184 * occurrence), the result is a no-op
1185 * (C1 | P1) & (<everything>) = C1 | P1
1186 *
1187 * Inverted, we have
1188 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1189 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1190 * <= (C1 & ~C2) | (P1 & ~P2)
1191 * */
1aa99e6b 1192
a0dd4231 1193 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1194 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1195 {
cdd87c1d 1196 unsigned int i;
8951c461 1197
cdd87c1d
KW
1198 ssc_intersection(ssc,
1199 anded_cp_list,
1200 FALSE /* Has already been inverted */
1201 );
c6b76537 1202
cdd87c1d
KW
1203 /* If either P1 or P2 is empty, the intersection will be also; can skip
1204 * the loop */
1205 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1206 ANYOF_POSIXL_ZERO(ssc);
1207 }
e0e1be5f 1208 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1209
1210 /* Note that the Posix class component P from 'and_with' actually
1211 * looks like:
1212 * P = Pa | Pb | ... | Pn
1213 * where each component is one posix class, such as in [\w\s].
1214 * Thus
1215 * ~P = ~(Pa | Pb | ... | Pn)
1216 * = ~Pa & ~Pb & ... & ~Pn
1217 * <= ~Pa | ~Pb | ... | ~Pn
1218 * The last is something we can easily calculate, but unfortunately
1219 * is likely to have many false positives. We could do better
1220 * in some (but certainly not all) instances if two classes in
1221 * P have known relationships. For example
1222 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1223 * So
1224 * :lower: & :print: = :lower:
1225 * And similarly for classes that must be disjoint. For example,
1226 * since \s and \w can have no elements in common based on rules in
1227 * the POSIX standard,
1228 * \w & ^\S = nothing
1229 * Unfortunately, some vendor locales do not meet the Posix
1230 * standard, in particular almost everything by Microsoft.
1231 * The loop below just changes e.g., \w into \W and vice versa */
1232
1ee208c4 1233 regnode_charclass_posixl temp;
cdd87c1d
KW
1234 int add = 1; /* To calculate the index of the complement */
1235
1236 ANYOF_POSIXL_ZERO(&temp);
1237 for (i = 0; i < ANYOF_MAX; i++) {
1238 assert(i % 2 != 0
7dcac5f6
KW
1239 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1240 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1241
7dcac5f6 1242 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1243 ANYOF_POSIXL_SET(&temp, i + add);
1244 }
1245 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1246 }
1247 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1248
cdd87c1d
KW
1249 } /* else ssc already has no posixes */
1250 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1251 in its initial state */
71068078 1252 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1253 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1254 {
1255 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1256 * copy it over 'ssc' */
1257 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1258 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1259 StructCopy(and_with, ssc, regnode_ssc);
1260 }
1261 else {
1262 ssc->invlist = anded_cp_list;
1263 ANYOF_POSIXL_ZERO(ssc);
1264 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
7dcac5f6 1265 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1266 }
1267 }
1268 }
e0e1be5f
KW
1269 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1270 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
cdd87c1d
KW
1271 {
1272 /* One or the other of P1, P2 is non-empty. */
1ea8b7fe
KW
1273 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1274 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1275 }
cdd87c1d
KW
1276 ssc_union(ssc, anded_cp_list, FALSE);
1277 }
1278 else { /* P1 = P2 = empty */
1279 ssc_intersection(ssc, anded_cp_list, FALSE);
1280 }
137165a6 1281 }
653099ff
GS
1282}
1283
653099ff 1284STATIC void
cdd87c1d 1285S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1286 const regnode_charclass *or_with)
653099ff 1287{
cdd87c1d
KW
1288 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1289 * another SSC or a regular ANYOF class. Can create false positives if
1290 * 'or_with' is to be inverted. */
7918f24d 1291
a0dd4231
KW
1292 SV* ored_cp_list;
1293 U8 ored_flags;
c6b76537 1294
cdd87c1d 1295 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1296
71068078 1297 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1298
1299 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1300 * the code point inversion list and just the relevant flags */
71068078 1301 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1302 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1303 ored_flags = ANYOF_FLAGS(or_with);
1304 }
1305 else {
5c0f85ef 1306 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1307 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1308 }
1309
1310 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1311
1312 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1313 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1314 * 'or_with' may be inverted. When not inverted, we have the simple
1315 * situation of computing:
1316 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1317 * If P1|P2 yields a situation with both a class and its complement are
1318 * set, like having both \w and \W, this matches all code points, and we
1319 * can delete these from the P component of the ssc going forward. XXX We
1320 * might be able to delete all the P components, but I (khw) am not certain
1321 * about this, and it is better to be safe.
1322 *
1323 * Inverted, we have
1324 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1325 * <= (C1 | P1) | ~C2
1326 * <= (C1 | ~C2) | P1
1327 * (which results in actually simpler code than the non-inverted case)
1328 * */
9826f543 1329
a0dd4231 1330 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1331 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1332 {
cdd87c1d 1333 /* We ignore P2, leaving P1 going forward */
1ea8b7fe
KW
1334 } /* else Not inverted */
1335 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
7dcac5f6 1336 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1337 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1338 unsigned int i;
1339 for (i = 0; i < ANYOF_MAX; i += 2) {
1340 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1341 {
1342 ssc_match_all_cp(ssc);
1343 ANYOF_POSIXL_CLEAR(ssc, i);
1344 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1345 }
1346 }
1347 }
1aa99e6b 1348 }
cdd87c1d
KW
1349
1350 ssc_union(ssc,
1351 ored_cp_list,
1352 FALSE /* Already has been inverted */
1353 );
653099ff
GS
1354}
1355
b423522f
KW
1356PERL_STATIC_INLINE void
1357S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1358{
1359 PERL_ARGS_ASSERT_SSC_UNION;
1360
71068078 1361 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1362
1363 _invlist_union_maybe_complement_2nd(ssc->invlist,
1364 invlist,
1365 invert2nd,
1366 &ssc->invlist);
1367}
1368
1369PERL_STATIC_INLINE void
1370S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1371 SV* const invlist,
1372 const bool invert2nd)
1373{
1374 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1375
71068078 1376 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1377
1378 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1379 invlist,
1380 invert2nd,
1381 &ssc->invlist);
1382}
1383
1384PERL_STATIC_INLINE void
1385S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1386{
1387 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1388
71068078 1389 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1390
1391 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1392}
1393
1394PERL_STATIC_INLINE void
1395S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1396{
1397 /* AND just the single code point 'cp' into the SSC 'ssc' */
1398
1399 SV* cp_list = _new_invlist(2);
1400
1401 PERL_ARGS_ASSERT_SSC_CP_AND;
1402
71068078 1403 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1404
1405 cp_list = add_cp_to_invlist(cp_list, cp);
1406 ssc_intersection(ssc, cp_list,
1407 FALSE /* Not inverted */
1408 );
1409 SvREFCNT_dec_NN(cp_list);
1410}
1411
1412PERL_STATIC_INLINE void
1413S_ssc_clear_locale(pTHX_ regnode_ssc *ssc)
1414{
1415 /* Set the SSC 'ssc' to not match any locale things */
1416
1417 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1418
71068078 1419 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1420
1421 ANYOF_POSIXL_ZERO(ssc);
1422 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1423}
1424
1425STATIC void
1426S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1427{
1428 /* The inversion list in the SSC is marked mortal; now we need a more
1429 * permanent copy, which is stored the same way that is done in a regular
1430 * ANYOF node, with the first 256 code points in a bit map */
1431
1432 SV* invlist = invlist_clone(ssc->invlist);
1433
1434 PERL_ARGS_ASSERT_SSC_FINALIZE;
1435
71068078 1436 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1437
a0dd4231
KW
1438 /* The code in this file assumes that all but these flags aren't relevant
1439 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1440 * time we reach here */
eff8b7dc 1441 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
a0dd4231 1442
b423522f
KW
1443 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1444
1ee208c4
KW
1445 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1446 NULL, NULL, NULL, FALSE);
b423522f 1447
85c8e306
KW
1448 /* Make sure is clone-safe */
1449 ssc->invlist = NULL;
1450
e0e1be5f 1451 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1462525b 1452 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
e0e1be5f 1453 }
1462525b
KW
1454
1455 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1456}
1457
a3621e74
YO
1458#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1459#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1460#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1461#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1462 ? (TRIE_LIST_CUR( idx ) - 1) \
1463 : 0 )
a3621e74 1464
3dab1dad
YO
1465
1466#ifdef DEBUGGING
07be1b83 1467/*
2b8b4781
NC
1468 dump_trie(trie,widecharmap,revcharmap)
1469 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1470 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1471
1472 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1473 The _interim_ variants are used for debugging the interim
1474 tables that are used to generate the final compressed
1475 representation which is what dump_trie expects.
1476
486ec47a 1477 Part of the reason for their existence is to provide a form
3dab1dad 1478 of documentation as to how the different representations function.
07be1b83
YO
1479
1480*/
3dab1dad
YO
1481
1482/*
3dab1dad
YO
1483 Dumps the final compressed table form of the trie to Perl_debug_log.
1484 Used for debugging make_trie().
1485*/
b9a59e08 1486
3dab1dad 1487STATIC void
2b8b4781
NC
1488S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1489 AV *revcharmap, U32 depth)
3dab1dad
YO
1490{
1491 U32 state;
ab3bbdeb 1492 SV *sv=sv_newmortal();
55eed653 1493 int colwidth= widecharmap ? 6 : 4;
2e64971a 1494 U16 word;
3dab1dad
YO
1495 GET_RE_DEBUG_FLAGS_DECL;
1496
7918f24d 1497 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1498
3dab1dad
YO
1499 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1500 (int)depth * 2 + 2,"",
1501 "Match","Base","Ofs" );
1502
1503 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1504 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1505 if ( tmp ) {
538e84ed 1506 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1507 colwidth,
538e84ed 1508 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1509 PL_colors[0], PL_colors[1],
1510 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1511 PERL_PV_ESCAPE_FIRSTCHAR
1512 )
ab3bbdeb 1513 );
3dab1dad
YO
1514 }
1515 }
1516 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1517 (int)depth * 2 + 2,"");
1518
1519 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1520 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1521 PerlIO_printf( Perl_debug_log, "\n");
1522
1e2e3d02 1523 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1524 const U32 base = trie->states[ state ].trans.base;
3dab1dad 1525
538e84ed
KW
1526 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1527 (int)depth * 2 + 2,"", (UV)state);
3dab1dad
YO
1528
1529 if ( trie->states[ state ].wordnum ) {
538e84ed
KW
1530 PerlIO_printf( Perl_debug_log, " W%4X",
1531 trie->states[ state ].wordnum );
3dab1dad
YO
1532 } else {
1533 PerlIO_printf( Perl_debug_log, "%6s", "" );
1534 }
1535
1536 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1537
1538 if ( base ) {
1539 U32 ofs = 0;
1540
1541 while( ( base + ofs < trie->uniquecharcount ) ||
1542 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
1543 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1544 != state))
3dab1dad
YO
1545 ofs++;
1546
1547 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1548
1549 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
1550 if ( ( base + ofs >= trie->uniquecharcount )
1551 && ( base + ofs - trie->uniquecharcount
1552 < trie->lasttrans )
1553 && trie->trans[ base + ofs
1554 - trie->uniquecharcount ].check == state )
3dab1dad 1555 {
ab3bbdeb
YO
1556 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1557 colwidth,
538e84ed
KW
1558 (UV)trie->trans[ base + ofs
1559 - trie->uniquecharcount ].next );
3dab1dad 1560 } else {
ab3bbdeb 1561 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1562 }
1563 }
1564
1565 PerlIO_printf( Perl_debug_log, "]");
1566
1567 }
1568 PerlIO_printf( Perl_debug_log, "\n" );
1569 }
538e84ed
KW
1570 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1571 (int)depth*2, "");
2e64971a
DM
1572 for (word=1; word <= trie->wordcount; word++) {
1573 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1574 (int)word, (int)(trie->wordinfo[word].prev),
1575 (int)(trie->wordinfo[word].len));
1576 }
1577 PerlIO_printf(Perl_debug_log, "\n" );
538e84ed 1578}
3dab1dad 1579/*
3dab1dad 1580 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 1581 List tries normally only are used for construction when the number of
3dab1dad
YO
1582 possible chars (trie->uniquecharcount) is very high.
1583 Used for debugging make_trie().
1584*/
1585STATIC void
55eed653 1586S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1587 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1588 U32 depth)
3dab1dad
YO
1589{
1590 U32 state;
ab3bbdeb 1591 SV *sv=sv_newmortal();
55eed653 1592 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1593 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1594
1595 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1596
3dab1dad 1597 /* print out the table precompression. */
ab3bbdeb
YO
1598 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1599 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1600 "------:-----+-----------------\n" );
538e84ed 1601
3dab1dad
YO
1602 for( state=1 ; state < next_alloc ; state ++ ) {
1603 U16 charid;
538e84ed 1604
ab3bbdeb 1605 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1606 (int)depth * 2 + 2,"", (UV)state );
1607 if ( ! trie->states[ state ].wordnum ) {
1608 PerlIO_printf( Perl_debug_log, "%5s| ","");
1609 } else {
1610 PerlIO_printf( Perl_debug_log, "W%4x| ",
1611 trie->states[ state ].wordnum
1612 );
1613 }
1614 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
1615 SV ** const tmp = av_fetch( revcharmap,
1616 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1617 if ( tmp ) {
1618 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1619 colwidth,
538e84ed
KW
1620 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1621 colwidth,
1622 PL_colors[0], PL_colors[1],
1623 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1624 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 1625 ) ,
1e2e3d02
YO
1626 TRIE_LIST_ITEM(state,charid).forid,
1627 (UV)TRIE_LIST_ITEM(state,charid).newstate
1628 );
538e84ed 1629 if (!(charid % 10))
664e119d
RGS
1630 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1631 (int)((depth * 2) + 14), "");
1e2e3d02 1632 }
ab3bbdeb
YO
1633 }
1634 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad 1635 }
538e84ed 1636}
3dab1dad
YO
1637
1638/*
3dab1dad 1639 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
1640 This is the normal DFA style state transition table, with a few
1641 twists to facilitate compression later.
3dab1dad
YO
1642 Used for debugging make_trie().
1643*/
1644STATIC void
55eed653 1645S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1646 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1647 U32 depth)
3dab1dad
YO
1648{
1649 U32 state;
1650 U16 charid;
ab3bbdeb 1651 SV *sv=sv_newmortal();
55eed653 1652 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1653 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1654
1655 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 1656
3dab1dad
YO
1657 /*
1658 print out the table precompression so that we can do a visual check
1659 that they are identical.
1660 */
538e84ed 1661
3dab1dad
YO
1662 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1663
1664 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1665 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1666 if ( tmp ) {
538e84ed 1667 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1668 colwidth,
538e84ed 1669 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1670 PL_colors[0], PL_colors[1],
1671 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1672 PERL_PV_ESCAPE_FIRSTCHAR
1673 )
ab3bbdeb 1674 );
3dab1dad
YO
1675 }
1676 }
1677
1678 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1679
1680 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1681 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1682 }
1683
1684 PerlIO_printf( Perl_debug_log, "\n" );
1685
1686 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1687
538e84ed 1688 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
3dab1dad
YO
1689 (int)depth * 2 + 2,"",
1690 (UV)TRIE_NODENUM( state ) );
1691
1692 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1693 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1694 if (v)
1695 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1696 else
1697 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1698 }
1699 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
538e84ed
KW
1700 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1701 (UV)trie->trans[ state ].check );
3dab1dad 1702 } else {
538e84ed
KW
1703 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1704 (UV)trie->trans[ state ].check,
3dab1dad
YO
1705 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1706 }
1707 }
07be1b83 1708}
3dab1dad
YO
1709
1710#endif
1711
2e64971a 1712
786e8c11
YO
1713/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1714 startbranch: the first branch in the whole branch sequence
1715 first : start branch of sequence of branch-exact nodes.
1716 May be the same as startbranch
1717 last : Thing following the last branch.
1718 May be the same as tail.
1719 tail : item following the branch sequence
1720 count : words in the sequence
1721 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1722 depth : indent depth
3dab1dad 1723
786e8c11 1724Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1725
786e8c11
YO
1726A trie is an N'ary tree where the branches are determined by digital
1727decomposition of the key. IE, at the root node you look up the 1st character and
1728follow that branch repeat until you find the end of the branches. Nodes can be
1729marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1730
786e8c11 1731 /he|she|his|hers/
72f13be8 1732
786e8c11
YO
1733would convert into the following structure. Numbers represent states, letters
1734following numbers represent valid transitions on the letter from that state, if
1735the number is in square brackets it represents an accepting state, otherwise it
1736will be in parenthesis.
07be1b83 1737
786e8c11
YO
1738 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1739 | |
1740 | (2)
1741 | |
1742 (1) +-i->(6)-+-s->[7]
1743 |
1744 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1745
786e8c11
YO
1746 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1747
1748This shows that when matching against the string 'hers' we will begin at state 1
1749read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1750then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1751is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1752single traverse. We store a mapping from accepting to state to which word was
1753matched, and then when we have multiple possibilities we try to complete the
1754rest of the regex in the order in which they occured in the alternation.
1755
1756The only prior NFA like behaviour that would be changed by the TRIE support is
1757the silent ignoring of duplicate alternations which are of the form:
1758
1759 / (DUPE|DUPE) X? (?{ ... }) Y /x
1760
4b714af6 1761Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1762and without the optimisation. With the optimisations dupes will be silently
486ec47a 1763ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1764the following demonstrates:
1765
1766 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1767
1768which prints out 'word' three times, but
1769
1770 'words'=~/(word|word|word)(?{ print $1 })S/
1771
1772which doesnt print it out at all. This is due to other optimisations kicking in.
1773
1774Example of what happens on a structural level:
1775
486ec47a 1776The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1777
1778 1: CURLYM[1] {1,32767}(18)
1779 5: BRANCH(8)
1780 6: EXACT <ac>(16)
1781 8: BRANCH(11)
1782 9: EXACT <ad>(16)
1783 11: BRANCH(14)
1784 12: EXACT <ab>(16)
1785 16: SUCCEED(0)
1786 17: NOTHING(18)
1787 18: END(0)
1788
1789This would be optimizable with startbranch=5, first=5, last=16, tail=16
1790and should turn into:
1791
1792 1: CURLYM[1] {1,32767}(18)
1793 5: TRIE(16)
1794 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1795 <ac>
1796 <ad>
1797 <ab>
1798 16: SUCCEED(0)
1799 17: NOTHING(18)
1800 18: END(0)
1801
1802Cases where tail != last would be like /(?foo|bar)baz/:
1803
1804 1: BRANCH(4)
1805 2: EXACT <foo>(8)
1806 4: BRANCH(7)
1807 5: EXACT <bar>(8)
1808 7: TAIL(8)
1809 8: EXACT <baz>(10)
1810 10: END(0)
1811
1812which would be optimizable with startbranch=1, first=1, last=7, tail=8
1813and would end up looking like:
1814
1815 1: TRIE(8)
1816 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1817 <foo>
1818 <bar>
1819 7: TAIL(8)
1820 8: EXACT <baz>(10)
1821 10: END(0)
1822
c80e42f3 1823 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
1824
1825is the recommended Unicode-aware way of saying
1826
1827 *(d++) = uv;
1828*/
1829
fab2782b 1830#define TRIE_STORE_REVCHAR(val) \
786e8c11 1831 STMT_START { \
73031816 1832 if (UTF) { \
fab2782b 1833 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1834 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 1835 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
1836 SvCUR_set(zlopp, kapow - flrbbbbb); \
1837 SvPOK_on(zlopp); \
1838 SvUTF8_on(zlopp); \
1839 av_push(revcharmap, zlopp); \
1840 } else { \
fab2782b 1841 char ooooff = (char)val; \
73031816
NC
1842 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1843 } \
1844 } STMT_END
786e8c11 1845
914a25d5
KW
1846/* This gets the next character from the input, folding it if not already
1847 * folded. */
1848#define TRIE_READ_CHAR STMT_START { \
1849 wordlen++; \
1850 if ( UTF ) { \
1851 /* if it is UTF then it is either already folded, or does not need \
1852 * folding */ \
1c1d615a 1853 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
1854 } \
1855 else if (folder == PL_fold_latin1) { \
7d006b13
KW
1856 /* This folder implies Unicode rules, which in the range expressible \
1857 * by not UTF is the lower case, with the two exceptions, one of \
1858 * which should have been taken care of before calling this */ \
1859 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1860 uvc = toLOWER_L1(*uc); \
1861 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1862 len = 1; \
914a25d5
KW
1863 } else { \
1864 /* raw data, will be folded later if needed */ \
1865 uvc = (U32)*uc; \
1866 len = 1; \
1867 } \
786e8c11
YO
1868} STMT_END
1869
1870
1871
1872#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1873 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1874 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1875 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1876 } \
1877 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1878 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1879 TRIE_LIST_CUR( state )++; \
1880} STMT_END
07be1b83 1881
786e8c11
YO
1882#define TRIE_LIST_NEW(state) STMT_START { \
1883 Newxz( trie->states[ state ].trans.list, \
1884 4, reg_trie_trans_le ); \
1885 TRIE_LIST_CUR( state ) = 1; \
1886 TRIE_LIST_LEN( state ) = 4; \
1887} STMT_END
07be1b83 1888
786e8c11
YO
1889#define TRIE_HANDLE_WORD(state) STMT_START { \
1890 U16 dupe= trie->states[ state ].wordnum; \
1891 regnode * const noper_next = regnext( noper ); \
1892 \
786e8c11
YO
1893 DEBUG_r({ \
1894 /* store the word for dumping */ \
1895 SV* tmp; \
1896 if (OP(noper) != NOTHING) \
740cce10 1897 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1898 else \
740cce10 1899 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1900 av_push( trie_words, tmp ); \
786e8c11
YO
1901 }); \
1902 \
1903 curword++; \
2e64971a
DM
1904 trie->wordinfo[curword].prev = 0; \
1905 trie->wordinfo[curword].len = wordlen; \
1906 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1907 \
1908 if ( noper_next < tail ) { \
1909 if (!trie->jump) \
538e84ed
KW
1910 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1911 sizeof(U16) ); \
7f69552c 1912 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1913 if (!jumper) \
1914 jumper = noper_next; \
1915 if (!nextbranch) \
1916 nextbranch= regnext(cur); \
1917 } \
1918 \
1919 if ( dupe ) { \
2e64971a
DM
1920 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1921 /* chain, so that when the bits of chain are later */\
1922 /* linked together, the dups appear in the chain */\
1923 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1924 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1925 } else { \
1926 /* we haven't inserted this word yet. */ \
1927 trie->states[ state ].wordnum = curword; \
1928 } \
1929} STMT_END
07be1b83 1930
3dab1dad 1931
786e8c11
YO
1932#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1933 ( ( base + charid >= ucharcount \
1934 && base + charid < ubound \
1935 && state == trie->trans[ base - ucharcount + charid ].check \
1936 && trie->trans[ base - ucharcount + charid ].next ) \
1937 ? trie->trans[ base - ucharcount + charid ].next \
1938 : ( state==1 ? special : 0 ) \
1939 )
3dab1dad 1940
786e8c11
YO
1941#define MADE_TRIE 1
1942#define MADE_JUMP_TRIE 2
1943#define MADE_EXACT_TRIE 4
3dab1dad 1944
a3621e74 1945STATIC I32
538e84ed
KW
1946S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1947 regnode *first, regnode *last, regnode *tail,
1948 U32 word_count, U32 flags, U32 depth)
a3621e74 1949{
27da23d5 1950 dVAR;
a3621e74
YO
1951 /* first pass, loop through and scan words */
1952 reg_trie_data *trie;
55eed653 1953 HV *widecharmap = NULL;
2b8b4781 1954 AV *revcharmap = newAV();
a3621e74 1955 regnode *cur;
a3621e74
YO
1956 STRLEN len = 0;
1957 UV uvc = 0;
1958 U16 curword = 0;
1959 U32 next_alloc = 0;
786e8c11
YO
1960 regnode *jumper = NULL;
1961 regnode *nextbranch = NULL;
7f69552c 1962 regnode *convert = NULL;
2e64971a 1963 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1964 /* we just use folder as a flag in utf8 */
1e696034 1965 const U8 * folder = NULL;
a3621e74 1966
2b8b4781 1967#ifdef DEBUGGING
cf78de0b 1968 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
1969 AV *trie_words = NULL;
1970 /* along with revcharmap, this only used during construction but both are
1971 * useful during debugging so we store them in the struct when debugging.
8e11feef 1972 */
2b8b4781 1973#else
cf78de0b 1974 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 1975 STRLEN trie_charcount=0;
3dab1dad 1976#endif
2b8b4781 1977 SV *re_trie_maxbuff;
a3621e74 1978 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1979
1980 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1981#ifndef DEBUGGING
1982 PERL_UNUSED_ARG(depth);
1983#endif
a3621e74 1984
1e696034 1985 switch (flags) {
79a81a6e 1986 case EXACT: break;
2f7f8cb1 1987 case EXACTFA:
fab2782b 1988 case EXACTFU_SS:
1e696034
KW
1989 case EXACTFU: folder = PL_fold_latin1; break;
1990 case EXACTF: folder = PL_fold; break;
fab2782b 1991 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1992 }
1993
c944940b 1994 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1995 trie->refcount = 1;
3dab1dad 1996 trie->startstate = 1;
786e8c11 1997 trie->wordcount = word_count;
f8fc2ecf 1998 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1999 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 2000 if (flags == EXACT)
c944940b 2001 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2002 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2003 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2004
a3621e74 2005 DEBUG_r({
2b8b4781 2006 trie_words = newAV();
a3621e74 2007 });
a3621e74 2008
0111c4fd 2009 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 2010 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2011 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2012 }
df826430 2013 DEBUG_TRIE_COMPILE_r({
538e84ed
KW
2014 PerlIO_printf( Perl_debug_log,
2015 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2016 (int)depth * 2 + 2, "",
2017 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2018 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2019 });
538e84ed 2020
7f69552c
YO
2021 /* Find the node we are going to overwrite */
2022 if ( first == startbranch && OP( last ) != BRANCH ) {
2023 /* whole branch chain */
2024 convert = first;
2025 } else {
2026 /* branch sub-chain */
2027 convert = NEXTOPER( first );
2028 }
538e84ed 2029
a3621e74
YO
2030 /* -- First loop and Setup --
2031
2032 We first traverse the branches and scan each word to determine if it
2033 contains widechars, and how many unique chars there are, this is
2034 important as we have to build a table with at least as many columns as we
2035 have unique chars.
2036
2037 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2038 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2039 the native representation of the character value as the key and IV's for
2040 the coded index.
a3621e74
YO
2041
2042 *TODO* If we keep track of how many times each character is used we can
2043 remap the columns so that the table compression later on is more
3b753521 2044 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2045 middle and the least common are on the outside. IMO this would be better
2046 than a most to least common mapping as theres a decent chance the most
2047 common letter will share a node with the least common, meaning the node
486ec47a 2048 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2049 case is when we have the least common nodes twice.
2050
2051 */
2052
a3621e74 2053 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2054 regnode *noper = NEXTOPER( cur );
e1ec3a88 2055 const U8 *uc = (U8*)STRING( noper );
df826430 2056 const U8 *e = uc + STR_LEN( noper );
bc031a7d 2057 int foldlen = 0;
07be1b83 2058 U32 wordlen = 0; /* required init */
bc031a7d
KW
2059 STRLEN minchars = 0;
2060 STRLEN maxchars = 0;
538e84ed
KW
2061 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2062 bitmap?*/
a3621e74 2063
3dab1dad 2064 if (OP(noper) == NOTHING) {
df826430
YO
2065 regnode *noper_next= regnext(noper);
2066 if (noper_next != tail && OP(noper_next) == flags) {
2067 noper = noper_next;
2068 uc= (U8*)STRING(noper);
2069 e= uc + STR_LEN(noper);
2070 trie->minlen= STR_LEN(noper);
2071 } else {
2072 trie->minlen= 0;
2073 continue;
2074 }
3dab1dad 2075 }
df826430 2076
fab2782b 2077 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2078 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2079 regardless of encoding */
fab2782b
YO
2080 if (OP( noper ) == EXACTFU_SS) {
2081 /* false positives are ok, so just set this */
0dc4a61d 2082 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2083 }
2084 }
bc031a7d
KW
2085 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2086 branch */
3dab1dad 2087 TRIE_CHARCOUNT(trie)++;
a3621e74 2088 TRIE_READ_CHAR;
645de4ce 2089
bc031a7d
KW
2090 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2091 * is in effect. Under /i, this character can match itself, or
2092 * anything that folds to it. If not under /i, it can match just
2093 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2094 * all fold to k, and all are single characters. But some folds
2095 * expand to more than one character, so for example LATIN SMALL
2096 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2097 * the string beginning at 'uc' is 'ffi', it could be matched by
2098 * three characters, or just by the one ligature character. (It
2099 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2100 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2101 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2102 * match.) The trie needs to know the minimum and maximum number
2103 * of characters that could match so that it can use size alone to
2104 * quickly reject many match attempts. The max is simple: it is
2105 * the number of folded characters in this branch (since a fold is
2106 * never shorter than what folds to it. */
2107
2108 maxchars++;
2109
2110 /* And the min is equal to the max if not under /i (indicated by
2111 * 'folder' being NULL), or there are no multi-character folds. If
2112 * there is a multi-character fold, the min is incremented just
2113 * once, for the character that folds to the sequence. Each
2114 * character in the sequence needs to be added to the list below of
2115 * characters in the trie, but we count only the first towards the
2116 * min number of characters needed. This is done through the
2117 * variable 'foldlen', which is returned by the macros that look
2118 * for these sequences as the number of bytes the sequence
2119 * occupies. Each time through the loop, we decrement 'foldlen' by
2120 * how many bytes the current char occupies. Only when it reaches
2121 * 0 do we increment 'minchars' or look for another multi-character
2122 * sequence. */
2123 if (folder == NULL) {
2124 minchars++;
2125 }
2126 else if (foldlen > 0) {
2127 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2128 }
2129 else {
bc031a7d
KW
2130 minchars++;
2131
2132 /* See if *uc is the beginning of a multi-character fold. If
2133 * so, we decrement the length remaining to look at, to account
2134 * for the current character this iteration. (We can use 'uc'
2135 * instead of the fold returned by TRIE_READ_CHAR because for
2136 * non-UTF, the latin1_safe macro is smart enough to account
2137 * for all the unfolded characters, and because for UTF, the
2138 * string will already have been folded earlier in the
2139 * compilation process */
2140 if (UTF) {
2141 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2142 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2143 }
2144 }
bc031a7d
KW
2145 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2146 foldlen--;
2147 }
645de4ce 2148 }
bc031a7d
KW
2149
2150 /* The current character (and any potential folds) should be added
2151 * to the possible matching characters for this position in this
2152 * branch */
a3621e74 2153 if ( uvc < 256 ) {
fab2782b
YO
2154 if ( folder ) {
2155 U8 folded= folder[ (U8) uvc ];
2156 if ( !trie->charmap[ folded ] ) {
2157 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2158 TRIE_STORE_REVCHAR( folded );
2159 }
2160 }
a3621e74
YO
2161 if ( !trie->charmap[ uvc ] ) {
2162 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2163 TRIE_STORE_REVCHAR( uvc );
a3621e74 2164 }
02daf0ab 2165 if ( set_bit ) {
62012aee
KW
2166 /* store the codepoint in the bitmap, and its folded
2167 * equivalent. */
fab2782b 2168 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2169
2170 /* store the folded codepoint */
fab2782b 2171 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2172
2173 if ( !UTF ) {
2174 /* store first byte of utf8 representation of
acdf4139 2175 variant codepoints */
6f2d5cbc 2176 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2177 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2178 }
2179 }
02daf0ab
YO
2180 set_bit = 0; /* We've done our bit :-) */
2181 }
a3621e74 2182 } else {
bc031a7d
KW
2183
2184 /* XXX We could come up with the list of code points that fold
2185 * to this using PL_utf8_foldclosures, except not for
2186 * multi-char folds, as there may be multiple combinations
2187 * there that could work, which needs to wait until runtime to
2188 * resolve (The comment about LIGATURE FFI above is such an
2189 * example */
2190
a3621e74 2191 SV** svpp;
55eed653
NC
2192 if ( !widecharmap )
2193 widecharmap = newHV();
a3621e74 2194
55eed653 2195 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2196
2197 if ( !svpp )
e4584336 2198 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2199
2200 if ( !SvTRUE( *svpp ) ) {
2201 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2202 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2203 }
2204 }
bc031a7d
KW
2205 } /* end loop through characters in this branch of the trie */
2206
2207 /* We take the min and max for this branch and combine to find the min
2208 * and max for all branches processed so far */
3dab1dad 2209 if( cur == first ) {
bc031a7d
KW
2210 trie->minlen = minchars;
2211 trie->maxlen = maxchars;
2212 } else if (minchars < trie->minlen) {
2213 trie->minlen = minchars;
2214 } else if (maxchars > trie->maxlen) {
2215 trie->maxlen = maxchars;
fab2782b 2216 }
a3621e74
YO
2217 } /* end first pass */
2218 DEBUG_TRIE_COMPILE_r(
538e84ed
KW
2219 PerlIO_printf( Perl_debug_log,
2220 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3dab1dad 2221 (int)depth * 2 + 2,"",
55eed653 2222 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2223 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2224 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2225 );
a3621e74
YO
2226
2227 /*
2228 We now know what we are dealing with in terms of unique chars and
2229 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2230 representation using a flat table will take. If it's over a reasonable
2231 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2232 conservative but potentially much slower representation using an array
2233 of lists.
2234
2235 At the end we convert both representations into the same compressed
2236 form that will be used in regexec.c for matching with. The latter
2237 is a form that cannot be used to construct with but has memory
2238 properties similar to the list form and access properties similar
2239 to the table form making it both suitable for fast searches and
2240 small enough that its feasable to store for the duration of a program.
2241
2242 See the comment in the code where the compressed table is produced
2243 inplace from the flat tabe representation for an explanation of how
2244 the compression works.
2245
2246 */
2247
2248
2e64971a
DM
2249 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2250 prev_states[1] = 0;
2251
538e84ed
KW
2252 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2253 > SvIV(re_trie_maxbuff) )
2254 {
a3621e74
YO
2255 /*
2256 Second Pass -- Array Of Lists Representation
2257
2258 Each state will be represented by a list of charid:state records
2259 (reg_trie_trans_le) the first such element holds the CUR and LEN
2260 points of the allocated array. (See defines above).
2261
2262 We build the initial structure using the lists, and then convert
2263 it into the compressed table form which allows faster lookups
2264 (but cant be modified once converted).
a3621e74
YO
2265 */
2266
a3621e74
YO
2267 STRLEN transcount = 1;
2268
538e84ed 2269 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2270 "%*sCompiling trie using list compiler\n",
2271 (int)depth * 2 + 2, ""));
686b73d4 2272
c944940b
JH
2273 trie->states = (reg_trie_state *)
2274 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2275 sizeof(reg_trie_state) );
a3621e74
YO
2276 TRIE_LIST_NEW(1);
2277 next_alloc = 2;
2278
2279 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2280
df826430 2281 regnode *noper = NEXTOPER( cur );
c445ea15 2282 U8 *uc = (U8*)STRING( noper );
df826430 2283 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
2284 U32 state = 1; /* required init */
2285 U16 charid = 0; /* sanity init */
07be1b83 2286 U32 wordlen = 0; /* required init */
c445ea15 2287
df826430
YO
2288 if (OP(noper) == NOTHING) {
2289 regnode *noper_next= regnext(noper);
2290 if (noper_next != tail && OP(noper_next) == flags) {
2291 noper = noper_next;
2292 uc= (U8*)STRING(noper);
2293 e= uc + STR_LEN(noper);
2294 }
2295 }
2296
3dab1dad 2297 if (OP(noper) != NOTHING) {
786e8c11 2298 for ( ; uc < e ; uc += len ) {
c445ea15 2299
786e8c11 2300 TRIE_READ_CHAR;
c445ea15 2301
786e8c11
YO
2302 if ( uvc < 256 ) {
2303 charid = trie->charmap[ uvc ];
c445ea15 2304 } else {
538e84ed
KW
2305 SV** const svpp = hv_fetch( widecharmap,
2306 (char*)&uvc,
2307 sizeof( UV ),
2308 0);
786e8c11
YO
2309 if ( !svpp ) {
2310 charid = 0;
2311 } else {
2312 charid=(U16)SvIV( *svpp );
2313 }
c445ea15 2314 }
538e84ed
KW
2315 /* charid is now 0 if we dont know the char read, or
2316 * nonzero if we do */
786e8c11 2317 if ( charid ) {
a3621e74 2318
786e8c11
YO
2319 U16 check;
2320 U32 newstate = 0;
a3621e74 2321
786e8c11
YO
2322 charid--;
2323 if ( !trie->states[ state ].trans.list ) {
2324 TRIE_LIST_NEW( state );
c445ea15 2325 }
538e84ed
KW
2326 for ( check = 1;
2327 check <= TRIE_LIST_USED( state );
2328 check++ )
2329 {
2330 if ( TRIE_LIST_ITEM( state, check ).forid
2331 == charid )
2332 {
786e8c11
YO
2333 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2334 break;
2335 }
2336 }
2337 if ( ! newstate ) {
2338 newstate = next_alloc++;
2e64971a 2339 prev_states[newstate] = state;
786e8c11
YO
2340 TRIE_LIST_PUSH( state, charid, newstate );
2341 transcount++;
2342 }
2343 state = newstate;
2344 } else {
2345 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2346 }
a28509cc 2347 }
c445ea15 2348 }
3dab1dad 2349 TRIE_HANDLE_WORD(state);
a3621e74
YO
2350
2351 } /* end second pass */
2352
1e2e3d02 2353 /* next alloc is the NEXT state to be allocated */
538e84ed 2354 trie->statecount = next_alloc;
c944940b
JH
2355 trie->states = (reg_trie_state *)
2356 PerlMemShared_realloc( trie->states,
2357 next_alloc
2358 * sizeof(reg_trie_state) );
a3621e74 2359
3dab1dad 2360 /* and now dump it out before we compress it */
2b8b4781
NC
2361 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2362 revcharmap, next_alloc,
2363 depth+1)
1e2e3d02 2364 );
a3621e74 2365
c944940b
JH
2366 trie->trans = (reg_trie_trans *)
2367 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2368 {
2369 U32 state;
a3621e74
YO
2370 U32 tp = 0;
2371 U32 zp = 0;
2372
2373
2374 for( state=1 ; state < next_alloc ; state ++ ) {
2375 U32 base=0;
2376
2377 /*
2378 DEBUG_TRIE_COMPILE_MORE_r(
2379 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2380 );
2381 */
2382
2383 if (trie->states[state].trans.list) {
2384 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2385 U16 maxid=minid;
a28509cc 2386 U16 idx;
a3621e74
YO
2387
2388 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2389 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2390 if ( forid < minid ) {
2391 minid=forid;
2392 } else if ( forid > maxid ) {
2393 maxid=forid;
2394 }
a3621e74
YO
2395 }
2396 if ( transcount < tp + maxid - minid + 1) {
2397 transcount *= 2;
c944940b
JH
2398 trie->trans = (reg_trie_trans *)
2399 PerlMemShared_realloc( trie->trans,
446bd890
NC
2400 transcount
2401 * sizeof(reg_trie_trans) );
538e84ed
KW
2402 Zero( trie->trans + (transcount / 2),
2403 transcount / 2,
2404 reg_trie_trans );
a3621e74
YO
2405 }
2406 base = trie->uniquecharcount + tp - minid;
2407 if ( maxid == minid ) {
2408 U32 set = 0;
2409 for ( ; zp < tp ; zp++ ) {
2410 if ( ! trie->trans[ zp ].next ) {
2411 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2412 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2413 1).newstate;
a3621e74
YO
2414 trie->trans[ zp ].check = state;
2415 set = 1;
2416 break;
2417 }
2418 }
2419 if ( !set ) {
538e84ed
KW
2420 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2421 1).newstate;
a3621e74
YO
2422 trie->trans[ tp ].check = state;
2423 tp++;
2424 zp = tp;
2425 }
2426 } else {
2427 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2428 const U32 tid = base
2429 - trie->uniquecharcount
2430 + TRIE_LIST_ITEM( state, idx ).forid;
2431 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2432 idx ).newstate;
a3621e74
YO
2433 trie->trans[ tid ].check = state;
2434 }
2435 tp += ( maxid - minid + 1 );
2436 }
2437 Safefree(trie->states[ state ].trans.list);
2438 }
2439 /*
2440 DEBUG_TRIE_COMPILE_MORE_r(
2441 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2442 );
2443 */
2444 trie->states[ state ].trans.base=base;
2445 }
cc601c31 2446 trie->lasttrans = tp + 1;
a3621e74
YO
2447 }
2448 } else {
2449 /*
2450 Second Pass -- Flat Table Representation.
2451
b423522f
KW
2452 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2453 each. We know that we will need Charcount+1 trans at most to store
2454 the data (one row per char at worst case) So we preallocate both
2455 structures assuming worst case.
a3621e74
YO
2456
2457 We then construct the trie using only the .next slots of the entry
2458 structs.
2459
b423522f
KW
2460 We use the .check field of the first entry of the node temporarily
2461 to make compression both faster and easier by keeping track of how
2462 many non zero fields are in the node.
a3621e74
YO
2463
2464 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2465 transition.
2466
b423522f
KW
2467 There are two terms at use here: state as a TRIE_NODEIDX() which is
2468 a number representing the first entry of the node, and state as a
2469 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2470 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2471 if there are 2 entrys per node. eg:
a3621e74
YO
2472
2473 A B A B
2474 1. 2 4 1. 3 7
2475 2. 0 3 3. 0 5
2476 3. 0 0 5. 0 0
2477 4. 0 0 7. 0 0
2478
b423522f
KW
2479 The table is internally in the right hand, idx form. However as we
2480 also have to deal with the states array which is indexed by nodenum
2481 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2482
2483 */
538e84ed 2484 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2485 "%*sCompiling trie using table compiler\n",
2486 (int)depth * 2 + 2, ""));
3dab1dad 2487
c944940b
JH
2488 trie->trans = (reg_trie_trans *)
2489 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2490 * trie->uniquecharcount + 1,
2491 sizeof(reg_trie_trans) );
2492 trie->states = (reg_trie_state *)
2493 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2494 sizeof(reg_trie_state) );
a3621e74
YO
2495 next_alloc = trie->uniquecharcount + 1;
2496
3dab1dad 2497
a3621e74
YO
2498 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2499
df826430 2500 regnode *noper = NEXTOPER( cur );
a28509cc 2501 const U8 *uc = (U8*)STRING( noper );
df826430 2502 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2503
2504 U32 state = 1; /* required init */
2505
2506 U16 charid = 0; /* sanity init */
2507 U32 accept_state = 0; /* sanity init */
a3621e74 2508
07be1b83 2509 U32 wordlen = 0; /* required init */
a3621e74 2510
df826430
YO
2511 if (OP(noper) == NOTHING) {
2512 regnode *noper_next= regnext(noper);
2513 if (noper_next != tail && OP(noper_next) == flags) {
2514 noper = noper_next;
2515 uc= (U8*)STRING(noper);
2516 e= uc + STR_LEN(noper);
2517 }
2518 }
fab2782b 2519
3dab1dad 2520 if ( OP(noper) != NOTHING ) {
786e8c11 2521 for ( ; uc < e ; uc += len ) {
a3621e74 2522
786e8c11 2523 TRIE_READ_CHAR;
a3621e74 2524
786e8c11
YO
2525 if ( uvc < 256 ) {
2526 charid = trie->charmap[ uvc ];
2527 } else {
538e84ed
KW
2528 SV* const * const svpp = hv_fetch( widecharmap,
2529 (char*)&uvc,
2530 sizeof( UV ),
2531 0);
786e8c11 2532 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2533 }
786e8c11
YO
2534 if ( charid ) {
2535 charid--;
2536 if ( !trie->trans[ state + charid ].next ) {
2537 trie->trans[ state + charid ].next = next_alloc;
2538 trie->trans[ state ].check++;
2e64971a
DM
2539 prev_states[TRIE_NODENUM(next_alloc)]
2540 = TRIE_NODENUM(state);
786e8c11
YO
2541 next_alloc += trie->uniquecharcount;
2542 }
2543 state = trie->trans[ state + charid ].next;
2544 } else {
2545 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2546 }
538e84ed
KW
2547 /* charid is now 0 if we dont know the char read, or
2548 * nonzero if we do */
a3621e74 2549 }
a3621e74 2550 }
3dab1dad
YO
2551 accept_state = TRIE_NODENUM( state );
2552 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2553
2554 } /* end second pass */
2555
3dab1dad 2556 /* and now dump it out before we compress it */
2b8b4781
NC
2557 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2558 revcharmap,
2559 next_alloc, depth+1));
a3621e74 2560
a3621e74
YO
2561 {
2562 /*
2563 * Inplace compress the table.*
2564
2565 For sparse data sets the table constructed by the trie algorithm will
2566 be mostly 0/FAIL transitions or to put it another way mostly empty.
2567 (Note that leaf nodes will not contain any transitions.)
2568
2569 This algorithm compresses the tables by eliminating most such
2570 transitions, at the cost of a modest bit of extra work during lookup:
2571
2572 - Each states[] entry contains a .base field which indicates the
2573 index in the state[] array wheres its transition data is stored.
2574
3b753521 2575 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2576
2577 - If .base is nonzero then charid is added to it to find an entry in
2578 the trans array.
2579
2580 -If trans[states[state].base+charid].check!=state then the
2581 transition is taken to be a 0/Fail transition. Thus if there are fail
2582 transitions at the front of the node then the .base offset will point
2583 somewhere inside the previous nodes data (or maybe even into a node
2584 even earlier), but the .check field determines if the transition is
2585 valid.
2586
786e8c11 2587 XXX - wrong maybe?
a3621e74 2588 The following process inplace converts the table to the compressed
3b753521 2589 table: We first do not compress the root node 1,and mark all its
a3621e74 2590 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2591 allows us to do a DFA construction from the compressed table later,
2592 and ensures that any .base pointers we calculate later are greater
2593 than 0.
a3621e74
YO
2594
2595 - We set 'pos' to indicate the first entry of the second node.
2596
2597 - We then iterate over the columns of the node, finding the first and
2598 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2599 and set the .check pointers accordingly, and advance pos
2600 appropriately and repreat for the next node. Note that when we copy
2601 the next pointers we have to convert them from the original
2602 NODEIDX form to NODENUM form as the former is not valid post
2603 compression.
2604
2605 - If a node has no transitions used we mark its base as 0 and do not
2606 advance the pos pointer.
2607
2608 - If a node only has one transition we use a second pointer into the
2609 structure to fill in allocated fail transitions from other states.
2610 This pointer is independent of the main pointer and scans forward
2611 looking for null transitions that are allocated to a state. When it
2612 finds one it writes the single transition into the "hole". If the
786e8c11 2613 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2614
2615 - Once compressed we can Renew/realloc the structures to release the
2616 excess space.
2617
2618 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2619 specifically Fig 3.47 and the associated pseudocode.
2620
2621 demq
2622 */
a3b680e6 2623 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2624 U32 state, charid;
a3621e74 2625 U32 pos = 0, zp=0;
1e2e3d02 2626 trie->statecount = laststate;
a3621e74
YO
2627
2628 for ( state = 1 ; state < laststate ; state++ ) {
2629 U8 flag = 0;
a28509cc
AL
2630 const U32 stateidx = TRIE_NODEIDX( state );
2631 const U32 o_used = trie->trans[ stateidx ].check;
2632 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2633 trie->trans[ stateidx ].check = 0;
2634
538e84ed
KW
2635 for ( charid = 0;
2636 used && charid < trie->uniquecharcount;
2637 charid++ )
2638 {
a3621e74
YO
2639 if ( flag || trie->trans[ stateidx + charid ].next ) {
2640 if ( trie->trans[ stateidx + charid ].next ) {
2641 if (o_used == 1) {
2642 for ( ; zp < pos ; zp++ ) {
2643 if ( ! trie->trans[ zp ].next ) {
2644 break;
2645 }
2646 }
538e84ed
KW
2647 trie->states[ state ].trans.base
2648 = zp
2649 + trie->uniquecharcount
2650 - charid ;
2651 trie->trans[ zp ].next
2652 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2653 + charid ].next );
a3621e74
YO
2654 trie->trans[ zp ].check = state;
2655 if ( ++zp > pos ) pos = zp;
2656 break;
2657 }
2658 used--;
2659 }
2660 if ( !flag ) {
2661 flag = 1;
538e84ed
KW
2662 trie->states[ state ].trans.base
2663 = pos + trie->uniquecharcount - charid ;
a3621e74 2664 }
538e84ed
KW
2665 trie->trans[ pos ].next
2666 = SAFE_TRIE_NODENUM(
2667 trie->trans[ stateidx + charid ].next );
a3621e74
YO
2668 trie->trans[ pos ].check = state;
2669 pos++;
2670 }
2671 }
2672 }
cc601c31 2673 trie->lasttrans = pos + 1;
c944940b
JH
2674 trie->states = (reg_trie_state *)
2675 PerlMemShared_realloc( trie->states, laststate
2676 * sizeof(reg_trie_state) );
a3621e74 2677 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2678 PerlIO_printf( Perl_debug_log,
2679 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2680 (int)depth * 2 + 2,"",
2681 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2682 + 1 ),
2683 (IV)next_alloc,
2684 (IV)pos,
2685 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
2686 );
2687
2688 } /* end table compress */
2689 }
1e2e3d02 2690 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2691 PerlIO_printf(Perl_debug_log,
2692 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1e2e3d02
YO
2693 (int)depth * 2 + 2, "",
2694 (UV)trie->statecount,
2695 (UV)trie->lasttrans)
2696 );
cc601c31 2697 /* resize the trans array to remove unused space */
c944940b
JH
2698 trie->trans = (reg_trie_trans *)
2699 PerlMemShared_realloc( trie->trans, trie->lasttrans
2700 * sizeof(reg_trie_trans) );
a3621e74 2701
538e84ed 2702 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2703 U8 nodetype =(U8)(flags & 0xFF);
2704 char *str=NULL;
538e84ed 2705
07be1b83 2706#ifdef DEBUGGING
e62cc96a 2707 regnode *optimize = NULL;
7122b237
YO
2708#ifdef RE_TRACK_PATTERN_OFFSETS
2709
b57a0404
JH
2710 U32 mjd_offset = 0;
2711 U32 mjd_nodelen = 0;
7122b237
YO
2712#endif /* RE_TRACK_PATTERN_OFFSETS */
2713#endif /* DEBUGGING */
a3621e74 2714 /*
3dab1dad
YO
2715 This means we convert either the first branch or the first Exact,
2716 depending on whether the thing following (in 'last') is a branch
2717 or not and whther first is the startbranch (ie is it a sub part of
2718 the alternation or is it the whole thing.)
3b753521 2719 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2720 the whole branch sequence, including the first.
a3621e74 2721 */
3dab1dad 2722 /* Find the node we are going to overwrite */
7f69552c 2723 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2724 /* branch sub-chain */
3dab1dad 2725 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2726#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2727 DEBUG_r({
2728 mjd_offset= Node_Offset((convert));
2729 mjd_nodelen= Node_Length((convert));
2730 });
7122b237 2731#endif
7f69552c 2732 /* whole branch chain */
7122b237
YO
2733 }
2734#ifdef RE_TRACK_PATTERN_OFFSETS
2735 else {
7f69552c
YO
2736 DEBUG_r({
2737 const regnode *nop = NEXTOPER( convert );
2738 mjd_offset= Node_Offset((nop));
2739 mjd_nodelen= Node_Length((nop));
2740 });
07be1b83
YO
2741 }
2742 DEBUG_OPTIMISE_r(
538e84ed
KW
2743 PerlIO_printf(Perl_debug_log,
2744 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
07be1b83 2745 (int)depth * 2 + 2, "",
786e8c11 2746 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2747 );
7122b237 2748#endif
538e84ed 2749 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
2750 split out as an EXACT and put in front of the TRIE node. */
2751 trie->startstate= 1;
55eed653 2752 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2753 U32 state;
1e2e3d02 2754 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2755 U32 ofs = 0;
8e11feef
RGS
2756 I32 idx = -1;
2757 U32 count = 0;
2758 const U32 base = trie->states[ state ].trans.base;
a3621e74 2759
3dab1dad 2760 if ( trie->states[state].wordnum )
8e11feef 2761 count = 1;
a3621e74 2762
8e11feef 2763 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2764 if ( ( base + ofs >= trie->uniquecharcount ) &&
2765 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2766 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2767 {
3dab1dad 2768 if ( ++count > 1 ) {
2b8b4781 2769 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2770 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2771 if ( state == 1 ) break;
3dab1dad
YO
2772 if ( count == 2 ) {
2773 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2774 DEBUG_OPTIMISE_r(
8e11feef
RGS
2775 PerlIO_printf(Perl_debug_log,
2776 "%*sNew Start State=%"UVuf" Class: [",
2777 (int)depth * 2 + 2, "",
786e8c11 2778 (UV)state));
be8e71aa 2779 if (idx >= 0) {
2b8b4781 2780 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2781 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2782
3dab1dad 2783 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2784 if ( folder )
2785 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2786 DEBUG_OPTIMISE_r(
f1f66076 2787 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2788 );
8e11feef
RGS
2789 }
2790 }
2791 TRIE_BITMAP_SET(trie,*ch);
2792 if ( folder )
2793 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2794 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2795 }
2796 idx = ofs;
2797 }
3dab1dad
YO
2798 }
2799 if ( count == 1 ) {
2b8b4781 2800 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2801 STRLEN len;
2802 char *ch = SvPV( *tmp, len );
de734bd5
A
2803 DEBUG_OPTIMISE_r({
2804 SV *sv=sv_newmortal();
8e11feef
RGS
2805 PerlIO_printf( Perl_debug_log,
2806 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2807 (int)depth * 2 + 2, "",
538e84ed
KW
2808 (UV)state, (UV)idx,
2809 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
2810 PL_colors[0], PL_colors[1],
2811 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 2812 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
2813 )
2814 );
2815 });
3dab1dad
YO
2816 if ( state==1 ) {
2817 OP( convert ) = nodetype;
2818 str=STRING(convert);
2819 STR_LEN(convert)=0;
2820 }
c490c714
YO
2821 STR_LEN(convert) += len;
2822 while (len--)
de734bd5 2823 *str++ = *ch++;
8e11feef 2824 } else {
538e84ed 2825#ifdef DEBUGGING
8e11feef
RGS
2826 if (state>1)
2827 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2828#endif
8e11feef
RGS
2829 break;
2830 }
2831 }
2e64971a 2832 trie->prefixlen = (state-1);
3dab1dad 2833 if (str) {
8e11feef 2834 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2835 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2836 trie->startstate = state;
07be1b83
YO
2837 trie->minlen -= (state - 1);
2838 trie->maxlen -= (state - 1);
33809eae
JH
2839#ifdef DEBUGGING
2840 /* At least the UNICOS C compiler choked on this
2841 * being argument to DEBUG_r(), so let's just have
2842 * it right here. */
2843 if (
2844#ifdef PERL_EXT_RE_BUILD
2845 1
2846#else
2847 DEBUG_r_TEST
2848#endif
2849 ) {
2850 regnode *fix = convert;
2851 U32 word = trie->wordcount;
2852 mjd_nodelen++;
2853 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2854 while( ++fix < n ) {
2855 Set_Node_Offset_Length(fix, 0, 0);
2856 }
2857 while (word--) {
2858 SV ** const tmp = av_fetch( trie_words, word, 0 );
2859 if (tmp) {
2860 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2861 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2862 else
2863 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2864 }
2865 }
2866 }
2867#endif
8e11feef
RGS
2868 if (trie->maxlen) {
2869 convert = n;
2870 } else {
3dab1dad 2871 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2872 DEBUG_r(optimize= n);
3dab1dad
YO
2873 }
2874 }
2875 }
538e84ed
KW
2876 if (!jumper)
2877 jumper = last;
3dab1dad 2878 if ( trie->maxlen ) {
8e11feef
RGS
2879 NEXT_OFF( convert ) = (U16)(tail - convert);
2880 ARG_SET( convert, data_slot );
538e84ed
KW
2881 /* Store the offset to the first unabsorbed branch in
2882 jump[0], which is otherwise unused by the jump logic.
786e8c11 2883 We use this when dumping a trie and during optimisation. */
538e84ed 2884 if (trie->jump)
7f69552c 2885 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 2886
6c48061a
YO
2887 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2888 * and there is a bitmap
2889 * and the first "jump target" node we found leaves enough room
2890 * then convert the TRIE node into a TRIEC node, with the bitmap
2891 * embedded inline in the opcode - this is hypothetically faster.
2892 */
2893 if ( !trie->states[trie->startstate].wordnum
2894 && trie->bitmap
2895 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2896 {
2897 OP( convert ) = TRIEC;
2898 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2899 PerlMemShared_free(trie->bitmap);
786e8c11 2900 trie->bitmap= NULL;
538e84ed 2901 } else
786e8c11 2902 OP( convert ) = TRIE;
a3621e74 2903
3dab1dad
YO
2904 /* store the type in the flags */
2905 convert->flags = nodetype;
a5ca303d 2906 DEBUG_r({
538e84ed
KW
2907 optimize = convert
2908 + NODE_STEP_REGNODE
a5ca303d
YO
2909 + regarglen[ OP( convert ) ];
2910 });
538e84ed 2911 /* XXX We really should free up the resource in trie now,
a5ca303d 2912 as we won't use them - (which resources?) dmq */
3dab1dad 2913 }
a3621e74 2914 /* needed for dumping*/
e62cc96a 2915 DEBUG_r(if (optimize) {
07be1b83 2916 regnode *opt = convert;
bcdf7404 2917
e62cc96a 2918 while ( ++opt < optimize) {
07be1b83
YO
2919 Set_Node_Offset_Length(opt,0,0);
2920 }
538e84ed
KW
2921 /*
2922 Try to clean up some of the debris left after the
786e8c11 2923 optimisation.
a3621e74 2924 */
786e8c11 2925 while( optimize < jumper ) {
07be1b83 2926 mjd_nodelen += Node_Length((optimize));
a3621e74 2927 OP( optimize ) = OPTIMIZED;
07be1b83 2928 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2929 optimize++;
2930 }
07be1b83 2931 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2932 });
2933 } /* end node insert */
2e64971a
DM
2934
2935 /* Finish populating the prev field of the wordinfo array. Walk back
2936 * from each accept state until we find another accept state, and if
2937 * so, point the first word's .prev field at the second word. If the
2938 * second already has a .prev field set, stop now. This will be the
2939 * case either if we've already processed that word's accept state,
3b753521
FN
2940 * or that state had multiple words, and the overspill words were
2941 * already linked up earlier.
2e64971a
DM
2942 */
2943 {
2944 U16 word;
2945 U32 state;
2946 U16 prev;
2947
2948 for (word=1; word <= trie->wordcount; word++) {
2949 prev = 0;
2950 if (trie->wordinfo[word].prev)
2951 continue;
2952 state = trie->wordinfo[word].accept;
2953 while (state) {
2954 state = prev_states[state];
2955 if (!state)
2956 break;
2957 prev = trie->states[state].wordnum;
2958 if (prev)
2959 break;
2960 }
2961 trie->wordinfo[word].prev = prev;
2962 }
2963 Safefree(prev_states);
2964 }
2965
2966
2967 /* and now dump out the compressed format */
2968 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2969
55eed653 2970 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2971#ifdef DEBUGGING
2972 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2973 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2974#else
03e70be4 2975 SvREFCNT_dec_NN(revcharmap);
07be1b83 2976#endif
538e84ed
KW
2977 return trie->jump
2978 ? MADE_JUMP_TRIE
2979 : trie->startstate>1
2980 ? MADE_EXACT_TRIE
786e8c11
YO
2981 : MADE_TRIE;
2982}
2983
2984STATIC void
2985S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2986{
b423522f
KW
2987/* The Trie is constructed and compressed now so we can build a fail array if
2988 * it's needed
786e8c11 2989
b423522f
KW
2990 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2991 3.32 in the
2992 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2993 Ullman 1985/88
786e8c11
YO
2994 ISBN 0-201-10088-6
2995
b423522f
KW
2996 We find the fail state for each state in the trie, this state is the longest
2997 proper suffix of the current state's 'word' that is also a proper prefix of
2998 another word in our trie. State 1 represents the word '' and is thus the
2999 default fail state. This allows the DFA not to have to restart after its
3000 tried and failed a word at a given point, it simply continues as though it
3001 had been matching the other word in the first place.
786e8c11
YO
3002 Consider
3003 'abcdgu'=~/abcdefg|cdgu/
b423522f
KW
3004 When we get to 'd' we are still matching the first word, we would encounter
3005 'g' which would fail, which would bring us to the state representing 'd' in
3006 the second word where we would try 'g' and succeed, proceeding to match
3007 'cdgu'.
786e8c11
YO
3008 */
3009 /* add a fail transition */
3251b653
NC
3010 const U32 trie_offset = ARG(source);
3011 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
3012 U32 *q;
3013 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 3014 const U32 numstates = trie->statecount;
786e8c11
YO
3015 const U32 ubound = trie->lasttrans + ucharcount;
3016 U32 q_read = 0;
3017 U32 q_write = 0;
3018 U32 charid;
3019 U32 base = trie->states[ 1 ].trans.base;
3020 U32 *fail;
3021 reg_ac_data *aho;
cf78de0b 3022 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
786e8c11 3023 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
3024
3025 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
3026#ifndef DEBUGGING
3027 PERL_UNUSED_ARG(depth);
3028#endif
3029
3030
3031 ARG_SET( stclass, data_slot );
c944940b 3032 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 3033 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 3034 aho->trie=trie_offset;
446bd890
NC
3035 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3036 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 3037 Newxz( q, numstates, U32);
c944940b 3038 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
3039 aho->refcount = 1;
3040 fail = aho->fail;
3041 /* initialize fail[0..1] to be 1 so that we always have
3042 a valid final fail state */
3043 fail[ 0 ] = fail[ 1 ] = 1;
3044
3045 for ( charid = 0; charid < ucharcount ; charid++ ) {
3046 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3047 if ( newstate ) {
3048 q[ q_write ] = newstate;
3049 /* set to point at the root */
3050 fail[ q[ q_write++ ] ]=1;
3051 }
3052 }
3053 while ( q_read < q_write) {
3054 const U32 cur = q[ q_read++ % numstates ];
3055 base = trie->states[ cur ].trans.base;
3056
3057 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3058 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3059 if (ch_state) {
3060 U32 fail_state = cur;
3061 U32 fail_base;
3062 do {
3063 fail_state = fail[ fail_state ];
3064 fail_base = aho->states[ fail_state ].trans.base;
3065 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3066
3067 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3068 fail[ ch_state ] = fail_state;
3069 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3070 {
3071 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3072 }
3073 q[ q_write++ % numstates] = ch_state;
3074 }
3075 }
3076 }
3077 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3078 when we fail in state 1, this allows us to use the
3079 charclass scan to find a valid start char. This is based on the principle
3080 that theres a good chance the string being searched contains lots of stuff
3081 that cant be a start char.
3082 */
3083 fail[ 0 ] = fail[ 1 ] = 0;
3084 DEBUG_TRIE_COMPILE_r({
6d99fb9b 3085 PerlIO_printf(Perl_debug_log,
538e84ed 3086 "%*sStclass Failtable (%"UVuf" states): 0",
6d99fb9b 3087 (int)(depth * 2), "", (UV)numstates
1e2e3d02 3088 );
786e8c11
YO
3089 for( q_read=1; q_read<numstates; q_read++ ) {
3090 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3091 }
3092 PerlIO_printf(Perl_debug_log, "\n");
3093 });
3094 Safefree(q);
e384d5c1 3095 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
a3621e74
YO
3096}
3097
786e8c11 3098
07be1b83 3099#define DEBUG_PEEP(str,scan,depth) \
b515a41d 3100 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
3101 SV * const mysv=sv_newmortal(); \
3102 regnode *Next = regnext(scan); \
2395827c 3103 regprop(RExC_rx, mysv, scan, NULL); \
7f69552c 3104 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
3105 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3106 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 3107 }});
07be1b83 3108
1de06328 3109
bb914485 3110/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
3111 * one. The regop may be changed if the node(s) contain certain sequences that
3112 * require special handling. The joining is only done if:
bb914485
KW
3113 * 1) there is room in the current conglomerated node to entirely contain the
3114 * next one.
3115 * 2) they are the exact same node type
3116 *
87b8b349 3117 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
3118 * these get optimized out
3119 *
0a982f06
KW
3120 * If a node is to match under /i (folded), the number of characters it matches
3121 * can be different than its character length if it contains a multi-character
31f05a37
KW
3122 * fold. *min_subtract is set to the total delta number of characters of the
3123 * input nodes.
bb914485 3124 *
fb29cc72 3125 * And *unfolded_multi_char is set to indicate whether or not the node contains
31f05a37
KW
3126 * an unfolded multi-char fold. This happens when whether the fold is valid or
3127 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3128 * SMALL LETTER SHARP S, as only if the target string being matched against
3129 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3130 * folding rules depend on the locale in force at runtime. (Multi-char folds
3131 * whose components are all above the Latin1 range are not run-time locale
3132 * dependent, and have already been folded by the time this function is
3133 * called.)
f758bddf 3134 *
bb914485 3135 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
3136 * multi-character fold sequences. It's been wrong in Perl for a very long
3137 * time. There are three code points in Unicode whose multi-character folds
3138 * were long ago discovered to mess things up. The previous designs for
3139 * dealing with these involved assigning a special node for them. This
538e84ed 3140 * approach doesn't always work, as evidenced by this example:
a0c4c608 3141 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
538e84ed 3142 * Both sides fold to "sss", but if the pattern is parsed to create a node that
0a982f06 3143 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
3144 * successful match would have to cross the node's boundary. The new approach
3145 * that hopefully generally solves the problem generates an EXACTFU_SS node
538e84ed 3146 * that is "sss" in this case.
bb914485 3147 *
0a982f06 3148 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
3149 * just these three. Now the code is general, for all such cases. The
3150 * approach taken is:
0a982f06 3151 * 1) This routine examines each EXACTFish node that could contain multi-
538e84ed
KW
3152 * character folded sequences. Since a single character can fold into
3153 * such a sequence, the minimum match length for this node is less than
3154 * the number of characters in the node. This routine returns in
31f05a37
KW
3155 * *min_subtract how many characters to subtract from the the actual
3156 * length of the string to get a real minimum match length; it is 0 if
3157 * there are no multi-char foldeds. This delta is used by the caller to
3158 * adjust the min length of the match, and the delta between min and max,
3159 * so that the optimizer doesn't reject these possibilities based on size
3160 * constraints.
cb117658 3161 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
3162 * is used for an EXACTFU node that contains at least one "ss" sequence in
3163 * it. For non-UTF-8 patterns and strings, this is the only case where
3164 * there is a possible fold length change. That means that a regular
3165 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3166 * with length changes, and so can be processed faster. regexec.c takes
3167 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
31f05a37
KW
3168 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3169 * known until runtime). This saves effort in regex matching. However,
3170 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3171 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3172 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3173 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
0a982f06
KW
3174 * possibilities for the non-UTF8 patterns are quite simple, except for
3175 * the sharp s. All the ones that don't involve a UTF-8 target string are
3176 * members of a fold-pair, and arrays are set up for all of them so that
3177 * the other member of the pair can be found quickly. Code elsewhere in
3178 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3179 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3180 * described in the next item.
31f05a37
KW
3181 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3182 * validity of the fold won't be known until runtime, and so must remain
3183 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3184 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3185 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3186 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3187 * The reason this is a problem is that the optimizer part of regexec.c
3188 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3189 * that a character in the pattern corresponds to at most a single
3190 * character in the target string. (And I do mean character, and not byte
3191 * here, unlike other parts of the documentation that have never been
3192 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3193 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3194 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3195 * nodes, violate the assumption, and they are the only instances where it
3196 * is violated. I'm reluctant to try to change the assumption, as the
3197 * code involved is impenetrable to me (khw), so instead the code here
3198 * punts. This routine examines EXACTFL nodes, and (when the pattern
3199 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3200 * boolean indicating whether or not the node contains such a fold. When
3201 * it is true, the caller sets a flag that later causes the optimizer in
3202 * this file to not set values for the floating and fixed string lengths,
3203 * and thus avoids the optimizer code in regexec.c that makes the invalid
1ca267a5 3204 * assumption. Thus, there is no optimization based on string lengths for
31f05a37
KW
3205 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3206 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3207 * assumption is wrong only in these cases is that all other non-UTF-8
3208 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3209 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3210 * EXACTF nodes because we don't know at compile time if it actually
3211 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3212 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3213 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3214 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3215 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3216 * string would require the pattern to be forced into UTF-8, the overhead
3217 * of which we want to avoid. Similarly the unfolded multi-char folds in
3218 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3219 * locale.)
098b07d5
KW
3220 *
3221 * Similarly, the code that generates tries doesn't currently handle
3222 * not-already-folded multi-char folds, and it looks like a pain to change
3223 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3224 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3225 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3226 * using /iaa matching will be doing so almost entirely with ASCII
3227 * strings, so this should rarely be encountered in practice */
1de06328 3228
fb29cc72 3229#define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
07be1b83 3230 if (PL_regkind[OP(scan)] == EXACT) \
fb29cc72 3231 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
07be1b83 3232
be8e71aa 3233STATIC U32
538e84ed 3234S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
fb29cc72 3235 UV *min_subtract, bool *unfolded_multi_char,
538e84ed
KW
3236 U32 flags,regnode *val, U32 depth)
3237{
07be1b83
YO
3238 /* Merge several consecutive EXACTish nodes into one. */
3239 regnode *n = regnext(scan);
3240 U32 stringok = 1;
3241 regnode *next = scan + NODE_SZ_STR(scan);
3242 U32 merged = 0;
3243 U32 stopnow = 0;
3244#ifdef DEBUGGING
3245 regnode *stop = scan;
72f13be8 3246 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 3247#else
d47053eb
RGS
3248 PERL_UNUSED_ARG(depth);
3249#endif
7918f24d
NC
3250
3251 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 3252#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
3253 PERL_UNUSED_ARG(flags);
3254 PERL_UNUSED_ARG(val);
07be1b83 3255#endif
07be1b83 3256 DEBUG_PEEP("join",scan,depth);
bb914485 3257
3f410cf6
KW
3258 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3259 * EXACT ones that are mergeable to the current one. */
3260 while (n
3261 && (PL_regkind[OP(n)] == NOTHING
3262 || (stringok && OP(n) == OP(scan)))
07be1b83 3263 && NEXT_OFF(n)
3f410cf6
KW
3264 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3265 {
538e84ed 3266
07be1b83
YO
3267 if (OP(n) == TAIL || n > next)
3268 stringok = 0;
3269 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
3270 DEBUG_PEEP("skip:",n,depth);
3271 NEXT_OFF(scan) += NEXT_OFF(n);
3272 next = n + NODE_STEP_REGNODE;
3273#ifdef DEBUGGING
3274 if (stringok)
3275 stop = n;
3276#endif
3277 n = regnext(n);
3278 }
3279 else if (stringok) {
786e8c11 3280 const unsigned int oldl = STR_LEN(scan);
07be1b83 3281 regnode * const nnext = regnext(n);
b2230d39 3282
baa60164
KW
3283 /* XXX I (khw) kind of doubt that this works on platforms (should
3284 * Perl ever run on one) where U8_MAX is above 255 because of lots
3285 * of other assumptions */
79a81a6e 3286 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
3287 if (oldl + STR_LEN(n) > U8_MAX)
3288 break;
538e84ed 3289
07be1b83 3290 DEBUG_PEEP("merg",n,depth);
07be1b83 3291 merged++;
b2230d39 3292
07be1b83
YO
3293 NEXT_OFF(scan) += NEXT_OFF(n);
3294 STR_LEN(scan) += STR_LEN(n);
3295 next = n + NODE_SZ_STR(n);
3296 /* Now we can overwrite *n : */
3297 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3298#ifdef DEBUGGING
3299 stop = next - 1;
3300#endif
3301 n = nnext;
3302 if (stopnow) break;
3303 }
3304
d47053eb
RGS
3305#ifdef EXPERIMENTAL_INPLACESCAN
3306 if (flags && !NEXT_OFF(n)) {
3307 DEBUG_PEEP("atch", val, depth);
3308 if (reg_off_by_arg[OP(n)]) {
3309 ARG_SET(n, val - n);
3310 }
3311 else {
3312 NEXT_OFF(n) = val - n;
3313 }
3314 stopnow = 1;
3315 }
07be1b83
YO
3316#endif
3317 }
2c2b7f86 3318
9d071ca8 3319 *min_subtract = 0;
fb29cc72 3320 *unfolded_multi_char = FALSE;
f646642f 3321
3f410cf6
KW
3322 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3323 * can now analyze for sequences of problematic code points. (Prior to
3324 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
3325 * hence missed). The sequences only happen in folding, hence for any
3326 * non-EXACT EXACTish node */
86d6fcad 3327 if (OP(scan) != EXACT) {
31f05a37
KW
3328 U8* s0 = (U8*) STRING(scan);
3329 U8* s = s0;
3330 U8* s_end = s0 + STR_LEN(scan);
3331
3332 int total_count_delta = 0; /* Total delta number of characters that
3333 multi-char folds expand to */
f758bddf
KW
3334
3335 /* One pass is made over the node's string looking for all the
baa60164 3336 * possibilities. To avoid some tests in the loop, there are two main
f758bddf
KW
3337 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3338 * non-UTF-8 */
3339 if (UTF) {
31f05a37
KW
3340 U8* folded = NULL;
3341
3342 if (OP(scan) == EXACTFL) {
3343 U8 *d;
3344
3345 /* An EXACTFL node would already have been changed to another
3346 * node type unless there is at least one character in it that
3347 * is problematic; likely a character whose fold definition
3348 * won't be known until runtime, and so has yet to be folded.
3349 * For all but the UTF-8 locale, folds are 1-1 in length, but
3350 * to handle the UTF-8 case, we need to create a temporary
3351 * folded copy using UTF-8 locale rules in order to analyze it.
3352 * This is because our macros that look to see if a sequence is
3353 * a multi-char fold assume everything is folded (otherwise the
3354 * tests in those macros would be too complicated and slow).
3355 * Note that here, the non-problematic folds will have already
3356 * been done, so we can just copy such characters. We actually
3357 * don't completely fold the EXACTFL string. We skip the
3358 * unfolded multi-char folds, as that would just create work
3359 * below to figure out the size they already are */
3360
3361 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3362 d = folded;
3363 while (s < s_end) {
3364 STRLEN s_len = UTF8SKIP(s);
3365 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3366 Copy(s, d, s_len, U8);
3367 d += s_len;
3368 }
3369 else if (is_FOLDS_TO_MULTI_utf8(s)) {
fb29cc72 3370 *unfolded_multi_char = TRUE;
31f05a37
KW
3371 Copy(s, d, s_len, U8);
3372 d += s_len;
3373 }
3374 else if (isASCII(*s)) {
3375 *(d++) = toFOLD(*s);
3376 }
3377 else {
3378 STRLEN len;
3379 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3380 d += len;
3381 }
3382 s += s_len;
3383 }
3384
3385 /* Point the remainder of the routine to look at our temporary
3386 * folded copy */
3387 s = folded;
3388 s_end = d;
3389 } /* End of creating folded copy of EXACTFL string */
86d6fcad 3390
0a982f06
KW
3391 /* Examine the string for a multi-character fold sequence. UTF-8
3392 * patterns have all characters pre-folded by the time this code is
3393 * executed */
3394 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3395 length sequence we are looking for is 2 */
86d6fcad 3396 {
538e84ed 3397 int count = 0; /* How many characters in a multi-char fold */
251b239f 3398 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
0a982f06
KW
3399 if (! len) { /* Not a multi-char fold: get next char */
3400 s += UTF8SKIP(s);
3401 continue;
3402 }
bb914485 3403
31f05a37
KW
3404 /* Nodes with 'ss' require special handling, except for
3405 * EXACTFA-ish for which there is no multi-char fold to this */
0a982f06 3406 if (len == 2 && *s == 's' && *(s+1) == 's'
098b07d5
KW
3407 && OP(scan) != EXACTFA
3408 && OP(scan) != EXACTFA_NO_TRIE)
0a982f06
KW
3409 {
3410 count = 2;
31f05a37
KW
3411 if (OP(scan) != EXACTFL) {
3412 OP(scan) = EXACTFU_SS;
3413 }
0a982f06
KW
3414 s += 2;
3415 }
0a982f06 3416 else { /* Here is a generic multi-char fold. */
31f05a37
KW
3417 U8* multi_end = s + len;
3418
3419 /* Count how many characters in it. In the case of /aa, no
3420 * folds which contain ASCII code points are allowed, so
3421 * check for those, and skip if found. */
3422 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
0a982f06
KW
3423 count = utf8_length(s, multi_end);
3424 s = multi_end;
3425 }
3426 else {
3427 while (s < multi_end) {
3428 if (isASCII(*s)) {
3429 s++;
3430 goto next_iteration;
3431 }
3432 else {
3433 s += UTF8SKIP(s);
3434 }
3435 count++;
3436 }
3437 }
3438 }
f758bddf 3439
0a982f06
KW
3440 /* The delta is how long the sequence is minus 1 (1 is how long
3441 * the character that folds to the sequence is) */
31f05a37
KW
3442 total_count_delta += count - 1;
3443 next_iteration: ;
bb914485 3444 }
31f05a37
KW
3445
3446 /* We created a temporary folded copy of the string in EXACTFL
3447 * nodes. Therefore we need to be sure it doesn't go below zero,
3448 * as the real string could be shorter */
3449 if (OP(scan) == EXACTFL) {
3450 int total_chars = utf8_length((U8*) STRING(scan),
3451 (U8*) STRING(scan) + STR_LEN(scan));
3452 if (total_count_delta > total_chars) {
3453 total_count_delta = total_chars;
3454 }
3455 }
3456
3457 *min_subtract += total_count_delta;
3458 Safefree(folded);
bb914485 3459 }
1ca267a5
KW
3460 else if (OP(scan) == EXACTFA) {
3461
3462 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3463 * fold to the ASCII range (and there are no existing ones in the
3464 * upper latin1 range). But, as outlined in the comments preceding
098b07d5
KW
3465 * this function, we need to flag any occurrences of the sharp s.
3466 * This character forbids trie formation (because of added
3467 * complexity) */
1ca267a5
KW
3468 while (s < s_end) {
3469 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
098b07d5 3470 OP(scan) = EXACTFA_NO_TRIE;
fb29cc72 3471 *unfolded_multi_char = TRUE;
1ca267a5
KW
3472 break;
3473 }
3474 s++;
3475 continue;
3476 }
3477 }
31f05a37
KW
3478 else {
3479
3480 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3481 * folds that are all Latin1. As explained in the comments
3482 * preceding this function, we look also for the sharp s in EXA