This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regexec.c: in debug show whether TRIE nodes have a jump table
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
afda64e8 84EXTERN_C const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
f7e03a10 89#include "dquote_inline.h"
b992490d 90#include "invlist_inline.h"
1b0f46bf 91#include "unicode_constants.h"
04e98a4d 92
538e84ed
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
95#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 97#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 98#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 99
a687059c
LW
100#ifndef STATIC
101#define STATIC static
102#endif
103
3f910e62
YO
104/* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
107struct scan_frame;
108typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
113 U32 is_top_frame; /* what flags do we use? */
114
115 struct scan_frame *this_prev_frame; /* this previous frame */
116 struct scan_frame *prev_frame; /* previous frame */
117 struct scan_frame *next_frame; /* next frame */
118} scan_frame;
119
f4ae5a27
KW
120/* Certain characters are output as a sequence with the first being a
121 * backslash. */
122#define isBACKSLASHED_PUNCT(c) \
123 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
124
125
09b2b2e6 126struct RExC_state_t {
514a91f1
DM
127 U32 flags; /* RXf_* are we folding, multilining? */
128 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 129 char *precomp; /* uncompiled string. */
711b303b 130 char *precomp_end; /* pointer to end of uncompiled string. */
288b8c02 131 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 132 regexp *rx; /* perl core regexp structure */
538e84ed
KW
133 regexp_internal *rxi; /* internal data for regexp object
134 pprivate field */
fac92740 135 char *start; /* Start of input for compile */
830247a4
IZ
136 char *end; /* End of input for compile */
137 char *parse; /* Input-scan pointer. */
285b5ca0
KW
138 char *adjusted_start; /* 'start', adjusted. See code use */
139 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
ea3daa5d 140 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 141 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
142 regnode *emit_bound; /* First regnode outside of the
143 allocated space */
f7c7e32a
DM
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
9a81a976
KW
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
149 scratch in pass1 */
830247a4
IZ
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
152 U32 seen;
ea3daa5d 153 SSize_t size; /* Code size. */
538e84ed
KW
154 I32 npar; /* Capture buffer count, (OPEN) plus
155 one. ("par" 0 is the whole
156 pattern)*/
157 I32 nestroot; /* root parens we are in - used by
158 accept */
830247a4
IZ
159 I32 extralen;
160 I32 seen_zerolen;
40d049e4
YO
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
d5a00e4a 163 regnode *end_op; /* END node in program */
02daf0ab
YO
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
e40e74fe
KW
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
170 utf8 */
81714fb9 171 HV *paren_names; /* Paren names */
538e84ed 172
40d049e4 173 regnode **recurse; /* Recurse regops */
d5a00e4a 174 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 176 through */
09a65838 177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 178 I32 in_lookbehind;
4624b182 179 I32 contains_locale;
cfafade5 180 I32 contains_i;
bb3f3ed2 181 I32 override_recoding;
b6d67071
KW
182#ifdef EBCDIC
183 I32 recode_x_to_native;
184#endif
9d53c457 185 I32 in_multi_char_class;
3d2bd50a 186 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 187 within pattern */
b1603ef8
DM
188 int num_code_blocks; /* size of code_blocks[] */
189 int code_index; /* next code_blocks[] slot */
ee273784 190 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
191 scan_frame *frame_head;
192 scan_frame *frame_last;
193 U32 frame_count;
7eec73eb 194 AV *warn_text;
dc6d7f5c 195#ifdef ADD_TO_REGEXEC
830247a4
IZ
196 char *starttry; /* -Dr: where regtry was called. */
197#define RExC_starttry (pRExC_state->starttry)
198#endif
d24ca0c5 199 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 200#ifdef DEBUGGING
be8e71aa 201 const char *lastparse;
3dab1dad 202 I32 lastnum;
1f1031fe 203 AV *paren_name_list; /* idx -> name */
d9a72fcc 204 U32 study_chunk_recursed_count;
c9f0d54c
YO
205 SV *mysv1;
206 SV *mysv2;
3dab1dad
YO
207#define RExC_lastparse (pRExC_state->lastparse)
208#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 209#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 210#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
211#define RExC_mysv (pRExC_state->mysv1)
212#define RExC_mysv1 (pRExC_state->mysv1)
213#define RExC_mysv2 (pRExC_state->mysv2)
214
3dab1dad 215#endif
512c0f5a 216 bool seen_unfolded_sharp_s;
911bd04e 217 bool strict;
da7cf1cc 218 bool study_started;
09b2b2e6 219};
830247a4 220
e2509266 221#define RExC_flags (pRExC_state->flags)
514a91f1 222#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 223#define RExC_precomp (pRExC_state->precomp)
285b5ca0
KW
224#define RExC_precomp_adj (pRExC_state->precomp_adj)
225#define RExC_adjusted_start (pRExC_state->adjusted_start)
711b303b 226#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 227#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 228#define RExC_rx (pRExC_state->rx)
f8fc2ecf 229#define RExC_rxi (pRExC_state->rxi)
fac92740 230#define RExC_start (pRExC_state->start)
830247a4
IZ
231#define RExC_end (pRExC_state->end)
232#define RExC_parse (pRExC_state->parse)
233#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
234
235/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
236 * EXACTF node, hence was parsed under /di rules. If later in the parse,
237 * something forces the pattern into using /ui rules, the sharp s should be
238 * folded into the sequence 'ss', which takes up more space than previously
239 * calculated. This means that the sizing pass needs to be restarted. (The
240 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
241 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
242 * so there is no need to resize [perl #125990]. */
243#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
244
7122b237 245#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
246#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
247 others */
7122b237 248#endif
830247a4 249#define RExC_emit (pRExC_state->emit)
f7c7e32a 250#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 251#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 252#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
253#define RExC_sawback (pRExC_state->sawback)
254#define RExC_seen (pRExC_state->seen)
255#define RExC_size (pRExC_state->size)
ee273784 256#define RExC_maxlen (pRExC_state->maxlen)
830247a4 257#define RExC_npar (pRExC_state->npar)
e2e6a0f1 258#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
259#define RExC_extralen (pRExC_state->extralen)
260#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 261#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 262#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 263#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
264#define RExC_open_parens (pRExC_state->open_parens)
265#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 266#define RExC_end_op (pRExC_state->end_op)
81714fb9 267#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
268#define RExC_recurse (pRExC_state->recurse)
269#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 270#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
271#define RExC_study_chunk_recursed_bytes \
272 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 273#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 274#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 275#define RExC_contains_i (pRExC_state->contains_i)
9d53c457 276#define RExC_override_recoding (pRExC_state->override_recoding)
b6d67071
KW
277#ifdef EBCDIC
278# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
279#endif
9d53c457 280#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
281#define RExC_frame_head (pRExC_state->frame_head)
282#define RExC_frame_last (pRExC_state->frame_last)
283#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 284#define RExC_strict (pRExC_state->strict)
da7cf1cc 285#define RExC_study_started (pRExC_state->study_started)
7eec73eb 286#define RExC_warn_text (pRExC_state->warn_text)
830247a4 287
99807a43
HS
288/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
289 * a flag to disable back-off on the fixed/floating substrings - if it's
290 * a high complexity pattern we assume the benefit of avoiding a full match
291 * is worth the cost of checking for the substrings even if they rarely help.
292 */
293#define RExC_naughty (pRExC_state->naughty)
294#define TOO_NAUGHTY (10)
295#define MARK_NAUGHTY(add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += (add)
298#define MARK_NAUGHTY_EXP(exp, add) \
299 if (RExC_naughty < TOO_NAUGHTY) \
300 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 301
a687059c
LW
302#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
303#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 304 ((*s) == '{' && regcurly(s)))
a687059c
LW
305
306/*
307 * Flags to be passed up and down.
308 */
a687059c 309#define WORST 0 /* Worst case. */
a3b492c3 310#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 311
e64f369d 312/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
313 * character. (There needs to be a case: in the switch statement in regexec.c
314 * for any node marked SIMPLE.) Note that this is not the same thing as
315 * REGNODE_SIMPLE */
fda99bee 316#define SIMPLE 0x02
e64f369d 317#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
318#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
319#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
320#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
321#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
322 calcuate sizes as UTF-8 */
a687059c 323
3dab1dad
YO
324#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
325
07be1b83
YO
326/* whether trie related optimizations are enabled */
327#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
328#define TRIE_STUDY_OPT
786e8c11 329#define FULL_TRIE_STUDY
07be1b83
YO
330#define TRIE_STCLASS
331#endif
1de06328
YO
332
333
40d049e4
YO
334
335#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
336#define PBITVAL(paren) (1 << ((paren) & 7))
337#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
338#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
339#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
340
82a6ada4 341#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 342 if (!UTF) { \
82a6ada4
KW
343 assert(PASS1); \
344 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
345 return NULL; \
346 } \
82a6ada4 347 } STMT_END
40d049e4 348
512c0f5a
KW
349/* Change from /d into /u rules, and restart the parse if we've already seen
350 * something whose size would increase as a result, by setting *flagp and
351 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
352 * we've change to /u during the parse. */
353#define REQUIRE_UNI_RULES(flagp, restart_retval) \
354 STMT_START { \
355 if (DEPENDS_SEMANTICS) { \
356 assert(PASS1); \
357 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
358 RExC_uni_semantics = 1; \
359 if (RExC_seen_unfolded_sharp_s) { \
360 *flagp |= RESTART_PASS1; \
361 return restart_retval; \
362 } \
363 } \
364 } STMT_END
365
f19b1a63
KW
366/* This converts the named class defined in regcomp.h to its equivalent class
367 * number defined in handy.h. */
368#define namedclass_to_classnum(class) ((int) ((class) / 2))
369#define classnum_to_namedclass(classnum) ((classnum) * 2)
370
de92f5e6
KW
371#define _invlist_union_complement_2nd(a, b, output) \
372 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
373#define _invlist_intersection_complement_2nd(a, b, output) \
374 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
375
1de06328
YO
376/* About scan_data_t.
377
378 During optimisation we recurse through the regexp program performing
379 various inplace (keyhole style) optimisations. In addition study_chunk
380 and scan_commit populate this data structure with information about
538e84ed 381 what strings MUST appear in the pattern. We look for the longest
3b753521 382 string that must appear at a fixed location, and we look for the
1de06328
YO
383 longest string that may appear at a floating location. So for instance
384 in the pattern:
538e84ed 385
1de06328 386 /FOO[xX]A.*B[xX]BAR/
538e84ed 387
1de06328
YO
388 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
389 strings (because they follow a .* construct). study_chunk will identify
390 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 391
1de06328 392 The strings can be composites, for instance
538e84ed 393
1de06328 394 /(f)(o)(o)/
538e84ed 395
1de06328 396 will result in a composite fixed substring 'foo'.
538e84ed 397
1de06328 398 For each string some basic information is maintained:
538e84ed 399
1de06328
YO
400 - offset or min_offset
401 This is the position the string must appear at, or not before.
402 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
403 characters must match before the string we are searching for.
404 Likewise when combined with minlenp and the length of the string it
538e84ed 405 tells us how many characters must appear after the string we have
1de06328 406 found.
538e84ed 407
1de06328
YO
408 - max_offset
409 Only used for floating strings. This is the rightmost point that
ea3daa5d 410 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 411 string can occur infinitely far to the right.
538e84ed 412
1de06328 413 - minlenp
2d608413
KW
414 A pointer to the minimum number of characters of the pattern that the
415 string was found inside. This is important as in the case of positive
538e84ed 416 lookahead or positive lookbehind we can have multiple patterns
1de06328 417 involved. Consider
538e84ed 418
1de06328 419 /(?=FOO).*F/
538e84ed 420
1de06328
YO
421 The minimum length of the pattern overall is 3, the minimum length
422 of the lookahead part is 3, but the minimum length of the part that
538e84ed 423 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 424 minimum length for the F is 1. This is important as the minimum length
538e84ed 425 is used to determine offsets in front of and behind the string being
1de06328 426 looked for. Since strings can be composites this is the length of the
486ec47a 427 pattern at the time it was committed with a scan_commit. Note that
1de06328 428 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 429 are not known until the full pattern has been compiled, thus the
1de06328 430 pointer to the value.
538e84ed 431
1de06328 432 - lookbehind
538e84ed 433
1de06328 434 In the case of lookbehind the string being searched for can be
538e84ed 435 offset past the start point of the final matching string.
1de06328
YO
436 If this value was just blithely removed from the min_offset it would
437 invalidate some of the calculations for how many chars must match
438 before or after (as they are derived from min_offset and minlen and
538e84ed 439 the length of the string being searched for).
1de06328
YO
440 When the final pattern is compiled and the data is moved from the
441 scan_data_t structure into the regexp structure the information
538e84ed
KW
442 about lookbehind is factored in, with the information that would
443 have been lost precalculated in the end_shift field for the
1de06328
YO
444 associated string.
445
446 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 447 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
448
449*/
2c2d71f5
JH
450
451typedef struct scan_data_t {
1de06328
YO
452 /*I32 len_min; unused */
453 /*I32 len_delta; unused */
49f55535 454 SSize_t pos_min;
ea3daa5d 455 SSize_t pos_delta;
2c2d71f5 456 SV *last_found;
ea3daa5d 457 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 458 SSize_t last_start_min;
ea3daa5d 459 SSize_t last_start_max;
1de06328
YO
460 SV **longest; /* Either &l_fixed, or &l_float. */
461 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 462 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 463 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
464 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
465 SV *longest_float; /* longest floating string found in pattern */
49f55535 466 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
467 SSize_t offset_float_max; /* latest point in string it can appear */
468 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 469 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
470 I32 flags;
471 I32 whilem_c;
49f55535 472 SSize_t *last_closep;
b8f7bb16 473 regnode_ssc *start_class;
2c2d71f5
JH
474} scan_data_t;
475
a687059c 476/*
e50aee73 477 * Forward declarations for pregcomp()'s friends.
a687059c 478 */
a0d0e21e 479
27da23d5 480static const scan_data_t zero_scan_data =
1de06328 481 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
482
483#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
484#define SF_BEFORE_SEOL 0x0001
485#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
486#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
487#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
488
44e3dfd2
BF
489#define SF_FIX_SHIFT_EOL (+2)
490#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
491
492#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
493#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
494
495#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
496#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
497#define SF_IS_INF 0x0040
498#define SF_HAS_PAR 0x0080
499#define SF_IN_PAR 0x0100
500#define SF_HAS_EVAL 0x0200
bc604ad8
KW
501
502
503/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
504 * longest substring in the pattern. When it is not set the optimiser keeps
505 * track of position, but does not keep track of the actual strings seen,
506 *
507 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
508 * /foo/i will not.
509 *
510 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
511 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
512 * turned off because of the alternation (BRANCH). */
07be1b83 513#define SCF_DO_SUBSTR 0x0400
bc604ad8 514
653099ff
GS
515#define SCF_DO_STCLASS_AND 0x0800
516#define SCF_DO_STCLASS_OR 0x1000
517#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 518#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 519
786e8c11 520#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 521#define SCF_SEEN_ACCEPT 0x8000
688e0391 522#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
523#define SCF_IN_DEFINE 0x20000
524
525
526
07be1b83 527
43fead97 528#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
529
530/* The enums for all these are ordered so things work out correctly */
a62b1201 531#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
532#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
533 == REGEX_DEPENDS_CHARSET)
00b27cfc 534#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
535#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
536 >= REGEX_UNICODE_CHARSET)
537#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
538 == REGEX_ASCII_RESTRICTED_CHARSET)
539#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
540 >= REGEX_ASCII_RESTRICTED_CHARSET)
541#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
542 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 543
43fead97 544#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 545
f2c2a6ab
KW
546/* For programs that want to be strictly Unicode compatible by dying if any
547 * attempt is made to match a non-Unicode code point against a Unicode
548 * property. */
549#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
550
93733859 551#define OOB_NAMEDCLASS -1
b8c5462f 552
8e661ac5
KW
553/* There is no code point that is out-of-bounds, so this is problematic. But
554 * its only current use is to initialize a variable that is always set before
555 * looked at. */
556#define OOB_UNICODE 0xDEADBEEF
557
a0ed51b3
LW
558#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
559#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
560
8615cb43 561
b45f050a
JF
562/* length of regex to show in messages that don't mark a position within */
563#define RegexLengthToShowInErrorMessages 127
564
565/*
566 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
567 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
568 * op/pragma/warn/regcomp.
569 */
7253e4e3
RK
570#define MARKER1 "<-- HERE" /* marker as it appears in the description */
571#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 572
538e84ed
KW
573#define REPORT_LOCATION " in regex; marked by " MARKER1 \
574 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 575
285b5ca0
KW
576/* The code in this file in places uses one level of recursion with parsing
577 * rebased to an alternate string constructed by us in memory. This can take
578 * the form of something that is completely different from the input, or
579 * something that uses the input as part of the alternate. In the first case,
580 * there should be no possibility of an error, as we are in complete control of
581 * the alternate string. But in the second case we don't control the input
582 * portion, so there may be errors in that. Here's an example:
583 * /[abc\x{DF}def]/ui
584 * is handled specially because \x{df} folds to a sequence of more than one
585 * character, 'ss'. What is done is to create and parse an alternate string,
586 * which looks like this:
587 * /(?:\x{DF}|[abc\x{DF}def])/ui
588 * where it uses the input unchanged in the middle of something it constructs,
589 * which is a branch for the DF outside the character class, and clustering
590 * parens around the whole thing. (It knows enough to skip the DF inside the
591 * class while in this substitute parse.) 'abc' and 'def' may have errors that
592 * need to be reported. The general situation looks like this:
593 *
594 * sI tI xI eI
595 * Input: ----------------------------------------------------
596 * Constructed: ---------------------------------------------------
597 * sC tC xC eC EC
598 *
599 * The input string sI..eI is the input pattern. The string sC..EC is the
600 * constructed substitute parse string. The portions sC..tC and eC..EC are
601 * constructed by us. The portion tC..eC is an exact duplicate of the input
602 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
603 * while parsing, we find an error at xC. We want to display a message showing
604 * the real input string. Thus we need to find the point xI in it which
605 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
606 * been constructed by us, and so shouldn't have errors. We get:
607 *
608 * xI = sI + (tI - sI) + (xC - tC)
609 *
610 * and, the offset into sI is:
611 *
612 * (xI - sI) = (tI - sI) + (xC - tC)
613 *
614 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
615 * and we save tC as RExC_adjusted_start.
903c858a
KW
616 *
617 * During normal processing of the input pattern, everything points to that,
618 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
285b5ca0
KW
619 */
620
621#define tI_sI RExC_precomp_adj
622#define tC RExC_adjusted_start
623#define sC RExC_precomp
624#define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
625#define xI(xC) (sC + xI_offset(xC))
626#define eC RExC_precomp_end
627
628#define REPORT_LOCATION_ARGS(xC) \
629 UTF8fARG(UTF, \
630 (xI(xC) > eC) /* Don't run off end */ \
631 ? eC - sC /* Length before the <--HERE */ \
632 : xI_offset(xC), \
633 sC), /* The input pattern printed up to the <--HERE */ \
634 UTF8fARG(UTF, \
635 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
636 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
c1d900c3 637
8a6d8ec6
HS
638/* Used to point after bad bytes for an error message, but avoid skipping
639 * past a nul byte. */
640#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
641
b45f050a
JF
642/*
643 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
644 * arg. Show regex, up to a maximum length. If it's too long, chop and add
645 * "...".
646 */
58e23c8d 647#define _FAIL(code) STMT_START { \
bfed75c6 648 const char *ellipses = ""; \
711b303b 649 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
650 \
651 if (!SIZE_ONLY) \
a5e7bc51 652 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
653 if (len > RegexLengthToShowInErrorMessages) { \
654 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
655 len = RegexLengthToShowInErrorMessages - 10; \
656 ellipses = "..."; \
657 } \
58e23c8d 658 code; \
ccb2c380 659} STMT_END
8615cb43 660
58e23c8d 661#define FAIL(msg) _FAIL( \
c1d900c3
BF
662 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
663 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
664
665#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
666 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
667 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 668
b45f050a 669/*
b45f050a
JF
670 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
671 */
ccb2c380 672#define Simple_vFAIL(m) STMT_START { \
ccb2c380 673 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 674 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 675} STMT_END
b45f050a
JF
676
677/*
678 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
679 */
ccb2c380
MP
680#define vFAIL(m) STMT_START { \
681 if (!SIZE_ONLY) \
a5e7bc51 682 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
683 Simple_vFAIL(m); \
684} STMT_END
b45f050a
JF
685
686/*
687 * Like Simple_vFAIL(), but accepts two arguments.
688 */
ccb2c380 689#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
690 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
691 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 692} STMT_END
b45f050a
JF
693
694/*
695 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
696 */
ccb2c380
MP
697#define vFAIL2(m,a1) STMT_START { \
698 if (!SIZE_ONLY) \
a5e7bc51 699 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
700 Simple_vFAIL2(m, a1); \
701} STMT_END
b45f050a
JF
702
703
704/*
705 * Like Simple_vFAIL(), but accepts three arguments.
706 */
ccb2c380 707#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 708 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 709 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 710} STMT_END
b45f050a
JF
711
712/*
713 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
714 */
ccb2c380
MP
715#define vFAIL3(m,a1,a2) STMT_START { \
716 if (!SIZE_ONLY) \
a5e7bc51 717 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
718 Simple_vFAIL3(m, a1, a2); \
719} STMT_END
b45f050a
JF
720
721/*
722 * Like Simple_vFAIL(), but accepts four arguments.
723 */
ccb2c380 724#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
725 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
726 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 727} STMT_END
b45f050a 728
95db3ffa
KW
729#define vFAIL4(m,a1,a2,a3) STMT_START { \
730 if (!SIZE_ONLY) \
731 SAVEFREESV(RExC_rx_sv); \
732 Simple_vFAIL4(m, a1, a2, a3); \
733} STMT_END
734
946095af 735/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
736#define vFAIL2utf8f(m, a1) STMT_START { \
737 if (!SIZE_ONLY) \
738 SAVEFREESV(RExC_rx_sv); \
739 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
740 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
741} STMT_END
742
3ba22297 743#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
744 if (!SIZE_ONLY) \
745 SAVEFREESV(RExC_rx_sv); \
746 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 747 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
748} STMT_END
749
499333dc
KW
750/* These have asserts in them because of [perl #122671] Many warnings in
751 * regcomp.c can occur twice. If they get output in pass1 and later in that
752 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
753 * would get output again. So they should be output in pass2, and these
754 * asserts make sure new warnings follow that paradigm. */
946095af 755
5e0a247b
KW
756/* m is not necessarily a "literal string", in this macro */
757#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
758 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
759 "%s" REPORT_LOCATION, \
760 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
761} STMT_END
762
668c081a 763#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
764 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
765 m REPORT_LOCATION, \
766 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
767} STMT_END
768
b927b7e9 769#define vWARN(loc, m) STMT_START { \
d528642a
KW
770 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
771 m REPORT_LOCATION, \
772 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
773} STMT_END
774
0d6106aa 775#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
776 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
777 m REPORT_LOCATION, \
778 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
779} STMT_END
780
147508a2 781#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
783 m REPORT_LOCATION, \
784 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
785} STMT_END
786
d528642a
KW
787#define ckWARNregdep(loc,m) STMT_START { \
788 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
789 WARN_REGEXP), \
790 m REPORT_LOCATION, \
791 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
792} STMT_END
793
d528642a
KW
794#define ckWARN2reg_d(loc,m, a1) STMT_START { \
795 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
796 m REPORT_LOCATION, \
797 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
798} STMT_END
799
d528642a
KW
800#define ckWARN2reg(loc, m, a1) STMT_START { \
801 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
802 m REPORT_LOCATION, \
803 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
804} STMT_END
805
d528642a
KW
806#define vWARN3(loc, m, a1, a2) STMT_START { \
807 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
808 m REPORT_LOCATION, \
809 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
810} STMT_END
811
d528642a
KW
812#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
813 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
814 m REPORT_LOCATION, \
815 a1, a2, \
816 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
817} STMT_END
818
ccb2c380 819#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
820 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
821 m REPORT_LOCATION, \
822 a1, a2, a3, \
823 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
824} STMT_END
825
668c081a 826#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
827 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
828 m REPORT_LOCATION, \
829 a1, a2, a3, \
830 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
831} STMT_END
832
ccb2c380 833#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
834 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
835 m REPORT_LOCATION, \
836 a1, a2, a3, a4, \
837 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 838} STMT_END
9d1d55b5 839
538e84ed 840/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
841 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
842 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
843 * Element 0 holds the number n.
07be1b83 844 * Position is 1 indexed.
fac92740 845 */
7122b237
YO
846#ifndef RE_TRACK_PATTERN_OFFSETS
847#define Set_Node_Offset_To_R(node,byte)
848#define Set_Node_Offset(node,byte)
849#define Set_Cur_Node_Offset
850#define Set_Node_Length_To_R(node,len)
851#define Set_Node_Length(node,len)
6a86c6ad 852#define Set_Node_Cur_Length(node,start)
538e84ed
KW
853#define Node_Offset(n)
854#define Node_Length(n)
7122b237
YO
855#define Set_Node_Offset_Length(node,offset,len)
856#define ProgLen(ri) ri->u.proglen
857#define SetProgLen(ri,x) ri->u.proglen = x
858#else
859#define ProgLen(ri) ri->u.offsets[0]
860#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
861#define Set_Node_Offset_To_R(node,byte) STMT_START { \
862 if (! SIZE_ONLY) { \
863 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 864 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 865 if((node) < 0) { \
538e84ed
KW
866 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
867 (int)(node)); \
ccb2c380
MP
868 } else { \
869 RExC_offsets[2*(node)-1] = (byte); \
870 } \
871 } \
872} STMT_END
873
874#define Set_Node_Offset(node,byte) \
875 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
876#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
877
878#define Set_Node_Length_To_R(node,len) STMT_START { \
879 if (! SIZE_ONLY) { \
880 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 881 __LINE__, (int)(node), (int)(len))); \
ccb2c380 882 if((node) < 0) { \
538e84ed
KW
883 Perl_croak(aTHX_ "value of node is %d in Length macro", \
884 (int)(node)); \
ccb2c380
MP
885 } else { \
886 RExC_offsets[2*(node)] = (len); \
887 } \
888 } \
889} STMT_END
890
891#define Set_Node_Length(node,len) \
892 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
893#define Set_Node_Cur_Length(node, start) \
894 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
895
896/* Get offsets and lengths */
897#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
898#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
899
07be1b83
YO
900#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
901 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
902 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
903} STMT_END
7122b237 904#endif
07be1b83
YO
905
906#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
907#define EXPERIMENTAL_INPLACESCAN
f427392e 908#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 909
cb41e5d6
YO
910#ifdef DEBUGGING
911int
6ad9a8ab 912Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
913{
914 va_list ap;
915 int result;
916 PerlIO *f= Perl_debug_log;
917 PERL_ARGS_ASSERT_RE_PRINTF;
918 va_start(ap, fmt);
919 result = PerlIO_vprintf(f, fmt, ap);
920 va_end(ap);
921 return result;
922}
923
924int
7b031478 925Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
926{
927 va_list ap;
928 int result;
929 PerlIO *f= Perl_debug_log;
930 PERL_ARGS_ASSERT_RE_INDENTF;
931 va_start(ap, depth);
daeb874b 932 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
933 result = PerlIO_vprintf(f, fmt, ap);
934 va_end(ap);
935 return result;
936}
937#endif /* DEBUGGING */
938
7b031478 939#define DEBUG_RExC_seen() \
538e84ed 940 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 941 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 942 \
e384d5c1 943 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 944 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 945 \
e384d5c1 946 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 947 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 948 \
e384d5c1 949 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 950 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 951 \
e384d5c1 952 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 953 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 954 \
7b031478 955 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 956 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 957 \
e384d5c1 958 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 959 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 960 \
e384d5c1 961 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 962 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 963 \
e384d5c1 964 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 965 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 966 \
e384d5c1 967 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 968 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 969 \
7b031478 970 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 971 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 972 \
6ad9a8ab 973 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
974 });
975
fdfb4f21 976#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 977 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21
YO
978
979#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
980 if ( ( flags ) ) { \
6ad9a8ab 981 Perl_re_printf( aTHX_ "%s", open_str); \
fdfb4f21
YO
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
993 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
994 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
995 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
996 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
6ad9a8ab 997 Perl_re_printf( aTHX_ "%s", close_str); \
fdfb4f21
YO
998 }
999
1000
304ee84b
YO
1001#define DEBUG_STUDYDATA(str,data,depth) \
1002DEBUG_OPTIMISE_MORE_r(if(data){ \
6ad9a8ab 1003 Perl_re_indentf( aTHX_ "" str "Pos:%"IVdf"/%"IVdf \
fdfb4f21 1004 " Flags: 0x%"UVXf, \
7b031478 1005 depth, \
1de06328
YO
1006 (IV)((data)->pos_min), \
1007 (IV)((data)->pos_delta), \
fdfb4f21
YO
1008 (UV)((data)->flags) \
1009 ); \
1010 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
6ad9a8ab 1011 Perl_re_printf( aTHX_ \
fdfb4f21 1012 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328 1013 (IV)((data)->whilem_c), \
304ee84b
YO
1014 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1015 is_inf ? "INF " : "" \
1de06328
YO
1016 ); \
1017 if ((data)->last_found) \
6ad9a8ab 1018 Perl_re_printf( aTHX_ \
1de06328
YO
1019 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
1020 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
1021 SvPVX_const((data)->last_found), \
1022 (IV)((data)->last_end), \
1023 (IV)((data)->last_start_min), \
1024 (IV)((data)->last_start_max), \
1025 ((data)->longest && \
1026 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1027 SvPVX_const((data)->longest_fixed), \
1028 (IV)((data)->offset_fixed), \
1029 ((data)->longest && \
1030 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1031 SvPVX_const((data)->longest_float), \
1032 (IV)((data)->offset_float_min), \
1033 (IV)((data)->offset_float_max) \
1034 ); \
6ad9a8ab 1035 Perl_re_printf( aTHX_ "\n"); \
1de06328
YO
1036});
1037
cb41e5d6 1038
c6871b76
KW
1039/* =========================================================
1040 * BEGIN edit_distance stuff.
1041 *
1042 * This calculates how many single character changes of any type are needed to
1043 * transform a string into another one. It is taken from version 3.1 of
1044 *
1045 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1046 */
1047
1048/* Our unsorted dictionary linked list. */
1049/* Note we use UVs, not chars. */
1050
1051struct dictionary{
1052 UV key;
1053 UV value;
1054 struct dictionary* next;
1055};
1056typedef struct dictionary item;
1057
1058
1059PERL_STATIC_INLINE item*
1060push(UV key,item* curr)
1061{
1062 item* head;
1063 Newxz(head, 1, item);
1064 head->key = key;
1065 head->value = 0;
1066 head->next = curr;
1067 return head;
1068}
1069
1070
1071PERL_STATIC_INLINE item*
1072find(item* head, UV key)
1073{
1074 item* iterator = head;
1075 while (iterator){
1076 if (iterator->key == key){
1077 return iterator;
1078 }
1079 iterator = iterator->next;
1080 }
1081
1082 return NULL;
1083}
1084
1085PERL_STATIC_INLINE item*
1086uniquePush(item* head,UV key)
1087{
1088 item* iterator = head;
1089
1090 while (iterator){
1091 if (iterator->key == key) {
1092 return head;
1093 }
1094 iterator = iterator->next;
1095 }
1096
1097 return push(key,head);
1098}
1099
1100PERL_STATIC_INLINE void
1101dict_free(item* head)
1102{
1103 item* iterator = head;
1104
1105 while (iterator) {
1106 item* temp = iterator;
1107 iterator = iterator->next;
1108 Safefree(temp);
1109 }
1110
1111 head = NULL;
1112}
1113
1114/* End of Dictionary Stuff */
1115
1116/* All calculations/work are done here */
1117STATIC int
1118S_edit_distance(const UV* src,
1119 const UV* tgt,
1120 const STRLEN x, /* length of src[] */
1121 const STRLEN y, /* length of tgt[] */
1122 const SSize_t maxDistance
1123)
1124{
1125 item *head = NULL;
1126 UV swapCount,swapScore,targetCharCount,i,j;
1127 UV *scores;
1128 UV score_ceil = x + y;
1129
1130 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1131
1132 /* intialize matrix start values */
1133 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1134 scores[0] = score_ceil;
1135 scores[1 * (y + 2) + 0] = score_ceil;
1136 scores[0 * (y + 2) + 1] = score_ceil;
1137 scores[1 * (y + 2) + 1] = 0;
1138 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1139
1140 /* work loops */
1141 /* i = src index */
1142 /* j = tgt index */
1143 for (i=1;i<=x;i++) {
1144 if (i < x)
1145 head = uniquePush(head,src[i]);
1146 scores[(i+1) * (y + 2) + 1] = i;
1147 scores[(i+1) * (y + 2) + 0] = score_ceil;
1148 swapCount = 0;
1149
1150 for (j=1;j<=y;j++) {
1151 if (i == 1) {
1152 if(j < y)
1153 head = uniquePush(head,tgt[j]);
1154 scores[1 * (y + 2) + (j + 1)] = j;
1155 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1156 }
1157
1158 targetCharCount = find(head,tgt[j-1])->value;
1159 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1160
1161 if (src[i-1] != tgt[j-1]){
1162 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1163 }
1164 else {
1165 swapCount = j;
1166 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1167 }
1168 }
1169
1170 find(head,src[i-1])->value = i;
1171 }
1172
1173 {
1174 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1175 dict_free(head);
1176 Safefree(scores);
1177 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1178 }
1179}
1180
1181/* END of edit_distance() stuff
1182 * ========================================================= */
1183
8e35b056
KW
1184/* is c a control character for which we have a mnemonic? */
1185#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1186
549b4e78
KW
1187STATIC const char *
1188S_cntrl_to_mnemonic(const U8 c)
1189{
1190 /* Returns the mnemonic string that represents character 'c', if one
1191 * exists; NULL otherwise. The only ones that exist for the purposes of
1192 * this routine are a few control characters */
1193
1194 switch (c) {
1195 case '\a': return "\\a";
1196 case '\b': return "\\b";
1197 case ESC_NATIVE: return "\\e";
1198 case '\f': return "\\f";
1199 case '\n': return "\\n";
1200 case '\r': return "\\r";
1201 case '\t': return "\\t";
1202 }
1203
1204 return NULL;
1205}
1206
653099ff 1207/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 1208 Update the longest found anchored substring and the longest found
653099ff
GS
1209 floating substrings if needed. */
1210
4327152a 1211STATIC void
ea3daa5d
FC
1212S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1213 SSize_t *minlenp, int is_inf)
c277df42 1214{
e1ec3a88
AL
1215 const STRLEN l = CHR_SVLEN(data->last_found);
1216 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 1217 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1218
7918f24d
NC
1219 PERL_ARGS_ASSERT_SCAN_COMMIT;
1220
c277df42 1221 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 1222 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
1223 if (*data->longest == data->longest_fixed) {
1224 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1225 if (data->flags & SF_BEFORE_EOL)
b81d288d 1226 data->flags
c277df42
IZ
1227 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1228 else
1229 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 1230 data->minlen_fixed=minlenp;
1de06328 1231 data->lookbehind_fixed=0;
a0ed51b3 1232 }
304ee84b 1233 else { /* *data->longest == data->longest_float */
c277df42 1234 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 1235 data->offset_float_max = (l
646e8787
DM
1236 ? data->last_start_max
1237 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1238 ? SSize_t_MAX
1239 : data->pos_min + data->pos_delta));
1240 if (is_inf
1241 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1242 data->offset_float_max = SSize_t_MAX;
c277df42 1243 if (data->flags & SF_BEFORE_EOL)
b81d288d 1244 data->flags
c277df42
IZ
1245 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1246 else
1247 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
1248 data->minlen_float=minlenp;
1249 data->lookbehind_float=0;
c277df42
IZ
1250 }
1251 }
1252 SvCUR_set(data->last_found, 0);
0eda9292 1253 {
a28509cc 1254 SV * const sv = data->last_found;
097eb12c
AL
1255 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1256 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1257 if (mg)
1258 mg->mg_len = 0;
1259 }
0eda9292 1260 }
c277df42
IZ
1261 data->last_end = -1;
1262 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1263 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1264}
1265
cdd87c1d
KW
1266/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1267 * list that describes which code points it matches */
1268
653099ff 1269STATIC void
3420edd7 1270S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1271{
cdd87c1d
KW
1272 /* Set the SSC 'ssc' to match an empty string or any code point */
1273
557bd3fb 1274 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1275
71068078 1276 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1277
0854ea0b
KW
1278 /* mortalize so won't leak */
1279 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1280 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1281}
1282
653099ff 1283STATIC int
dc3bf405 1284S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1285{
c144baaa
KW
1286 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1287 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1288 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1289 * in any way, so there's no point in using it */
cdd87c1d
KW
1290
1291 UV start, end;
1292 bool ret;
653099ff 1293
557bd3fb 1294 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1295
71068078 1296 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1297
93e92956 1298 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1299 return FALSE;
1300 }
1301
1302 /* See if the list consists solely of the range 0 - Infinity */
1303 invlist_iterinit(ssc->invlist);
1304 ret = invlist_iternext(ssc->invlist, &start, &end)
1305 && start == 0
1306 && end == UV_MAX;
1307
1308 invlist_iterfinish(ssc->invlist);
1309
1310 if (ret) {
1311 return TRUE;
1312 }
1313
1314 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1315 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1316 int i;
1317 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1318 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1319 return TRUE;
1320 }
1321 }
1322 }
1323
1324 return FALSE;
653099ff
GS
1325}
1326
653099ff 1327STATIC void
cdd87c1d 1328S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1329{
cdd87c1d
KW
1330 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1331 * string, any code point, or any posix class under locale */
1332
557bd3fb 1333 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1334
557bd3fb 1335 Zero(ssc, 1, regnode_ssc);
71068078 1336 set_ANYOF_SYNTHETIC(ssc);
93e92956 1337 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1338 ssc_anything(ssc);
cdd87c1d 1339
2f306ab9
KW
1340 /* If any portion of the regex is to operate under locale rules that aren't
1341 * fully known at compile time, initialization includes it. The reason
1342 * this isn't done for all regexes is that the optimizer was written under
1343 * the assumption that locale was all-or-nothing. Given the complexity and
1344 * lack of documentation in the optimizer, and that there are inadequate
1345 * test cases for locale, many parts of it may not work properly, it is
1346 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1347 if (RExC_contains_locale) {
1348 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1349 }
1350 else {
1351 ANYOF_POSIXL_ZERO(ssc);
1352 }
653099ff
GS
1353}
1354
b423522f 1355STATIC int
dc3bf405
BF
1356S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1357 const regnode_ssc *ssc)
b423522f
KW
1358{
1359 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1360 * to the list of code points matched, and locale posix classes; hence does
1361 * not check its flags) */
1362
1363 UV start, end;
1364 bool ret;
1365
1366 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1367
71068078 1368 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1369
1370 invlist_iterinit(ssc->invlist);
1371 ret = invlist_iternext(ssc->invlist, &start, &end)
1372 && start == 0
1373 && end == UV_MAX;
1374
1375 invlist_iterfinish(ssc->invlist);
1376
1377 if (! ret) {
1378 return FALSE;
1379 }
1380
e0e1be5f 1381 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1382 return FALSE;
b423522f
KW
1383 }
1384
1385 return TRUE;
1386}
1387
1388STATIC SV*
1389S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1390 const regnode_charclass* const node)
b423522f
KW
1391{
1392 /* Returns a mortal inversion list defining which code points are matched
1393 * by 'node', which is of type ANYOF. Handles complementing the result if
1394 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1395 * returned list must, and will, contain every code point that is a
1396 * possibility. */
b423522f 1397
e2506fa7 1398 SV* invlist = NULL;
1ee208c4 1399 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1400 unsigned int i;
1401 const U32 n = ARG(node);
31f05a37 1402 bool new_node_has_latin1 = FALSE;
b423522f
KW
1403
1404 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1405
1406 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1407 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1408 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1409 AV * const av = MUTABLE_AV(SvRV(rv));
1410 SV **const ary = AvARRAY(av);
1411 assert(RExC_rxi->data->what[n] == 's');
1412
1413 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1414 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1415 }
1416 else if (ary[0] && ary[0] != &PL_sv_undef) {
1417
1418 /* Here, no compile-time swash, and there are things that won't be
1419 * known until runtime -- we have to assume it could be anything */
e2506fa7 1420 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1421 return _add_range_to_invlist(invlist, 0, UV_MAX);
1422 }
1ee208c4 1423 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1424
1425 /* Here no compile-time swash, and no run-time only data. Use the
1426 * node's inversion list */
1ee208c4
KW
1427 invlist = sv_2mortal(invlist_clone(ary[3]));
1428 }
1429
1430 /* Get the code points valid only under UTF-8 locales */
037715a6 1431 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1432 && ary[2] && ary[2] != &PL_sv_undef)
1433 {
1434 only_utf8_locale_invlist = ary[2];
b423522f
KW
1435 }
1436 }
1437
e2506fa7
KW
1438 if (! invlist) {
1439 invlist = sv_2mortal(_new_invlist(0));
1440 }
1441
dcb20b36
KW
1442 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1443 * code points, and an inversion list for the others, but if there are code
1444 * points that should match only conditionally on the target string being
1445 * UTF-8, those are placed in the inversion list, and not the bitmap.
1446 * Since there are circumstances under which they could match, they are
1447 * included in the SSC. But if the ANYOF node is to be inverted, we have
1448 * to exclude them here, so that when we invert below, the end result
1449 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1450 * have to do this here before we add the unconditionally matched code
1451 * points */
b423522f
KW
1452 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1453 _invlist_intersection_complement_2nd(invlist,
1454 PL_UpperLatin1,
1455 &invlist);
1456 }
1457
1458 /* Add in the points from the bit map */
dcb20b36 1459 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1460 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1461 unsigned int start = i++;
1462
1463 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1464 /* empty */
1465 }
1466 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1467 new_node_has_latin1 = TRUE;
b423522f
KW
1468 }
1469 }
1470
1471 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1472 * as well. But don't add them if inverting, as when that gets done below,
1473 * it would exclude all these characters, including the ones it shouldn't
1474 * that were added just above */
1475 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1476 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1477 {
b423522f
KW
1478 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1479 }
1480
1481 /* Similarly for these */
93e92956 1482 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1483 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1484 }
1485
1486 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1487 _invlist_invert(invlist);
1488 }
037715a6 1489 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1490
1491 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1492 * locale. We can skip this if there are no 0-255 at all. */
1493 _invlist_union(invlist, PL_Latin1, &invlist);
1494 }
1495
1ee208c4
KW
1496 /* Similarly add the UTF-8 locale possible matches. These have to be
1497 * deferred until after the non-UTF-8 locale ones are taken care of just
1498 * above, or it leads to wrong results under ANYOF_INVERT */
1499 if (only_utf8_locale_invlist) {
31f05a37 1500 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1501 only_utf8_locale_invlist,
31f05a37
KW
1502 ANYOF_FLAGS(node) & ANYOF_INVERT,
1503 &invlist);
1504 }
b423522f
KW
1505
1506 return invlist;
1507}
1508
1051e1c4 1509/* These two functions currently do the exact same thing */
557bd3fb 1510#define ssc_init_zero ssc_init
653099ff 1511
cdd87c1d
KW
1512#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1513#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1514
557bd3fb 1515/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1516 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1517 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1518
653099ff 1519STATIC void
b423522f 1520S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1521 const regnode_charclass *and_with)
653099ff 1522{
cdd87c1d
KW
1523 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1524 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1525
a0dd4231
KW
1526 SV* anded_cp_list;
1527 U8 anded_flags;
1e6ade67 1528
cdd87c1d 1529 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1530
71068078 1531 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1532
1533 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1534 * the code point inversion list and just the relevant flags */
71068078 1535 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1536 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1537 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1538
1539 /* XXX This is a kludge around what appears to be deficiencies in the
1540 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1541 * there are paths through the optimizer where it doesn't get weeded
1542 * out when it should. And if we don't make some extra provision for
1543 * it like the code just below, it doesn't get added when it should.
1544 * This solution is to add it only when AND'ing, which is here, and
1545 * only when what is being AND'ed is the pristine, original node
1546 * matching anything. Thus it is like adding it to ssc_anything() but
1547 * only when the result is to be AND'ed. Probably the same solution
1548 * could be adopted for the same problem we have with /l matching,
1549 * which is solved differently in S_ssc_init(), and that would lead to
1550 * fewer false positives than that solution has. But if this solution
1551 * creates bugs, the consequences are only that a warning isn't raised
1552 * that should be; while the consequences for having /l bugs is
1553 * incorrect matches */
7dcac5f6 1554 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1555 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1556 }
a0dd4231
KW
1557 }
1558 else {
5c0f85ef 1559 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1560 if (OP(and_with) == ANYOFD) {
1561 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1562 }
1563 else {
1564 anded_flags = ANYOF_FLAGS(and_with)
1565 &( ANYOF_COMMON_FLAGS
108316fb
KW
1566 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1567 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1568 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1569 anded_flags &=
1570 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1571 }
f240c685 1572 }
a0dd4231
KW
1573 }
1574
1575 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1576
1577 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1578 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1579 * 'and_with' may be inverted. When not inverted, we have the situation of
1580 * computing:
1581 * (C1 | P1) & (C2 | P2)
1582 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1583 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1584 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1585 * <= ((C1 & C2) | P1 | P2)
1586 * Alternatively, the last few steps could be:
1587 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1588 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1589 * <= (C1 | C2 | (P1 & P2))
1590 * We favor the second approach if either P1 or P2 is non-empty. This is
1591 * because these components are a barrier to doing optimizations, as what
1592 * they match cannot be known until the moment of matching as they are
1593 * dependent on the current locale, 'AND"ing them likely will reduce or
1594 * eliminate them.
1595 * But we can do better if we know that C1,P1 are in their initial state (a
1596 * frequent occurrence), each matching everything:
1597 * (<everything>) & (C2 | P2) = C2 | P2
1598 * Similarly, if C2,P2 are in their initial state (again a frequent
1599 * occurrence), the result is a no-op
1600 * (C1 | P1) & (<everything>) = C1 | P1
1601 *
1602 * Inverted, we have
1603 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1604 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1605 * <= (C1 & ~C2) | (P1 & ~P2)
1606 * */
1aa99e6b 1607
a0dd4231 1608 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1609 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1610 {
cdd87c1d 1611 unsigned int i;
8951c461 1612
cdd87c1d
KW
1613 ssc_intersection(ssc,
1614 anded_cp_list,
1615 FALSE /* Has already been inverted */
1616 );
c6b76537 1617
cdd87c1d
KW
1618 /* If either P1 or P2 is empty, the intersection will be also; can skip
1619 * the loop */
93e92956 1620 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1621 ANYOF_POSIXL_ZERO(ssc);
1622 }
e0e1be5f 1623 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1624
1625 /* Note that the Posix class component P from 'and_with' actually
1626 * looks like:
1627 * P = Pa | Pb | ... | Pn
1628 * where each component is one posix class, such as in [\w\s].
1629 * Thus
1630 * ~P = ~(Pa | Pb | ... | Pn)
1631 * = ~Pa & ~Pb & ... & ~Pn
1632 * <= ~Pa | ~Pb | ... | ~Pn
1633 * The last is something we can easily calculate, but unfortunately
1634 * is likely to have many false positives. We could do better
1635 * in some (but certainly not all) instances if two classes in
1636 * P have known relationships. For example
1637 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1638 * So
1639 * :lower: & :print: = :lower:
1640 * And similarly for classes that must be disjoint. For example,
1641 * since \s and \w can have no elements in common based on rules in
1642 * the POSIX standard,
1643 * \w & ^\S = nothing
1644 * Unfortunately, some vendor locales do not meet the Posix
1645 * standard, in particular almost everything by Microsoft.
1646 * The loop below just changes e.g., \w into \W and vice versa */
1647
1ee208c4 1648 regnode_charclass_posixl temp;
cdd87c1d
KW
1649 int add = 1; /* To calculate the index of the complement */
1650
1651 ANYOF_POSIXL_ZERO(&temp);
1652 for (i = 0; i < ANYOF_MAX; i++) {
1653 assert(i % 2 != 0
7dcac5f6
KW
1654 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1655 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1656
7dcac5f6 1657 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1658 ANYOF_POSIXL_SET(&temp, i + add);
1659 }
1660 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1661 }
1662 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1663
cdd87c1d
KW
1664 } /* else ssc already has no posixes */
1665 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1666 in its initial state */
71068078 1667 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1668 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1669 {
1670 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1671 * copy it over 'ssc' */
1672 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1673 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1674 StructCopy(and_with, ssc, regnode_ssc);
1675 }
1676 else {
1677 ssc->invlist = anded_cp_list;
1678 ANYOF_POSIXL_ZERO(ssc);
93e92956 1679 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1680 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1681 }
1682 }
1683 }
e0e1be5f 1684 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1685 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1686 {
1687 /* One or the other of P1, P2 is non-empty. */
93e92956 1688 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1689 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1690 }
cdd87c1d
KW
1691 ssc_union(ssc, anded_cp_list, FALSE);
1692 }
1693 else { /* P1 = P2 = empty */
1694 ssc_intersection(ssc, anded_cp_list, FALSE);
1695 }
137165a6 1696 }
653099ff
GS
1697}
1698
653099ff 1699STATIC void
cdd87c1d 1700S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1701 const regnode_charclass *or_with)
653099ff 1702{
cdd87c1d
KW
1703 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1704 * another SSC or a regular ANYOF class. Can create false positives if
1705 * 'or_with' is to be inverted. */
7918f24d 1706
a0dd4231
KW
1707 SV* ored_cp_list;
1708 U8 ored_flags;
c6b76537 1709
cdd87c1d 1710 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1711
71068078 1712 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1713
1714 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1715 * the code point inversion list and just the relevant flags */
71068078 1716 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1717 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1718 ored_flags = ANYOF_FLAGS(or_with);
1719 }
1720 else {
5c0f85ef 1721 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1722 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1723 if (OP(or_with) != ANYOFD) {
1724 ored_flags
1725 |= ANYOF_FLAGS(or_with)
108316fb
KW
1726 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1727 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1728 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1729 ored_flags |=
1730 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1731 }
f240c685 1732 }
a0dd4231
KW
1733 }
1734
1735 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1736
1737 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1738 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1739 * 'or_with' may be inverted. When not inverted, we have the simple
1740 * situation of computing:
1741 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1742 * If P1|P2 yields a situation with both a class and its complement are
1743 * set, like having both \w and \W, this matches all code points, and we
1744 * can delete these from the P component of the ssc going forward. XXX We
1745 * might be able to delete all the P components, but I (khw) am not certain
1746 * about this, and it is better to be safe.
1747 *
1748 * Inverted, we have
1749 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1750 * <= (C1 | P1) | ~C2
1751 * <= (C1 | ~C2) | P1
1752 * (which results in actually simpler code than the non-inverted case)
1753 * */
9826f543 1754
a0dd4231 1755 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1756 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1757 {
cdd87c1d 1758 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1759 } /* else Not inverted */
93e92956 1760 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1761 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1762 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1763 unsigned int i;
1764 for (i = 0; i < ANYOF_MAX; i += 2) {
1765 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1766 {
1767 ssc_match_all_cp(ssc);
1768 ANYOF_POSIXL_CLEAR(ssc, i);
1769 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1770 }
1771 }
1772 }
1aa99e6b 1773 }
cdd87c1d
KW
1774
1775 ssc_union(ssc,
1776 ored_cp_list,
1777 FALSE /* Already has been inverted */
1778 );
653099ff
GS
1779}
1780
b423522f
KW
1781PERL_STATIC_INLINE void
1782S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1783{
1784 PERL_ARGS_ASSERT_SSC_UNION;
1785
71068078 1786 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1787
1788 _invlist_union_maybe_complement_2nd(ssc->invlist,
1789 invlist,
1790 invert2nd,
1791 &ssc->invlist);
1792}
1793
1794PERL_STATIC_INLINE void
1795S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1796 SV* const invlist,
1797 const bool invert2nd)
1798{
1799 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1800
71068078 1801 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1802
1803 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1804 invlist,
1805 invert2nd,
1806 &ssc->invlist);
1807}
1808
1809PERL_STATIC_INLINE void
1810S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1811{
1812 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1813
71068078 1814 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1815
1816 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1817}
1818
1819PERL_STATIC_INLINE void
1820S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1821{
1822 /* AND just the single code point 'cp' into the SSC 'ssc' */
1823
1824 SV* cp_list = _new_invlist(2);
1825
1826 PERL_ARGS_ASSERT_SSC_CP_AND;
1827
71068078 1828 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1829
1830 cp_list = add_cp_to_invlist(cp_list, cp);
1831 ssc_intersection(ssc, cp_list,
1832 FALSE /* Not inverted */
1833 );
1834 SvREFCNT_dec_NN(cp_list);
1835}
1836
1837PERL_STATIC_INLINE void
dc3bf405 1838S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1839{
1840 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1841 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1842
71068078 1843 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1844
1845 ANYOF_POSIXL_ZERO(ssc);
1846 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1847}
1848
b35552de
KW
1849#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1850
1851STATIC bool
1852S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1853{
1854 /* The synthetic start class is used to hopefully quickly winnow down
1855 * places where a pattern could start a match in the target string. If it
1856 * doesn't really narrow things down that much, there isn't much point to
1857 * having the overhead of using it. This function uses some very crude
1858 * heuristics to decide if to use the ssc or not.
1859 *
1860 * It returns TRUE if 'ssc' rules out more than half what it considers to
1861 * be the "likely" possible matches, but of course it doesn't know what the
1862 * actual things being matched are going to be; these are only guesses
1863 *
1864 * For /l matches, it assumes that the only likely matches are going to be
1865 * in the 0-255 range, uniformly distributed, so half of that is 127
1866 * For /a and /d matches, it assumes that the likely matches will be just
1867 * the ASCII range, so half of that is 63
1868 * For /u and there isn't anything matching above the Latin1 range, it
1869 * assumes that that is the only range likely to be matched, and uses
1870 * half that as the cut-off: 127. If anything matches above Latin1,
1871 * it assumes that all of Unicode could match (uniformly), except for
1872 * non-Unicode code points and things in the General Category "Other"
1873 * (unassigned, private use, surrogates, controls and formats). This
1874 * is a much large number. */
1875
b35552de
KW
1876 U32 count = 0; /* Running total of number of code points matched by
1877 'ssc' */
1878 UV start, end; /* Start and end points of current range in inversion
1879 list */
72400949
KW
1880 const U32 max_code_points = (LOC)
1881 ? 256
1882 : (( ! UNI_SEMANTICS
1883 || invlist_highest(ssc->invlist) < 256)
1884 ? 128
1885 : NON_OTHER_COUNT);
1886 const U32 max_match = max_code_points / 2;
b35552de
KW
1887
1888 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1889
1890 invlist_iterinit(ssc->invlist);
1891 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1892 if (start >= max_code_points) {
1893 break;
b35552de 1894 }
72400949 1895 end = MIN(end, max_code_points - 1);
b35552de 1896 count += end - start + 1;
72400949 1897 if (count >= max_match) {
b35552de
KW
1898 invlist_iterfinish(ssc->invlist);
1899 return FALSE;
1900 }
1901 }
1902
1903 return TRUE;
1904}
1905
1906
b423522f
KW
1907STATIC void
1908S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1909{
1910 /* The inversion list in the SSC is marked mortal; now we need a more
1911 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1912 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1913 * map */
b423522f
KW
1914
1915 SV* invlist = invlist_clone(ssc->invlist);
1916
1917 PERL_ARGS_ASSERT_SSC_FINALIZE;
1918
71068078 1919 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1920
a0dd4231 1921 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1922 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1923 * by the time we reach here */
f240c685
KW
1924 assert(! (ANYOF_FLAGS(ssc)
1925 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
1926 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1927 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 1928
b423522f
KW
1929 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1930
1ee208c4
KW
1931 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1932 NULL, NULL, NULL, FALSE);
b423522f 1933
85c8e306
KW
1934 /* Make sure is clone-safe */
1935 ssc->invlist = NULL;
1936
e0e1be5f 1937 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1938 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1939 }
1462525b 1940
b2e90ddf
KW
1941 if (RExC_contains_locale) {
1942 OP(ssc) = ANYOFL;
1943 }
1944
1462525b 1945 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1946}
1947
a3621e74
YO
1948#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1949#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1950#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1951#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1952 ? (TRIE_LIST_CUR( idx ) - 1) \
1953 : 0 )
a3621e74 1954
3dab1dad
YO
1955
1956#ifdef DEBUGGING
07be1b83 1957/*
2b8b4781
NC
1958 dump_trie(trie,widecharmap,revcharmap)
1959 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1960 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1961
1962 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1963 The _interim_ variants are used for debugging the interim
1964 tables that are used to generate the final compressed
1965 representation which is what dump_trie expects.
1966
486ec47a 1967 Part of the reason for their existence is to provide a form
3dab1dad 1968 of documentation as to how the different representations function.
07be1b83
YO
1969
1970*/
3dab1dad
YO
1971
1972/*
3dab1dad
YO
1973 Dumps the final compressed table form of the trie to Perl_debug_log.
1974 Used for debugging make_trie().
1975*/
b9a59e08 1976
3dab1dad 1977STATIC void
2b8b4781
NC
1978S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1979 AV *revcharmap, U32 depth)
3dab1dad
YO
1980{
1981 U32 state;
ab3bbdeb 1982 SV *sv=sv_newmortal();
55eed653 1983 int colwidth= widecharmap ? 6 : 4;
2e64971a 1984 U16 word;
3dab1dad
YO
1985 GET_RE_DEBUG_FLAGS_DECL;
1986
7918f24d 1987 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1988
6ad9a8ab 1989 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 1990 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
1991
1992 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1993 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1994 if ( tmp ) {
6ad9a8ab 1995 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 1996 colwidth,
538e84ed 1997 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1998 PL_colors[0], PL_colors[1],
1999 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2000 PERL_PV_ESCAPE_FIRSTCHAR
2001 )
ab3bbdeb 2002 );
3dab1dad
YO
2003 }
2004 }
1e37780e
YO
2005 Perl_re_printf( aTHX_ "\n");
2006 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2007
2008 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2009 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2010 Perl_re_printf( aTHX_ "\n");
3dab1dad 2011
1e2e3d02 2012 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2013 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2014
1e37780e 2015 Perl_re_indentf( aTHX_ "#%4"UVXf"|", depth+1, (UV)state);
3dab1dad
YO
2016
2017 if ( trie->states[ state ].wordnum ) {
1e37780e 2018 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2019 } else {
6ad9a8ab 2020 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2021 }
2022
6ad9a8ab 2023 Perl_re_printf( aTHX_ " @%4"UVXf" ", (UV)base );
3dab1dad
YO
2024
2025 if ( base ) {
2026 U32 ofs = 0;
2027
2028 while( ( base + ofs < trie->uniquecharcount ) ||
2029 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2030 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2031 != state))
3dab1dad
YO
2032 ofs++;
2033
6ad9a8ab 2034 Perl_re_printf( aTHX_ "+%2"UVXf"[ ", (UV)ofs);
3dab1dad
YO
2035
2036 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2037 if ( ( base + ofs >= trie->uniquecharcount )
2038 && ( base + ofs - trie->uniquecharcount
2039 < trie->lasttrans )
2040 && trie->trans[ base + ofs
2041 - trie->uniquecharcount ].check == state )
3dab1dad 2042 {
1e37780e
YO
2043 Perl_re_printf( aTHX_ "%*"UVXf, colwidth,
2044 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2045 );
3dab1dad 2046 } else {
6ad9a8ab 2047 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2048 }
2049 }
2050
6ad9a8ab 2051 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2052
2053 }
6ad9a8ab 2054 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2055 }
6ad9a8ab 2056 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2057 depth);
2e64971a 2058 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2059 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2060 (int)word, (int)(trie->wordinfo[word].prev),
2061 (int)(trie->wordinfo[word].len));
2062 }
6ad9a8ab 2063 Perl_re_printf( aTHX_ "\n" );
538e84ed 2064}
3dab1dad 2065/*
3dab1dad 2066 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2067 List tries normally only are used for construction when the number of
3dab1dad
YO
2068 possible chars (trie->uniquecharcount) is very high.
2069 Used for debugging make_trie().
2070*/
2071STATIC void
55eed653 2072S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2073 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2074 U32 depth)
3dab1dad
YO
2075{
2076 U32 state;
ab3bbdeb 2077 SV *sv=sv_newmortal();
55eed653 2078 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2079 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2080
2081 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2082
3dab1dad 2083 /* print out the table precompression. */
6ad9a8ab 2084 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2085 depth+1 );
6ad9a8ab 2086 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2087 depth+1, "------:-----+-----------------\n" );
538e84ed 2088
3dab1dad
YO
2089 for( state=1 ; state < next_alloc ; state ++ ) {
2090 U16 charid;
538e84ed 2091
6ad9a8ab 2092 Perl_re_indentf( aTHX_ " %4"UVXf" :",
cb41e5d6 2093 depth+1, (UV)state );
3dab1dad 2094 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2095 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2096 } else {
6ad9a8ab 2097 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2098 trie->states[ state ].wordnum
2099 );
2100 }
2101 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2102 SV ** const tmp = av_fetch( revcharmap,
2103 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2104 if ( tmp ) {
6ad9a8ab 2105 Perl_re_printf( aTHX_ "%*s:%3X=%4"UVXf" | ",
ab3bbdeb 2106 colwidth,
538e84ed
KW
2107 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2108 colwidth,
2109 PL_colors[0], PL_colors[1],
2110 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2111 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2112 ) ,
1e2e3d02
YO
2113 TRIE_LIST_ITEM(state,charid).forid,
2114 (UV)TRIE_LIST_ITEM(state,charid).newstate
2115 );
538e84ed 2116 if (!(charid % 10))
6ad9a8ab 2117 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2118 (int)((depth * 2) + 14), "");
1e2e3d02 2119 }
ab3bbdeb 2120 }
6ad9a8ab 2121 Perl_re_printf( aTHX_ "\n");
3dab1dad 2122 }
538e84ed 2123}
3dab1dad
YO
2124
2125/*
3dab1dad 2126 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2127 This is the normal DFA style state transition table, with a few
2128 twists to facilitate compression later.
3dab1dad
YO
2129 Used for debugging make_trie().
2130*/
2131STATIC void
55eed653 2132S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2133 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2134 U32 depth)
3dab1dad
YO
2135{
2136 U32 state;
2137 U16 charid;
ab3bbdeb 2138 SV *sv=sv_newmortal();
55eed653 2139 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2140 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2141
2142 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2143
3dab1dad
YO
2144 /*
2145 print out the table precompression so that we can do a visual check
2146 that they are identical.
2147 */
538e84ed 2148
6ad9a8ab 2149 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2150
2151 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2152 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2153 if ( tmp ) {
6ad9a8ab 2154 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2155 colwidth,
538e84ed 2156 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2157 PL_colors[0], PL_colors[1],
2158 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2159 PERL_PV_ESCAPE_FIRSTCHAR
2160 )
ab3bbdeb 2161 );
3dab1dad
YO
2162 }
2163 }
2164
4aaafc03
YO
2165 Perl_re_printf( aTHX_ "\n");
2166 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2167
2168 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2169 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2170 }
2171
6ad9a8ab 2172 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2173
2174 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2175
6ad9a8ab 2176 Perl_re_indentf( aTHX_ "%4"UVXf" : ",
cb41e5d6 2177 depth+1,
3dab1dad
YO
2178 (UV)TRIE_NODENUM( state ) );
2179
2180 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2181 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2182 if (v)
6ad9a8ab 2183 Perl_re_printf( aTHX_ "%*"UVXf, colwidth, v );
ab3bbdeb 2184 else
6ad9a8ab 2185 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2186 }
2187 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
6ad9a8ab 2188 Perl_re_printf( aTHX_ " (%4"UVXf")\n",
538e84ed 2189 (UV)trie->trans[ state ].check );
3dab1dad 2190 } else {
6ad9a8ab 2191 Perl_re_printf( aTHX_ " (%4"UVXf") W%4X\n",
538e84ed 2192 (UV)trie->trans[ state ].check,
3dab1dad
YO
2193 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2194 }
2195 }
07be1b83 2196}
3dab1dad
YO
2197
2198#endif
2199
2e64971a 2200
786e8c11
YO
2201/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2202 startbranch: the first branch in the whole branch sequence
2203 first : start branch of sequence of branch-exact nodes.
2204 May be the same as startbranch
2205 last : Thing following the last branch.
2206 May be the same as tail.
2207 tail : item following the branch sequence
2208 count : words in the sequence
a4525e78 2209 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2210 depth : indent depth
3dab1dad 2211
786e8c11 2212Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2213
786e8c11
YO
2214A trie is an N'ary tree where the branches are determined by digital
2215decomposition of the key. IE, at the root node you look up the 1st character and
2216follow that branch repeat until you find the end of the branches. Nodes can be
2217marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2218
786e8c11 2219 /he|she|his|hers/
72f13be8 2220
786e8c11
YO
2221would convert into the following structure. Numbers represent states, letters
2222following numbers represent valid transitions on the letter from that state, if
2223the number is in square brackets it represents an accepting state, otherwise it
2224will be in parenthesis.
07be1b83 2225
786e8c11
YO
2226 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2227 | |
2228 | (2)
2229 | |
2230 (1) +-i->(6)-+-s->[7]
2231 |
2232 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2233
786e8c11
YO
2234 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2235
2236This shows that when matching against the string 'hers' we will begin at state 1
2237read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2238then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2239is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2240single traverse. We store a mapping from accepting to state to which word was
2241matched, and then when we have multiple possibilities we try to complete the
b8fda935 2242rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2243
2244The only prior NFA like behaviour that would be changed by the TRIE support is
2245the silent ignoring of duplicate alternations which are of the form:
2246
2247 / (DUPE|DUPE) X? (?{ ... }) Y /x
2248
4b714af6 2249Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2250and without the optimisation. With the optimisations dupes will be silently
486ec47a 2251ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2252the following demonstrates:
2253
2254 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2255
2256which prints out 'word' three times, but
2257
2258 'words'=~/(word|word|word)(?{ print $1 })S/
2259
2260which doesnt print it out at all. This is due to other optimisations kicking in.
2261
2262Example of what happens on a structural level:
2263
486ec47a 2264The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2265
2266 1: CURLYM[1] {1,32767}(18)
2267 5: BRANCH(8)
2268 6: EXACT <ac>(16)
2269 8: BRANCH(11)
2270 9: EXACT <ad>(16)
2271 11: BRANCH(14)
2272 12: EXACT <ab>(16)
2273 16: SUCCEED(0)
2274 17: NOTHING(18)
2275 18: END(0)
2276
2277This would be optimizable with startbranch=5, first=5, last=16, tail=16
2278and should turn into:
2279
2280 1: CURLYM[1] {1,32767}(18)
2281 5: TRIE(16)
2282 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2283 <ac>
2284 <ad>
2285 <ab>
2286 16: SUCCEED(0)
2287 17: NOTHING(18)
2288 18: END(0)
2289
2290Cases where tail != last would be like /(?foo|bar)baz/:
2291
2292 1: BRANCH(4)
2293 2: EXACT <foo>(8)
2294 4: BRANCH(7)
2295 5: EXACT <bar>(8)
2296 7: TAIL(8)
2297 8: EXACT <baz>(10)
2298 10: END(0)
2299
2300which would be optimizable with startbranch=1, first=1, last=7, tail=8
2301and would end up looking like:
2302
2303 1: TRIE(8)
2304 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2305 <foo>
2306 <bar>
2307 7: TAIL(8)
2308 8: EXACT <baz>(10)
2309 10: END(0)
2310
c80e42f3 2311 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2312
2313is the recommended Unicode-aware way of saying
2314
2315 *(d++) = uv;
2316*/
2317
fab2782b 2318#define TRIE_STORE_REVCHAR(val) \
786e8c11 2319 STMT_START { \
73031816 2320 if (UTF) { \
668fcfea 2321 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2322 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2323 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2324 SvCUR_set(zlopp, kapow - flrbbbbb); \
2325 SvPOK_on(zlopp); \
2326 SvUTF8_on(zlopp); \
2327 av_push(revcharmap, zlopp); \
2328 } else { \
fab2782b 2329 char ooooff = (char)val; \
73031816
NC
2330 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2331 } \
2332 } STMT_END
786e8c11 2333
914a25d5
KW
2334/* This gets the next character from the input, folding it if not already
2335 * folded. */
2336#define TRIE_READ_CHAR STMT_START { \
2337 wordlen++; \
2338 if ( UTF ) { \
2339 /* if it is UTF then it is either already folded, or does not need \
2340 * folding */ \
1c1d615a 2341 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2342 } \
2343 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2344 /* This folder implies Unicode rules, which in the range expressible \
2345 * by not UTF is the lower case, with the two exceptions, one of \
2346 * which should have been taken care of before calling this */ \
2347 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2348 uvc = toLOWER_L1(*uc); \
2349 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2350 len = 1; \
914a25d5
KW
2351 } else { \
2352 /* raw data, will be folded later if needed */ \
2353 uvc = (U32)*uc; \
2354 len = 1; \
2355 } \
786e8c11
YO
2356} STMT_END
2357
2358
2359
2360#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2361 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
2362 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2363 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
2364 } \
2365 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2366 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2367 TRIE_LIST_CUR( state )++; \
2368} STMT_END
07be1b83 2369
786e8c11
YO
2370#define TRIE_LIST_NEW(state) STMT_START { \
2371 Newxz( trie->states[ state ].trans.list, \
2372 4, reg_trie_trans_le ); \
2373 TRIE_LIST_CUR( state ) = 1; \
2374 TRIE_LIST_LEN( state ) = 4; \
2375} STMT_END
07be1b83 2376
786e8c11
YO
2377#define TRIE_HANDLE_WORD(state) STMT_START { \
2378 U16 dupe= trie->states[ state ].wordnum; \
2379 regnode * const noper_next = regnext( noper ); \
2380 \
786e8c11
YO
2381 DEBUG_r({ \
2382 /* store the word for dumping */ \
2383 SV* tmp; \
2384 if (OP(noper) != NOTHING) \
740cce10 2385 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2386 else \
740cce10 2387 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2388 av_push( trie_words, tmp ); \
786e8c11
YO
2389 }); \
2390 \
2391 curword++; \
2e64971a
DM
2392 trie->wordinfo[curword].prev = 0; \
2393 trie->wordinfo[curword].len = wordlen; \
2394 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2395 \
2396 if ( noper_next < tail ) { \
2397 if (!trie->jump) \
538e84ed
KW
2398 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2399 sizeof(U16) ); \
7f69552c 2400 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2401 if (!jumper) \
2402 jumper = noper_next; \
2403 if (!nextbranch) \
2404 nextbranch= regnext(cur); \
2405 } \
2406 \
2407 if ( dupe ) { \
2e64971a
DM
2408 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2409 /* chain, so that when the bits of chain are later */\
2410 /* linked together, the dups appear in the chain */\
2411 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2412 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2413 } else { \
2414 /* we haven't inserted this word yet. */ \
2415 trie->states[ state ].wordnum = curword; \
2416 } \
2417} STMT_END
07be1b83 2418
3dab1dad 2419
786e8c11
YO
2420#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2421 ( ( base + charid >= ucharcount \
2422 && base + charid < ubound \
2423 && state == trie->trans[ base - ucharcount + charid ].check \
2424 && trie->trans[ base - ucharcount + charid ].next ) \
2425 ? trie->trans[ base - ucharcount + charid ].next \
2426 : ( state==1 ? special : 0 ) \
2427 )
3dab1dad 2428
786e8c11
YO
2429#define MADE_TRIE 1
2430#define MADE_JUMP_TRIE 2
2431#define MADE_EXACT_TRIE 4
3dab1dad 2432
a3621e74 2433STATIC I32
538e84ed
KW
2434S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2435 regnode *first, regnode *last, regnode *tail,
2436 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2437{
2438 /* first pass, loop through and scan words */
2439 reg_trie_data *trie;
55eed653 2440 HV *widecharmap = NULL;
2b8b4781 2441 AV *revcharmap = newAV();
a3621e74 2442 regnode *cur;
a3621e74
YO
2443 STRLEN len = 0;
2444 UV uvc = 0;
2445 U16 curword = 0;
2446 U32 next_alloc = 0;
786e8c11
YO
2447 regnode *jumper = NULL;
2448 regnode *nextbranch = NULL;
7f69552c 2449 regnode *convert = NULL;
2e64971a 2450 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2451 /* we just use folder as a flag in utf8 */
1e696034 2452 const U8 * folder = NULL;
a3621e74 2453
2b8b4781 2454#ifdef DEBUGGING
cf78de0b 2455 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2456 AV *trie_words = NULL;
2457 /* along with revcharmap, this only used during construction but both are
2458 * useful during debugging so we store them in the struct when debugging.
8e11feef 2459 */
2b8b4781 2460#else
cf78de0b 2461 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2462 STRLEN trie_charcount=0;
3dab1dad 2463#endif
2b8b4781 2464 SV *re_trie_maxbuff;
a3621e74 2465 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2466
2467 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2468#ifndef DEBUGGING
2469 PERL_UNUSED_ARG(depth);
2470#endif
a3621e74 2471
1e696034 2472 switch (flags) {
a4525e78 2473 case EXACT: case EXACTL: break;
2f7f8cb1 2474 case EXACTFA:
fab2782b 2475 case EXACTFU_SS:
a4525e78
KW
2476 case EXACTFU:
2477 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2478 case EXACTF: folder = PL_fold; break;
fab2782b 2479 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2480 }
2481
c944940b 2482 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2483 trie->refcount = 1;
3dab1dad 2484 trie->startstate = 1;
786e8c11 2485 trie->wordcount = word_count;
f8fc2ecf 2486 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2487 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2488 if (flags == EXACT || flags == EXACTL)
c944940b 2489 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2490 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2491 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2492
a3621e74 2493 DEBUG_r({
2b8b4781 2494 trie_words = newAV();
a3621e74 2495 });
a3621e74 2496
0111c4fd 2497 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2498 assert(re_trie_maxbuff);
a3621e74 2499 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2500 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2501 }
df826430 2502 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2503 Perl_re_indentf( aTHX_
cb41e5d6
YO
2504 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2505 depth+1,
538e84ed
KW
2506 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2507 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2508 });
538e84ed 2509
7f69552c
YO
2510 /* Find the node we are going to overwrite */
2511 if ( first == startbranch && OP( last ) != BRANCH ) {
2512 /* whole branch chain */
2513 convert = first;
2514 } else {
2515 /* branch sub-chain */
2516 convert = NEXTOPER( first );
2517 }
538e84ed 2518
a3621e74
YO
2519 /* -- First loop and Setup --
2520
2521 We first traverse the branches and scan each word to determine if it
2522 contains widechars, and how many unique chars there are, this is
2523 important as we have to build a table with at least as many columns as we
2524 have unique chars.
2525
2526 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2527 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2528 the native representation of the character value as the key and IV's for
2529 the coded index.
a3621e74
YO
2530
2531 *TODO* If we keep track of how many times each character is used we can
2532 remap the columns so that the table compression later on is more
3b753521 2533 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2534 middle and the least common are on the outside. IMO this would be better
2535 than a most to least common mapping as theres a decent chance the most
2536 common letter will share a node with the least common, meaning the node
486ec47a 2537 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2538 case is when we have the least common nodes twice.
2539
2540 */
2541
a3621e74 2542 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2543 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2544 const U8 *uc;
2545 const U8 *e;
bc031a7d 2546 int foldlen = 0;
07be1b83 2547 U32 wordlen = 0; /* required init */
bc031a7d
KW
2548 STRLEN minchars = 0;
2549 STRLEN maxchars = 0;
538e84ed
KW
2550 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2551 bitmap?*/
a3621e74 2552
3dab1dad 2553 if (OP(noper) == NOTHING) {
df826430 2554 regnode *noper_next= regnext(noper);
944e05e3
YO
2555 if (noper_next < tail)
2556 noper= noper_next;
2557 }
2558
2559 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2560 uc= (U8*)STRING(noper);
2561 e= uc + STR_LEN(noper);
2562 } else {
2563 trie->minlen= 0;
2564 continue;
3dab1dad 2565 }
df826430 2566
944e05e3 2567
fab2782b 2568 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2569 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2570 regardless of encoding */
fab2782b
YO
2571 if (OP( noper ) == EXACTFU_SS) {
2572 /* false positives are ok, so just set this */
0dc4a61d 2573 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2574 }
2575 }
bc031a7d
KW
2576 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2577 branch */
3dab1dad 2578 TRIE_CHARCOUNT(trie)++;
a3621e74 2579 TRIE_READ_CHAR;
645de4ce 2580
bc031a7d
KW
2581 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2582 * is in effect. Under /i, this character can match itself, or
2583 * anything that folds to it. If not under /i, it can match just
2584 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2585 * all fold to k, and all are single characters. But some folds
2586 * expand to more than one character, so for example LATIN SMALL
2587 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2588 * the string beginning at 'uc' is 'ffi', it could be matched by
2589 * three characters, or just by the one ligature character. (It
2590 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2591 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2592 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2593 * match.) The trie needs to know the minimum and maximum number
2594 * of characters that could match so that it can use size alone to
2595 * quickly reject many match attempts. The max is simple: it is
2596 * the number of folded characters in this branch (since a fold is
2597 * never shorter than what folds to it. */
2598
2599 maxchars++;
2600
2601 /* And the min is equal to the max if not under /i (indicated by
2602 * 'folder' being NULL), or there are no multi-character folds. If
2603 * there is a multi-character fold, the min is incremented just
2604 * once, for the character that folds to the sequence. Each
2605 * character in the sequence needs to be added to the list below of
2606 * characters in the trie, but we count only the first towards the
2607 * min number of characters needed. This is done through the
2608 * variable 'foldlen', which is returned by the macros that look
2609 * for these sequences as the number of bytes the sequence
2610 * occupies. Each time through the loop, we decrement 'foldlen' by
2611 * how many bytes the current char occupies. Only when it reaches
2612 * 0 do we increment 'minchars' or look for another multi-character
2613 * sequence. */
2614 if (folder == NULL) {
2615 minchars++;
2616 }
2617 else if (foldlen > 0) {
2618 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2619 }
2620 else {
bc031a7d
KW
2621 minchars++;
2622
2623 /* See if *uc is the beginning of a multi-character fold. If
2624 * so, we decrement the length remaining to look at, to account
2625 * for the current character this iteration. (We can use 'uc'
2626 * instead of the fold returned by TRIE_READ_CHAR because for
2627 * non-UTF, the latin1_safe macro is smart enough to account
2628 * for all the unfolded characters, and because for UTF, the
2629 * string will already have been folded earlier in the
2630 * compilation process */
2631 if (UTF) {
2632 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2633 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2634 }
2635 }
bc031a7d
KW
2636 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2637 foldlen--;
2638 }
645de4ce 2639 }
bc031a7d
KW
2640
2641 /* The current character (and any potential folds) should be added
2642 * to the possible matching characters for this position in this
2643 * branch */
a3621e74 2644 if ( uvc < 256 ) {
fab2782b
YO
2645 if ( folder ) {
2646 U8 folded= folder[ (U8) uvc ];
2647 if ( !trie->charmap[ folded ] ) {
2648 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2649 TRIE_STORE_REVCHAR( folded );
2650 }
2651 }
a3621e74
YO
2652 if ( !trie->charmap[ uvc ] ) {
2653 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2654 TRIE_STORE_REVCHAR( uvc );
a3621e74 2655 }
02daf0ab 2656 if ( set_bit ) {
62012aee
KW
2657 /* store the codepoint in the bitmap, and its folded
2658 * equivalent. */
fab2782b 2659 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2660
2661 /* store the folded codepoint */
fab2782b 2662 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2663
2664 if ( !UTF ) {
2665 /* store first byte of utf8 representation of
acdf4139 2666 variant codepoints */
6f2d5cbc 2667 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2668 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2669 }
2670 }
02daf0ab
YO
2671 set_bit = 0; /* We've done our bit :-) */
2672 }
a3621e74 2673 } else {
bc031a7d
KW
2674
2675 /* XXX We could come up with the list of code points that fold
2676 * to this using PL_utf8_foldclosures, except not for
2677 * multi-char folds, as there may be multiple combinations
2678 * there that could work, which needs to wait until runtime to
2679 * resolve (The comment about LIGATURE FFI above is such an
2680 * example */
2681
a3621e74 2682 SV** svpp;
55eed653
NC
2683 if ( !widecharmap )
2684 widecharmap = newHV();
a3621e74 2685
55eed653 2686 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2687
2688 if ( !svpp )
e4584336 2689 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2690
2691 if ( !SvTRUE( *svpp ) ) {
2692 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2693 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2694 }
2695 }
bc031a7d
KW
2696 } /* end loop through characters in this branch of the trie */
2697
2698 /* We take the min and max for this branch and combine to find the min
2699 * and max for all branches processed so far */
3dab1dad 2700 if( cur == first ) {
bc031a7d
KW
2701 trie->minlen = minchars;
2702 trie->maxlen = maxchars;
2703 } else if (minchars < trie->minlen) {
2704 trie->minlen = minchars;
2705 } else if (maxchars > trie->maxlen) {
2706 trie->maxlen = maxchars;
fab2782b 2707 }
a3621e74
YO
2708 } /* end first pass */
2709 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2710 Perl_re_indentf( aTHX_
cb41e5d6
YO
2711 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2712 depth+1,
55eed653 2713 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2714 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2715 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2716 );
a3621e74
YO
2717
2718 /*
2719 We now know what we are dealing with in terms of unique chars and
2720 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2721 representation using a flat table will take. If it's over a reasonable
2722 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2723 conservative but potentially much slower representation using an array
2724 of lists.
2725
2726 At the end we convert both representations into the same compressed
2727 form that will be used in regexec.c for matching with. The latter
2728 is a form that cannot be used to construct with but has memory
2729 properties similar to the list form and access properties similar
2730 to the table form making it both suitable for fast searches and
2731 small enough that its feasable to store for the duration of a program.
2732
2733 See the comment in the code where the compressed table is produced
2734 inplace from the flat tabe representation for an explanation of how
2735 the compression works.
2736
2737 */
2738
2739
2e64971a
DM
2740 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2741 prev_states[1] = 0;
2742
538e84ed
KW
2743 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2744 > SvIV(re_trie_maxbuff) )
2745 {
a3621e74
YO
2746 /*
2747 Second Pass -- Array Of Lists Representation
2748
2749 Each state will be represented by a list of charid:state records
2750 (reg_trie_trans_le) the first such element holds the CUR and LEN
2751 points of the allocated array. (See defines above).
2752
2753 We build the initial structure using the lists, and then convert
2754 it into the compressed table form which allows faster lookups
2755 (but cant be modified once converted).
a3621e74
YO
2756 */
2757
a3621e74
YO
2758 STRLEN transcount = 1;
2759
6ad9a8ab 2760 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2761 depth+1));
686b73d4 2762
c944940b
JH
2763 trie->states = (reg_trie_state *)
2764 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2765 sizeof(reg_trie_state) );
a3621e74
YO
2766 TRIE_LIST_NEW(1);
2767 next_alloc = 2;
2768
2769 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2770
df826430 2771 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2772 U32 state = 1; /* required init */
2773 U16 charid = 0; /* sanity init */
07be1b83 2774 U32 wordlen = 0; /* required init */
c445ea15 2775
df826430
YO
2776 if (OP(noper) == NOTHING) {
2777 regnode *noper_next= regnext(noper);
944e05e3
YO
2778 if (noper_next < tail)
2779 noper= noper_next;
df826430
YO
2780 }
2781
944e05e3
YO
2782 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2783 const U8 *uc= (U8*)STRING(noper);
2784 const U8 *e= uc + STR_LEN(noper);
2785
786e8c11 2786 for ( ; uc < e ; uc += len ) {
c445ea15 2787
786e8c11 2788 TRIE_READ_CHAR;
c445ea15 2789
786e8c11
YO
2790 if ( uvc < 256 ) {
2791 charid = trie->charmap[ uvc ];
c445ea15 2792 } else {
538e84ed
KW
2793 SV** const svpp = hv_fetch( widecharmap,
2794 (char*)&uvc,
2795 sizeof( UV ),
2796 0);
786e8c11
YO
2797 if ( !svpp ) {
2798 charid = 0;
2799 } else {
2800 charid=(U16)SvIV( *svpp );
2801 }
c445ea15 2802 }
538e84ed
KW
2803 /* charid is now 0 if we dont know the char read, or
2804 * nonzero if we do */
786e8c11 2805 if ( charid ) {
a3621e74 2806
786e8c11
YO
2807 U16 check;
2808 U32 newstate = 0;
a3621e74 2809
786e8c11
YO
2810 charid--;
2811 if ( !trie->states[ state ].trans.list ) {
2812 TRIE_LIST_NEW( state );
c445ea15 2813 }
538e84ed
KW
2814 for ( check = 1;
2815 check <= TRIE_LIST_USED( state );
2816 check++ )
2817 {
2818 if ( TRIE_LIST_ITEM( state, check ).forid
2819 == charid )
2820 {
786e8c11
YO
2821 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2822 break;
2823 }
2824 }
2825 if ( ! newstate ) {
2826 newstate = next_alloc++;
2e64971a 2827 prev_states[newstate] = state;
786e8c11
YO
2828 TRIE_LIST_PUSH( state, charid, newstate );
2829 transcount++;
2830 }
2831 state = newstate;
2832 } else {
2833 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2834 }
a28509cc 2835 }
c445ea15 2836 }
3dab1dad 2837 TRIE_HANDLE_WORD(state);
a3621e74
YO
2838
2839 } /* end second pass */
2840
1e2e3d02 2841 /* next alloc is the NEXT state to be allocated */
538e84ed 2842 trie->statecount = next_alloc;
c944940b
JH
2843 trie->states = (reg_trie_state *)
2844 PerlMemShared_realloc( trie->states,
2845 next_alloc
2846 * sizeof(reg_trie_state) );
a3621e74 2847
3dab1dad 2848 /* and now dump it out before we compress it */
2b8b4781
NC
2849 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2850 revcharmap, next_alloc,
2851 depth+1)
1e2e3d02 2852 );
a3621e74 2853
c944940b
JH
2854 trie->trans = (reg_trie_trans *)
2855 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2856 {
2857 U32 state;
a3621e74
YO
2858 U32 tp = 0;
2859 U32 zp = 0;
2860
2861
2862 for( state=1 ; state < next_alloc ; state ++ ) {
2863 U32 base=0;
2864
2865 /*
2866 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2867 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
2868 );
2869 */
2870
2871 if (trie->states[state].trans.list) {
2872 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2873 U16 maxid=minid;
a28509cc 2874 U16 idx;
a3621e74
YO
2875
2876 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2877 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2878 if ( forid < minid ) {
2879 minid=forid;
2880 } else if ( forid > maxid ) {
2881 maxid=forid;
2882 }
a3621e74
YO
2883 }
2884 if ( transcount < tp + maxid - minid + 1) {
2885 transcount *= 2;
c944940b
JH
2886 trie->trans = (reg_trie_trans *)
2887 PerlMemShared_realloc( trie->trans,
446bd890
NC
2888 transcount
2889 * sizeof(reg_trie_trans) );
538e84ed
KW
2890 Zero( trie->trans + (transcount / 2),
2891 transcount / 2,
2892 reg_trie_trans );
a3621e74
YO
2893 }
2894 base = trie->uniquecharcount + tp - minid;
2895 if ( maxid == minid ) {
2896 U32 set = 0;
2897 for ( ; zp < tp ; zp++ ) {
2898 if ( ! trie->trans[ zp ].next ) {
2899 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2900 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2901 1).newstate;
a3621e74
YO
2902 trie->trans[ zp ].check = state;
2903 set = 1;
2904 break;
2905 }
2906 }
2907 if ( !set ) {
538e84ed
KW
2908 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2909 1).newstate;
a3621e74
YO
2910 trie->trans[ tp ].check = state;
2911 tp++;
2912 zp = tp;
2913 }
2914 } else {
2915 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2916 const U32 tid = base
2917 - trie->uniquecharcount
2918 + TRIE_LIST_ITEM( state, idx ).forid;
2919 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2920 idx ).newstate;
a3621e74
YO
2921 trie->trans[ tid ].check = state;
2922 }
2923 tp += ( maxid - minid + 1 );
2924 }
2925 Safefree(trie->states[ state ].trans.list);
2926 }
2927 /*
2928 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2929 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
2930 );
2931 */
2932 trie->states[ state ].trans.base=base;
2933 }
cc601c31 2934 trie->lasttrans = tp + 1;
a3621e74
YO
2935 }
2936 } else {
2937 /*
2938 Second Pass -- Flat Table Representation.
2939
b423522f
KW
2940 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2941 each. We know that we will need Charcount+1 trans at most to store
2942 the data (one row per char at worst case) So we preallocate both
2943 structures assuming worst case.
a3621e74
YO
2944
2945 We then construct the trie using only the .next slots of the entry
2946 structs.
2947
b423522f
KW
2948 We use the .check field of the first entry of the node temporarily
2949 to make compression both faster and easier by keeping track of how
2950 many non zero fields are in the node.
a3621e74
YO
2951
2952 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2953 transition.
2954
b423522f
KW
2955 There are two terms at use here: state as a TRIE_NODEIDX() which is
2956 a number representing the first entry of the node, and state as a
2957 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2958 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2959 if there are 2 entrys per node. eg:
a3621e74
YO
2960
2961 A B A B
2962 1. 2 4 1. 3 7
2963 2. 0 3 3. 0 5
2964 3. 0 0 5. 0 0
2965 4. 0 0 7. 0 0
2966
b423522f
KW
2967 The table is internally in the right hand, idx form. However as we
2968 also have to deal with the states array which is indexed by nodenum
2969 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2970
2971 */
6ad9a8ab 2972 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 2973 depth+1));
3dab1dad 2974
c944940b
JH
2975 trie->trans = (reg_trie_trans *)
2976 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2977 * trie->uniquecharcount + 1,
2978 sizeof(reg_trie_trans) );
2979 trie->states = (reg_trie_state *)
2980 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2981 sizeof(reg_trie_state) );
a3621e74
YO
2982 next_alloc = trie->uniquecharcount + 1;
2983
3dab1dad 2984
a3621e74
YO
2985 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2986
df826430 2987 regnode *noper = NEXTOPER( cur );
a3621e74
YO
2988
2989 U32 state = 1; /* required init */
2990
2991 U16 charid = 0; /* sanity init */
2992 U32 accept_state = 0; /* sanity init */
a3621e74 2993
07be1b83 2994 U32 wordlen = 0; /* required init */
a3621e74 2995
df826430
YO
2996 if (OP(noper) == NOTHING) {
2997 regnode *noper_next= regnext(noper);
944e05e3
YO
2998 if (noper_next < tail)
2999 noper= noper_next;
df826430 3000 }
fab2782b 3001
944e05e3
YO
3002 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3003 const U8 *uc= (U8*)STRING(noper);
3004 const U8 *e= uc + STR_LEN(noper);
3005
786e8c11 3006 for ( ; uc < e ; uc += len ) {
a3621e74 3007
786e8c11 3008 TRIE_READ_CHAR;
a3621e74 3009
786e8c11
YO
3010 if ( uvc < 256 ) {
3011 charid = trie->charmap[ uvc ];
3012 } else {
538e84ed
KW
3013 SV* const * const svpp = hv_fetch( widecharmap,
3014 (char*)&uvc,
3015 sizeof( UV ),
3016 0);
786e8c11 3017 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3018 }
786e8c11
YO
3019 if ( charid ) {
3020 charid--;
3021 if ( !trie->trans[ state + charid ].next ) {
3022 trie->trans[ state + charid ].next = next_alloc;
3023 trie->trans[ state ].check++;
2e64971a
DM
3024 prev_states[TRIE_NODENUM(next_alloc)]
3025 = TRIE_NODENUM(state);
786e8c11
YO
3026 next_alloc += trie->uniquecharcount;
3027 }
3028 state = trie->trans[ state + charid ].next;
3029 } else {
3030 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
3031 }
538e84ed
KW
3032 /* charid is now 0 if we dont know the char read, or
3033 * nonzero if we do */
a3621e74 3034 }
a3621e74 3035 }
3dab1dad
YO
3036 accept_state = TRIE_NODENUM( state );
3037 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3038
3039 } /* end second pass */
3040
3dab1dad 3041 /* and now dump it out before we compress it */
2b8b4781
NC
3042 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3043 revcharmap,
3044 next_alloc, depth+1));
a3621e74 3045
a3621e74
YO
3046 {
3047 /*
3048 * Inplace compress the table.*
3049
3050 For sparse data sets the table constructed by the trie algorithm will
3051 be mostly 0/FAIL transitions or to put it another way mostly empty.
3052 (Note that leaf nodes will not contain any transitions.)
3053
3054 This algorithm compresses the tables by eliminating most such
3055 transitions, at the cost of a modest bit of extra work during lookup:
3056
3057 - Each states[] entry contains a .base field which indicates the
3058 index in the state[] array wheres its transition data is stored.
3059
3b753521 3060 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3061
3062 - If .base is nonzero then charid is added to it to find an entry in
3063 the trans array.
3064
3065 -If trans[states[state].base+charid].check!=state then the
3066 transition is taken to be a 0/Fail transition. Thus if there are fail
3067 transitions at the front of the node then the .base offset will point
3068 somewhere inside the previous nodes data (or maybe even into a node
3069 even earlier), but the .check field determines if the transition is
3070 valid.
3071
786e8c11 3072 XXX - wrong maybe?
a3621e74 3073 The following process inplace converts the table to the compressed
3b753521 3074 table: We first do not compress the root node 1,and mark all its
a3621e74 3075 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3076 allows us to do a DFA construction from the compressed table later,
3077 and ensures that any .base pointers we calculate later are greater
3078 than 0.
a3621e74
YO
3079
3080 - We set 'pos' to indicate the first entry of the second node.
3081
3082 - We then iterate over the columns of the node, finding the first and
3083 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3084 and set the .check pointers accordingly, and advance pos
3085 appropriately and repreat for the next node. Note that when we copy
3086 the next pointers we have to convert them from the original
3087 NODEIDX form to NODENUM form as the former is not valid post
3088 compression.
3089
3090 - If a node has no transitions used we mark its base as 0 and do not
3091 advance the pos pointer.
3092
3093 - If a node only has one transition we use a second pointer into the
3094 structure to fill in allocated fail transitions from other states.
3095 This pointer is independent of the main pointer and scans forward
3096 looking for null transitions that are allocated to a state. When it
3097 finds one it writes the single transition into the "hole". If the
786e8c11 3098 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3099
3100 - Once compressed we can Renew/realloc the structures to release the
3101 excess space.
3102
3103 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3104 specifically Fig 3.47 and the associated pseudocode.
3105
3106 demq
3107 */
a3b680e6 3108 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3109 U32 state, charid;
a3621e74 3110 U32 pos = 0, zp=0;
1e2e3d02 3111 trie->statecount = laststate;
a3621e74
YO
3112
3113 for ( state = 1 ; state < laststate ; state++ ) {
3114 U8 flag = 0;
a28509cc
AL
3115 const U32 stateidx = TRIE_NODEIDX( state );
3116 const U32 o_used = trie->trans[ stateidx ].check;
3117 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3118 trie->trans[ stateidx ].check = 0;
3119
538e84ed
KW
3120 for ( charid = 0;
3121 used && charid < trie->uniquecharcount;
3122 charid++ )
3123 {
a3621e74
YO
3124 if ( flag || trie->trans[ stateidx + charid ].next ) {
3125 if ( trie->trans[ stateidx + charid ].next ) {
3126 if (o_used == 1) {
3127 for ( ; zp < pos ; zp++ ) {
3128 if ( ! trie->trans[ zp ].next ) {
3129 break;
3130 }
3131 }
538e84ed
KW
3132 trie->states[ state ].trans.base
3133 = zp
3134 + trie->uniquecharcount
3135 - charid ;
3136 trie->trans[ zp ].next
3137 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3138 + charid ].next );
a3621e74
YO
3139 trie->trans[ zp ].check = state;
3140 if ( ++zp > pos ) pos = zp;
3141 break;
3142 }
3143 used--;
3144 }
3145 if ( !flag ) {
3146 flag = 1;
538e84ed
KW
3147 trie->states[ state ].trans.base
3148 = pos + trie->uniquecharcount - charid ;
a3621e74 3149 }
538e84ed
KW
3150 trie->trans[ pos ].next
3151 = SAFE_TRIE_NODENUM(
3152 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3153 trie->trans[ pos ].check = state;
3154 pos++;
3155 }
3156 }
3157 }
cc601c31 3158 trie->lasttrans = pos + 1;
c944940b
JH
3159 trie->states = (reg_trie_state *)
3160 PerlMemShared_realloc( trie->states, laststate
3161 * sizeof(reg_trie_state) );
a3621e74 3162 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3163 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
cb41e5d6 3164 depth+1,
538e84ed
KW
3165 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3166 + 1 ),
3167 (IV)next_alloc,
3168 (IV)pos,
3169 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3170 );
3171
3172 } /* end table compress */
3173 }
1e2e3d02 3174 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3175 Perl_re_indentf( aTHX_ "Statecount:%"UVxf" Lasttrans:%"UVxf"\n",
cb41e5d6 3176 depth+1,
1e2e3d02
YO
3177 (UV)trie->statecount,
3178 (UV)trie->lasttrans)
3179 );
cc601c31 3180 /* resize the trans array to remove unused space */
c944940b
JH
3181 trie->trans = (reg_trie_trans *)
3182 PerlMemShared_realloc( trie->trans, trie->lasttrans
3183 * sizeof(reg_trie_trans) );
a3621e74 3184
538e84ed 3185 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3186 U8 nodetype =(U8)(flags & 0xFF);
3187 char *str=NULL;
538e84ed 3188
07be1b83 3189#ifdef DEBUGGING
e62cc96a 3190 regnode *optimize = NULL;
7122b237
YO
3191#ifdef RE_TRACK_PATTERN_OFFSETS
3192
b57a0404
JH
3193 U32 mjd_offset = 0;
3194 U32 mjd_nodelen = 0;
7122b237
YO
3195#endif /* RE_TRACK_PATTERN_OFFSETS */
3196#endif /* DEBUGGING */
a3621e74 3197 /*
3dab1dad
YO
3198 This means we convert either the first branch or the first Exact,
3199 depending on whether the thing following (in 'last') is a branch
3200 or not and whther first is the startbranch (ie is it a sub part of
3201 the alternation or is it the whole thing.)
3b753521 3202 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3203 the whole branch sequence, including the first.
a3621e74 3204 */
3dab1dad 3205 /* Find the node we are going to overwrite */
7f69552c 3206 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3207 /* branch sub-chain */
3dab1dad 3208 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3209#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3210 DEBUG_r({
3211 mjd_offset= Node_Offset((convert));
3212 mjd_nodelen= Node_Length((convert));
3213 });
7122b237 3214#endif
7f69552c 3215 /* whole branch chain */
7122b237
YO
3216 }
3217#ifdef RE_TRACK_PATTERN_OFFSETS
3218 else {
7f69552c
YO
3219 DEBUG_r({
3220 const regnode *nop = NEXTOPER( convert );
3221 mjd_offset= Node_Offset((nop));
3222 mjd_nodelen= Node_Length((nop));
3223 });
07be1b83
YO
3224 }
3225 DEBUG_OPTIMISE_r(
6ad9a8ab 3226 Perl_re_indentf( aTHX_ "MJD offset:%"UVuf" MJD length:%"UVuf"\n",
cb41e5d6 3227 depth+1,
786e8c11 3228 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3229 );
7122b237 3230#endif
538e84ed 3231 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3232 split out as an EXACT and put in front of the TRIE node. */
3233 trie->startstate= 1;
55eed653 3234 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 3235 U32 state;
1e2e3d02 3236 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3237 U32 ofs = 0;
8e11feef
RGS
3238 I32 idx = -1;
3239 U32 count = 0;
3240 const U32 base = trie->states[ state ].trans.base;
a3621e74 3241
3dab1dad 3242 if ( trie->states[state].wordnum )
8e11feef 3243 count = 1;
a3621e74 3244
8e11feef 3245 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3246 if ( ( base + ofs >= trie->uniquecharcount ) &&
3247 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3248 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3249 {
3dab1dad 3250 if ( ++count > 1 ) {
2b8b4781 3251 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 3252 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3253 if ( state == 1 ) break;
3dab1dad
YO
3254 if ( count == 2 ) {
3255 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3256 DEBUG_OPTIMISE_r(
6ad9a8ab 3257 Perl_re_indentf( aTHX_ "New Start State=%"UVuf" Class: [",
cb41e5d6 3258 depth+1,
786e8c11 3259 (UV)state));
be8e71aa 3260 if (idx >= 0) {
2b8b4781 3261 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 3262 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3263
3dab1dad 3264 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
3265 if ( folder )
3266 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 3267 DEBUG_OPTIMISE_r(
6ad9a8ab 3268 Perl_re_printf( aTHX_ "%s", (char*)ch)
3dab1dad 3269 );
8e11feef
RGS
3270 }
3271 }
3272 TRIE_BITMAP_SET(trie,*ch);
3273 if ( folder )
3274 TRIE_BITMAP_SET(trie,folder[ *ch ]);
6ad9a8ab 3275 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
8e11feef
RGS
3276 }
3277 idx = ofs;
3278 }
3dab1dad
YO
3279 }
3280 if ( count == 1 ) {
2b8b4781 3281 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
3282 STRLEN len;
3283 char *ch = SvPV( *tmp, len );
de734bd5
A
3284 DEBUG_OPTIMISE_r({
3285 SV *sv=sv_newmortal();
6ad9a8ab 3286 Perl_re_indentf( aTHX_ "Prefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
cb41e5d6 3287 depth+1,
538e84ed
KW
3288 (UV)state, (UV)idx,
3289 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3290 PL_colors[0], PL_colors[1],
3291 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3292 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3293 )
3294 );
3295 });
3dab1dad
YO
3296 if ( state==1 ) {
3297 OP( convert ) = nodetype;
3298 str=STRING(convert);
3299 STR_LEN(convert)=0;
3300 }
c490c714
YO
3301 STR_LEN(convert) += len;
3302 while (len--)
de734bd5 3303 *str++ = *ch++;
8e11feef 3304 } else {
538e84ed 3305#ifdef DEBUGGING
8e11feef 3306 if (state>1)
6ad9a8ab 3307 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
f9049ba1 3308#endif
8e11feef
RGS
3309 break;
3310 }
3311 }
2e64971a 3312 trie->prefixlen = (state-1);
3dab1dad 3313 if (str) {
8e11feef 3314 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3315 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3316 trie->startstate = state;
07be1b83
YO
3317 trie->minlen -= (state - 1);
3318 trie->maxlen -= (state - 1);
33809eae
JH
3319#ifdef DEBUGGING
3320 /* At least the UNICOS C compiler choked on this
3321 * being argument to DEBUG_r(), so let's just have
3322 * it right here. */
3323 if (
3324#ifdef PERL_EXT_RE_BUILD
3325 1
3326#else
3327 DEBUG_r_TEST
3328#endif
3329 ) {
3330 regnode *fix = convert;
3331 U32 word = trie->wordcount;
3332 mjd_nodelen++;
3333 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3334 while( ++fix < n ) {
3335 Set_Node_Offset_Length(fix, 0, 0);
3336 }
3337 while (word--) {
3338 SV ** const tmp = av_fetch( trie_words, word, 0 );
3339 if (tmp) {
3340 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3341 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3342 else
3343 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3344 }
3345 }
3346 }
3347#endif
8e11feef
RGS
3348 if (trie->maxlen) {
3349 convert = n;
3350 } else {
3dab1dad 3351 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3352 DEBUG_r(optimize= n);
3dab1dad
YO
3353 }
3354 }
3355 }
538e84ed
KW
3356 if (!jumper)
3357 jumper = last;
3dab1dad 3358 if ( trie->maxlen ) {
8e11feef
RGS
3359 NEXT_OFF( convert ) = (U16)(tail - convert);
3360 ARG_SET( convert, data_slot );
538e84ed
KW
3361 /* Store the offset to the first unabsorbed branch in
3362 jump[0], which is otherwise unused by the jump logic.
786e8c11 3363 We use this when dumping a trie and during optimisation. */
538e84ed 3364 if (trie->jump)
7f69552c 3365 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3366
6c48061a
YO
3367 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3368 * and there is a bitmap
3369 * and the first "jump target" node we found leaves enough room
3370 * then convert the TRIE node into a TRIEC node, with the bitmap
3371 * embedded inline in the opcode - this is hypothetically faster.
3372 */
3373 if ( !trie->states[trie->startstate].wordnum
3374 && trie->bitmap
3375 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3376 {
3377 OP( convert ) = TRIEC;
3378 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3379 PerlMemShared_free(trie->bitmap);
786e8c11 3380 trie->bitmap= NULL;
538e84ed 3381 } else
786e8c11 3382 OP( convert ) = TRIE;
a3621e74 3383
3dab1dad
YO
3384 /* store the type in the flags */
3385 convert->flags = nodetype;
a5ca303d 3386 DEBUG_r({
538e84ed
KW
3387 optimize = convert
3388 + NODE_STEP_REGNODE
a5ca303d
YO
3389 + regarglen[ OP( convert ) ];
3390 });
538e84ed 3391 /* XXX We really should free up the resource in trie now,
a5ca303d 3392 as we won't use them - (which resources?) dmq */
3dab1dad 3393 }
a3621e74 3394 /* needed for dumping*/
e62cc96a 3395 DEBUG_r(if (optimize) {
07be1b83 3396 regnode *opt = convert;
bcdf7404 3397
e62cc96a 3398 while ( ++opt < optimize) {
07be1b83
YO
3399 Set_Node_Offset_Length(opt,0,0);
3400 }
538e84ed
KW
3401 /*
3402 Try to clean up some of the debris left after the
786e8c11 3403 optimisation.
a3621e74 3404 */
786e8c11 3405 while( optimize < jumper ) {
07be1b83 3406 mjd_nodelen += Node_Length((optimize));
a3621e74 3407 OP( optimize ) = OPTIMIZED;
07be1b83 3408 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3409 optimize++;
3410 }
07be1b83 3411 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3412 });
3413 } /* end node insert */
2e64971a
DM
3414
3415 /* Finish populating the prev field of the wordinfo array. Walk back
3416 * from each accept state until we find another accept state, and if
3417 * so, point the first word's .prev field at the second word. If the
3418 * second already has a .prev field set, stop now. This will be the
3419 * case either if we've already processed that word's accept state,
3b753521
FN
3420 * or that state had multiple words, and the overspill words were
3421 * already linked up earlier.
2e64971a
DM
3422 */
3423 {
3424 U16 word;
3425 U32 state;
3426 U16 prev;
3427
3428 for (word=1; word <= trie->wordcount; word++) {
3429 prev = 0;
3430 if (trie->wordinfo[word].prev)
3431 continue;
3432 state = trie->wordinfo[word].accept;
3433 while (state) {
3434 state = prev_states[state];
3435 if (!state)
3436 break;
3437 prev = trie->states[state].wordnum;
3438 if (prev)
3439 break;
3440 }
3441 trie->wordinfo[word].prev = prev;
3442 }
3443 Safefree(prev_states);
3444 }
3445
3446
3447 /* and now dump out the compressed format */
3448 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3449
55eed653 3450 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
3451#ifdef DEBUGGING
3452 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3453 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3454#else
03e70be4 3455 SvREFCNT_dec_NN(revcharmap);
07be1b83 3456#endif
538e84ed
KW
3457 return trie->jump
3458 ? MADE_JUMP_TRIE
3459 : trie->startstate>1
3460 ? MADE_EXACT_TRIE
786e8c11
YO
3461 : MADE_TRIE;
3462}
3463
615a2e7f
YO
3464STATIC regnode *
3465S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
786e8c11 3466{
b423522f
KW
3467/* The Trie is constructed and compressed now so we can build a fail array if
3468 * it's needed
786e8c11 3469
b423522f
KW
3470 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3471 3.32 in the
3472 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3473 Ullman 1985/88
786e8c11
YO
3474 ISBN 0-201-10088-6
3475
b423522f
KW
3476 We find the fail state for each state in the trie, this state is the longest
3477 proper suffix of the current state's 'word' that is also a proper prefix of
3478 another word in our trie. State 1 represents the word '' and is thus the
3479 default fail state. This allows the DFA not to have to restart after its
3480 tried and failed a word at a given point, it simply continues as though it