This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
update size after Renew
[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;
bb3f3ed2 180 I32 override_recoding;
b6d67071
KW
181#ifdef EBCDIC
182 I32 recode_x_to_native;
183#endif
9d53c457 184 I32 in_multi_char_class;
1acab4c5 185 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 186 within pattern */
b1603ef8 187 int code_index; /* next code_blocks[] slot */
ee273784 188 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
189 scan_frame *frame_head;
190 scan_frame *frame_last;
191 U32 frame_count;
7eec73eb 192 AV *warn_text;
dc6d7f5c 193#ifdef ADD_TO_REGEXEC
830247a4
IZ
194 char *starttry; /* -Dr: where regtry was called. */
195#define RExC_starttry (pRExC_state->starttry)
196#endif
d24ca0c5 197 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 198#ifdef DEBUGGING
be8e71aa 199 const char *lastparse;
3dab1dad 200 I32 lastnum;
1f1031fe 201 AV *paren_name_list; /* idx -> name */
d9a72fcc 202 U32 study_chunk_recursed_count;
c9f0d54c
YO
203 SV *mysv1;
204 SV *mysv2;
3dab1dad
YO
205#define RExC_lastparse (pRExC_state->lastparse)
206#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 207#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 208#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
209#define RExC_mysv (pRExC_state->mysv1)
210#define RExC_mysv1 (pRExC_state->mysv1)
211#define RExC_mysv2 (pRExC_state->mysv2)
212
3dab1dad 213#endif
512c0f5a 214 bool seen_unfolded_sharp_s;
911bd04e 215 bool strict;
da7cf1cc 216 bool study_started;
09b2b2e6 217};
830247a4 218
e2509266 219#define RExC_flags (pRExC_state->flags)
514a91f1 220#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 221#define RExC_precomp (pRExC_state->precomp)
285b5ca0
KW
222#define RExC_precomp_adj (pRExC_state->precomp_adj)
223#define RExC_adjusted_start (pRExC_state->adjusted_start)
711b303b 224#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 225#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 226#define RExC_rx (pRExC_state->rx)
f8fc2ecf 227#define RExC_rxi (pRExC_state->rxi)
fac92740 228#define RExC_start (pRExC_state->start)
830247a4
IZ
229#define RExC_end (pRExC_state->end)
230#define RExC_parse (pRExC_state->parse)
231#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
232
233/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
234 * EXACTF node, hence was parsed under /di rules. If later in the parse,
235 * something forces the pattern into using /ui rules, the sharp s should be
236 * folded into the sequence 'ss', which takes up more space than previously
237 * calculated. This means that the sizing pass needs to be restarted. (The
238 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
239 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
240 * so there is no need to resize [perl #125990]. */
241#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
242
7122b237 243#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
244#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
245 others */
7122b237 246#endif
830247a4 247#define RExC_emit (pRExC_state->emit)
f7c7e32a 248#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 249#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 250#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
251#define RExC_sawback (pRExC_state->sawback)
252#define RExC_seen (pRExC_state->seen)
253#define RExC_size (pRExC_state->size)
ee273784 254#define RExC_maxlen (pRExC_state->maxlen)
830247a4 255#define RExC_npar (pRExC_state->npar)
e2e6a0f1 256#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
257#define RExC_extralen (pRExC_state->extralen)
258#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 259#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 260#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 261#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
262#define RExC_open_parens (pRExC_state->open_parens)
263#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 264#define RExC_end_op (pRExC_state->end_op)
81714fb9 265#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
266#define RExC_recurse (pRExC_state->recurse)
267#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 268#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
269#define RExC_study_chunk_recursed_bytes \
270 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 271#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 272#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
273#ifdef EBCDIC
274# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275#endif
9d53c457 276#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
277#define RExC_frame_head (pRExC_state->frame_head)
278#define RExC_frame_last (pRExC_state->frame_last)
279#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 280#define RExC_strict (pRExC_state->strict)
da7cf1cc 281#define RExC_study_started (pRExC_state->study_started)
7eec73eb 282#define RExC_warn_text (pRExC_state->warn_text)
830247a4 283
99807a43
HS
284/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
288 */
289#define RExC_naughty (pRExC_state->naughty)
290#define TOO_NAUGHTY (10)
291#define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294#define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 297
a687059c
LW
298#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 300 ((*s) == '{' && regcurly(s)))
a687059c
LW
301
302/*
303 * Flags to be passed up and down.
304 */
a687059c 305#define WORST 0 /* Worst case. */
a3b492c3 306#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 307
e64f369d 308/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
311 * REGNODE_SIMPLE */
fda99bee 312#define SIMPLE 0x02
e64f369d 313#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
314#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
316#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
a687059c 319
3dab1dad
YO
320#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
321
07be1b83
YO
322/* whether trie related optimizations are enabled */
323#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324#define TRIE_STUDY_OPT
786e8c11 325#define FULL_TRIE_STUDY
07be1b83
YO
326#define TRIE_STCLASS
327#endif
1de06328
YO
328
329
40d049e4
YO
330
331#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332#define PBITVAL(paren) (1 << ((paren) & 7))
333#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
336
82a6ada4 337#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 338 if (!UTF) { \
82a6ada4
KW
339 assert(PASS1); \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
341 return NULL; \
342 } \
82a6ada4 343 } STMT_END
40d049e4 344
512c0f5a
KW
345/* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
348 * we've change to /u during the parse. */
349#define REQUIRE_UNI_RULES(flagp, restart_retval) \
350 STMT_START { \
351 if (DEPENDS_SEMANTICS) { \
352 assert(PASS1); \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
358 } \
359 } \
360 } STMT_END
361
f19b1a63
KW
362/* This converts the named class defined in regcomp.h to its equivalent class
363 * number defined in handy.h. */
364#define namedclass_to_classnum(class) ((int) ((class) / 2))
365#define classnum_to_namedclass(classnum) ((classnum) * 2)
366
de92f5e6
KW
367#define _invlist_union_complement_2nd(a, b, output) \
368 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
369#define _invlist_intersection_complement_2nd(a, b, output) \
370 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
371
1de06328
YO
372/* About scan_data_t.
373
374 During optimisation we recurse through the regexp program performing
375 various inplace (keyhole style) optimisations. In addition study_chunk
376 and scan_commit populate this data structure with information about
538e84ed 377 what strings MUST appear in the pattern. We look for the longest
3b753521 378 string that must appear at a fixed location, and we look for the
1de06328
YO
379 longest string that may appear at a floating location. So for instance
380 in the pattern:
538e84ed 381
1de06328 382 /FOO[xX]A.*B[xX]BAR/
538e84ed 383
1de06328
YO
384 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
385 strings (because they follow a .* construct). study_chunk will identify
386 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 387
1de06328 388 The strings can be composites, for instance
538e84ed 389
1de06328 390 /(f)(o)(o)/
538e84ed 391
1de06328 392 will result in a composite fixed substring 'foo'.
538e84ed 393
1de06328 394 For each string some basic information is maintained:
538e84ed 395
1de06328
YO
396 - offset or min_offset
397 This is the position the string must appear at, or not before.
398 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
399 characters must match before the string we are searching for.
400 Likewise when combined with minlenp and the length of the string it
538e84ed 401 tells us how many characters must appear after the string we have
1de06328 402 found.
538e84ed 403
1de06328
YO
404 - max_offset
405 Only used for floating strings. This is the rightmost point that
ea3daa5d 406 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 407 string can occur infinitely far to the right.
538e84ed 408
1de06328 409 - minlenp
2d608413
KW
410 A pointer to the minimum number of characters of the pattern that the
411 string was found inside. This is important as in the case of positive
538e84ed 412 lookahead or positive lookbehind we can have multiple patterns
1de06328 413 involved. Consider
538e84ed 414
1de06328 415 /(?=FOO).*F/
538e84ed 416
1de06328
YO
417 The minimum length of the pattern overall is 3, the minimum length
418 of the lookahead part is 3, but the minimum length of the part that
538e84ed 419 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 420 minimum length for the F is 1. This is important as the minimum length
538e84ed 421 is used to determine offsets in front of and behind the string being
1de06328 422 looked for. Since strings can be composites this is the length of the
486ec47a 423 pattern at the time it was committed with a scan_commit. Note that
1de06328 424 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 425 are not known until the full pattern has been compiled, thus the
1de06328 426 pointer to the value.
538e84ed 427
1de06328 428 - lookbehind
538e84ed 429
1de06328 430 In the case of lookbehind the string being searched for can be
538e84ed 431 offset past the start point of the final matching string.
1de06328
YO
432 If this value was just blithely removed from the min_offset it would
433 invalidate some of the calculations for how many chars must match
434 before or after (as they are derived from min_offset and minlen and
538e84ed 435 the length of the string being searched for).
1de06328
YO
436 When the final pattern is compiled and the data is moved from the
437 scan_data_t structure into the regexp structure the information
538e84ed
KW
438 about lookbehind is factored in, with the information that would
439 have been lost precalculated in the end_shift field for the
1de06328
YO
440 associated string.
441
442 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 443 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
444
445*/
2c2d71f5
JH
446
447typedef struct scan_data_t {
1de06328
YO
448 /*I32 len_min; unused */
449 /*I32 len_delta; unused */
49f55535 450 SSize_t pos_min;
ea3daa5d 451 SSize_t pos_delta;
2c2d71f5 452 SV *last_found;
ea3daa5d 453 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 454 SSize_t last_start_min;
ea3daa5d 455 SSize_t last_start_max;
1de06328
YO
456 SV **longest; /* Either &l_fixed, or &l_float. */
457 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 458 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 459 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
460 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
461 SV *longest_float; /* longest floating string found in pattern */
49f55535 462 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
463 SSize_t offset_float_max; /* latest point in string it can appear */
464 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 465 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
466 I32 flags;
467 I32 whilem_c;
49f55535 468 SSize_t *last_closep;
b8f7bb16 469 regnode_ssc *start_class;
2c2d71f5
JH
470} scan_data_t;
471
a687059c 472/*
e50aee73 473 * Forward declarations for pregcomp()'s friends.
a687059c 474 */
a0d0e21e 475
27da23d5 476static const scan_data_t zero_scan_data =
1de06328 477 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
478
479#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
480#define SF_BEFORE_SEOL 0x0001
481#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
482#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
483#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
484
44e3dfd2
BF
485#define SF_FIX_SHIFT_EOL (+2)
486#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
487
488#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
489#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
490
491#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
492#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
493#define SF_IS_INF 0x0040
494#define SF_HAS_PAR 0x0080
495#define SF_IN_PAR 0x0100
496#define SF_HAS_EVAL 0x0200
bc604ad8
KW
497
498
499/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
500 * longest substring in the pattern. When it is not set the optimiser keeps
501 * track of position, but does not keep track of the actual strings seen,
502 *
503 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
504 * /foo/i will not.
505 *
506 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
507 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
508 * turned off because of the alternation (BRANCH). */
07be1b83 509#define SCF_DO_SUBSTR 0x0400
bc604ad8 510
653099ff
GS
511#define SCF_DO_STCLASS_AND 0x0800
512#define SCF_DO_STCLASS_OR 0x1000
513#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 514#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 515
786e8c11 516#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 517#define SCF_SEEN_ACCEPT 0x8000
688e0391 518#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
519#define SCF_IN_DEFINE 0x20000
520
521
522
07be1b83 523
43fead97 524#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
525
526/* The enums for all these are ordered so things work out correctly */
a62b1201 527#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
528#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
529 == REGEX_DEPENDS_CHARSET)
00b27cfc 530#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
531#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
532 >= REGEX_UNICODE_CHARSET)
533#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
534 == REGEX_ASCII_RESTRICTED_CHARSET)
535#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
536 >= REGEX_ASCII_RESTRICTED_CHARSET)
537#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
538 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 539
43fead97 540#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 541
f2c2a6ab
KW
542/* For programs that want to be strictly Unicode compatible by dying if any
543 * attempt is made to match a non-Unicode code point against a Unicode
544 * property. */
545#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
546
93733859 547#define OOB_NAMEDCLASS -1
b8c5462f 548
8e661ac5
KW
549/* There is no code point that is out-of-bounds, so this is problematic. But
550 * its only current use is to initialize a variable that is always set before
551 * looked at. */
552#define OOB_UNICODE 0xDEADBEEF
553
a0ed51b3 554#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 555
8615cb43 556
b45f050a
JF
557/* length of regex to show in messages that don't mark a position within */
558#define RegexLengthToShowInErrorMessages 127
559
560/*
561 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
562 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
563 * op/pragma/warn/regcomp.
564 */
7253e4e3
RK
565#define MARKER1 "<-- HERE" /* marker as it appears in the description */
566#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 567
538e84ed 568#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 569 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 570
285b5ca0
KW
571/* The code in this file in places uses one level of recursion with parsing
572 * rebased to an alternate string constructed by us in memory. This can take
573 * the form of something that is completely different from the input, or
574 * something that uses the input as part of the alternate. In the first case,
575 * there should be no possibility of an error, as we are in complete control of
576 * the alternate string. But in the second case we don't control the input
577 * portion, so there may be errors in that. Here's an example:
578 * /[abc\x{DF}def]/ui
579 * is handled specially because \x{df} folds to a sequence of more than one
580 * character, 'ss'. What is done is to create and parse an alternate string,
581 * which looks like this:
582 * /(?:\x{DF}|[abc\x{DF}def])/ui
583 * where it uses the input unchanged in the middle of something it constructs,
584 * which is a branch for the DF outside the character class, and clustering
585 * parens around the whole thing. (It knows enough to skip the DF inside the
586 * class while in this substitute parse.) 'abc' and 'def' may have errors that
587 * need to be reported. The general situation looks like this:
588 *
589 * sI tI xI eI
590 * Input: ----------------------------------------------------
591 * Constructed: ---------------------------------------------------
592 * sC tC xC eC EC
593 *
594 * The input string sI..eI is the input pattern. The string sC..EC is the
595 * constructed substitute parse string. The portions sC..tC and eC..EC are
596 * constructed by us. The portion tC..eC is an exact duplicate of the input
597 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
598 * while parsing, we find an error at xC. We want to display a message showing
599 * the real input string. Thus we need to find the point xI in it which
600 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
601 * been constructed by us, and so shouldn't have errors. We get:
602 *
603 * xI = sI + (tI - sI) + (xC - tC)
604 *
605 * and, the offset into sI is:
606 *
607 * (xI - sI) = (tI - sI) + (xC - tC)
608 *
609 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
610 * and we save tC as RExC_adjusted_start.
903c858a
KW
611 *
612 * During normal processing of the input pattern, everything points to that,
613 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
285b5ca0
KW
614 */
615
616#define tI_sI RExC_precomp_adj
617#define tC RExC_adjusted_start
618#define sC RExC_precomp
619#define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
620#define xI(xC) (sC + xI_offset(xC))
621#define eC RExC_precomp_end
622
623#define REPORT_LOCATION_ARGS(xC) \
624 UTF8fARG(UTF, \
625 (xI(xC) > eC) /* Don't run off end */ \
626 ? eC - sC /* Length before the <--HERE */ \
627 : xI_offset(xC), \
628 sC), /* The input pattern printed up to the <--HERE */ \
629 UTF8fARG(UTF, \
630 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
631 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
c1d900c3 632
8a6d8ec6
HS
633/* Used to point after bad bytes for an error message, but avoid skipping
634 * past a nul byte. */
635#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
636
b45f050a
JF
637/*
638 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
639 * arg. Show regex, up to a maximum length. If it's too long, chop and add
640 * "...".
641 */
58e23c8d 642#define _FAIL(code) STMT_START { \
bfed75c6 643 const char *ellipses = ""; \
711b303b 644 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
645 \
646 if (!SIZE_ONLY) \
a5e7bc51 647 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
648 if (len > RegexLengthToShowInErrorMessages) { \
649 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
650 len = RegexLengthToShowInErrorMessages - 10; \
651 ellipses = "..."; \
652 } \
58e23c8d 653 code; \
ccb2c380 654} STMT_END
8615cb43 655
58e23c8d 656#define FAIL(msg) _FAIL( \
147e3846 657 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 658 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
659
660#define FAIL2(msg,arg) _FAIL( \
147e3846 661 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 662 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 663
b45f050a 664/*
b45f050a
JF
665 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
666 */
ccb2c380 667#define Simple_vFAIL(m) STMT_START { \
ccb2c380 668 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 669 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 670} STMT_END
b45f050a
JF
671
672/*
673 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
674 */
ccb2c380
MP
675#define vFAIL(m) STMT_START { \
676 if (!SIZE_ONLY) \
a5e7bc51 677 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
678 Simple_vFAIL(m); \
679} STMT_END
b45f050a
JF
680
681/*
682 * Like Simple_vFAIL(), but accepts two arguments.
683 */
ccb2c380 684#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
685 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
686 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 687} STMT_END
b45f050a
JF
688
689/*
690 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
691 */
ccb2c380
MP
692#define vFAIL2(m,a1) STMT_START { \
693 if (!SIZE_ONLY) \
a5e7bc51 694 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
695 Simple_vFAIL2(m, a1); \
696} STMT_END
b45f050a
JF
697
698
699/*
700 * Like Simple_vFAIL(), but accepts three arguments.
701 */
ccb2c380 702#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 703 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 704 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 705} STMT_END
b45f050a
JF
706
707/*
708 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
709 */
ccb2c380
MP
710#define vFAIL3(m,a1,a2) STMT_START { \
711 if (!SIZE_ONLY) \
a5e7bc51 712 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
713 Simple_vFAIL3(m, a1, a2); \
714} STMT_END
b45f050a
JF
715
716/*
717 * Like Simple_vFAIL(), but accepts four arguments.
718 */
ccb2c380 719#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
720 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
721 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 722} STMT_END
b45f050a 723
95db3ffa
KW
724#define vFAIL4(m,a1,a2,a3) STMT_START { \
725 if (!SIZE_ONLY) \
726 SAVEFREESV(RExC_rx_sv); \
727 Simple_vFAIL4(m, a1, a2, a3); \
728} STMT_END
729
946095af 730/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
731#define vFAIL2utf8f(m, a1) STMT_START { \
732 if (!SIZE_ONLY) \
733 SAVEFREESV(RExC_rx_sv); \
734 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
735 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
736} STMT_END
737
3ba22297 738#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
739 if (!SIZE_ONLY) \
740 SAVEFREESV(RExC_rx_sv); \
741 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 742 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
743} STMT_END
744
499333dc
KW
745/* These have asserts in them because of [perl #122671] Many warnings in
746 * regcomp.c can occur twice. If they get output in pass1 and later in that
747 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
748 * would get output again. So they should be output in pass2, and these
749 * asserts make sure new warnings follow that paradigm. */
946095af 750
5e0a247b
KW
751/* m is not necessarily a "literal string", in this macro */
752#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
753 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
754 "%s" REPORT_LOCATION, \
755 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
756} STMT_END
757
668c081a 758#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
759 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
760 m REPORT_LOCATION, \
761 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
762} STMT_END
763
b927b7e9 764#define vWARN(loc, m) STMT_START { \
d528642a
KW
765 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
766 m REPORT_LOCATION, \
767 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
768} STMT_END
769
0d6106aa 770#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
771 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
772 m REPORT_LOCATION, \
773 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
774} STMT_END
775
147508a2 776#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
777 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
778 m REPORT_LOCATION, \
779 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
780} STMT_END
781
d528642a
KW
782#define ckWARNregdep(loc,m) STMT_START { \
783 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
784 WARN_REGEXP), \
785 m REPORT_LOCATION, \
786 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
787} STMT_END
788
d528642a
KW
789#define ckWARN2reg_d(loc,m, a1) STMT_START { \
790 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
791 m REPORT_LOCATION, \
792 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
793} STMT_END
794
d528642a
KW
795#define ckWARN2reg(loc, m, a1) STMT_START { \
796 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
797 m REPORT_LOCATION, \
798 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
799} STMT_END
800
d528642a
KW
801#define vWARN3(loc, m, a1, a2) STMT_START { \
802 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
803 m REPORT_LOCATION, \
804 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
805} STMT_END
806
d528642a
KW
807#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
808 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
809 m REPORT_LOCATION, \
810 a1, a2, \
811 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
812} STMT_END
813
ccb2c380 814#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
815 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
816 m REPORT_LOCATION, \
817 a1, a2, a3, \
818 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
819} STMT_END
820
668c081a 821#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
822 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
823 m REPORT_LOCATION, \
824 a1, a2, a3, \
825 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
826} STMT_END
827
ccb2c380 828#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
829 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
830 m REPORT_LOCATION, \
831 a1, a2, a3, a4, \
832 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 833} STMT_END
9d1d55b5 834
538e84ed 835/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
836 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
837 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
838 * Element 0 holds the number n.
07be1b83 839 * Position is 1 indexed.
fac92740 840 */
7122b237
YO
841#ifndef RE_TRACK_PATTERN_OFFSETS
842#define Set_Node_Offset_To_R(node,byte)
843#define Set_Node_Offset(node,byte)
844#define Set_Cur_Node_Offset
845#define Set_Node_Length_To_R(node,len)
846#define Set_Node_Length(node,len)
6a86c6ad 847#define Set_Node_Cur_Length(node,start)
538e84ed
KW
848#define Node_Offset(n)
849#define Node_Length(n)
7122b237
YO
850#define Set_Node_Offset_Length(node,offset,len)
851#define ProgLen(ri) ri->u.proglen
852#define SetProgLen(ri,x) ri->u.proglen = x
853#else
854#define ProgLen(ri) ri->u.offsets[0]
855#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
856#define Set_Node_Offset_To_R(node,byte) STMT_START { \
857 if (! SIZE_ONLY) { \
858 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 859 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 860 if((node) < 0) { \
538e84ed
KW
861 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
862 (int)(node)); \
ccb2c380
MP
863 } else { \
864 RExC_offsets[2*(node)-1] = (byte); \
865 } \
866 } \
867} STMT_END
868
869#define Set_Node_Offset(node,byte) \
870 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
871#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
872
873#define Set_Node_Length_To_R(node,len) STMT_START { \
874 if (! SIZE_ONLY) { \
875 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 876 __LINE__, (int)(node), (int)(len))); \
ccb2c380 877 if((node) < 0) { \
538e84ed
KW
878 Perl_croak(aTHX_ "value of node is %d in Length macro", \
879 (int)(node)); \
ccb2c380
MP
880 } else { \
881 RExC_offsets[2*(node)] = (len); \
882 } \
883 } \
884} STMT_END
885
886#define Set_Node_Length(node,len) \
887 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
888#define Set_Node_Cur_Length(node, start) \
889 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
890
891/* Get offsets and lengths */
892#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
893#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
894
07be1b83
YO
895#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
896 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
897 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
898} STMT_END
7122b237 899#endif
07be1b83
YO
900
901#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
902#define EXPERIMENTAL_INPLACESCAN
f427392e 903#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 904
cb41e5d6
YO
905#ifdef DEBUGGING
906int
6ad9a8ab 907Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
908{
909 va_list ap;
910 int result;
911 PerlIO *f= Perl_debug_log;
912 PERL_ARGS_ASSERT_RE_PRINTF;
913 va_start(ap, fmt);
914 result = PerlIO_vprintf(f, fmt, ap);
915 va_end(ap);
916 return result;
917}
918
919int
7b031478 920Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
921{
922 va_list ap;
923 int result;
924 PerlIO *f= Perl_debug_log;
925 PERL_ARGS_ASSERT_RE_INDENTF;
926 va_start(ap, depth);
daeb874b 927 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
928 result = PerlIO_vprintf(f, fmt, ap);
929 va_end(ap);
930 return result;
931}
932#endif /* DEBUGGING */
933
7b031478 934#define DEBUG_RExC_seen() \
538e84ed 935 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 936 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 937 \
e384d5c1 938 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 939 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 940 \
e384d5c1 941 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 942 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 943 \
e384d5c1 944 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 945 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 946 \
e384d5c1 947 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 948 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 949 \
7b031478 950 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 951 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 952 \
e384d5c1 953 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 954 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 955 \
e384d5c1 956 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 957 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 958 \
e384d5c1 959 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 960 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 961 \
e384d5c1 962 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 963 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 964 \
7b031478 965 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 966 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 967 \
6ad9a8ab 968 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
969 });
970
fdfb4f21 971#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 972 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21
YO
973
974#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
975 if ( ( flags ) ) { \
6ad9a8ab 976 Perl_re_printf( aTHX_ "%s", open_str); \
fdfb4f21
YO
977 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
978 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
6ad9a8ab 992 Perl_re_printf( aTHX_ "%s", close_str); \
fdfb4f21
YO
993 }
994
995
304ee84b
YO
996#define DEBUG_STUDYDATA(str,data,depth) \
997DEBUG_OPTIMISE_MORE_r(if(data){ \
147e3846
KW
998 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
999 " Flags: 0x%" UVXf, \
7b031478 1000 depth, \
1de06328
YO
1001 (IV)((data)->pos_min), \
1002 (IV)((data)->pos_delta), \
fdfb4f21
YO
1003 (UV)((data)->flags) \
1004 ); \
1005 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
147e3846
KW
1006 Perl_re_printf( aTHX_ \
1007 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1de06328 1008 (IV)((data)->whilem_c), \
304ee84b
YO
1009 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1010 is_inf ? "INF " : "" \
1de06328
YO
1011 ); \
1012 if ((data)->last_found) \
147e3846
KW
1013 Perl_re_printf( aTHX_ \
1014 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1015 " %sFixed:'%s' @ %" IVdf \
1016 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1de06328
YO
1017 SvPVX_const((data)->last_found), \
1018 (IV)((data)->last_end), \
1019 (IV)((data)->last_start_min), \
1020 (IV)((data)->last_start_max), \
1021 ((data)->longest && \
1022 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1023 SvPVX_const((data)->longest_fixed), \
1024 (IV)((data)->offset_fixed), \
1025 ((data)->longest && \
1026 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1027 SvPVX_const((data)->longest_float), \
1028 (IV)((data)->offset_float_min), \
1029 (IV)((data)->offset_float_max) \
1030 ); \
6ad9a8ab 1031 Perl_re_printf( aTHX_ "\n"); \
1de06328
YO
1032});
1033
cb41e5d6 1034
c6871b76
KW
1035/* =========================================================
1036 * BEGIN edit_distance stuff.
1037 *
1038 * This calculates how many single character changes of any type are needed to
1039 * transform a string into another one. It is taken from version 3.1 of
1040 *
1041 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1042 */
1043
1044/* Our unsorted dictionary linked list. */
1045/* Note we use UVs, not chars. */
1046
1047struct dictionary{
1048 UV key;
1049 UV value;
1050 struct dictionary* next;
1051};
1052typedef struct dictionary item;
1053
1054
1055PERL_STATIC_INLINE item*
1056push(UV key,item* curr)
1057{
1058 item* head;
1059 Newxz(head, 1, item);
1060 head->key = key;
1061 head->value = 0;
1062 head->next = curr;
1063 return head;
1064}
1065
1066
1067PERL_STATIC_INLINE item*
1068find(item* head, UV key)
1069{
1070 item* iterator = head;
1071 while (iterator){
1072 if (iterator->key == key){
1073 return iterator;
1074 }
1075 iterator = iterator->next;
1076 }
1077
1078 return NULL;
1079}
1080
1081PERL_STATIC_INLINE item*
1082uniquePush(item* head,UV key)
1083{
1084 item* iterator = head;
1085
1086 while (iterator){
1087 if (iterator->key == key) {
1088 return head;
1089 }
1090 iterator = iterator->next;
1091 }
1092
1093 return push(key,head);
1094}
1095
1096PERL_STATIC_INLINE void
1097dict_free(item* head)
1098{
1099 item* iterator = head;
1100
1101 while (iterator) {
1102 item* temp = iterator;
1103 iterator = iterator->next;
1104 Safefree(temp);
1105 }
1106
1107 head = NULL;
1108}
1109
1110/* End of Dictionary Stuff */
1111
1112/* All calculations/work are done here */
1113STATIC int
1114S_edit_distance(const UV* src,
1115 const UV* tgt,
1116 const STRLEN x, /* length of src[] */
1117 const STRLEN y, /* length of tgt[] */
1118 const SSize_t maxDistance
1119)
1120{
1121 item *head = NULL;
1122 UV swapCount,swapScore,targetCharCount,i,j;
1123 UV *scores;
1124 UV score_ceil = x + y;
1125
1126 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1127
1128 /* intialize matrix start values */
1129 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1130 scores[0] = score_ceil;
1131 scores[1 * (y + 2) + 0] = score_ceil;
1132 scores[0 * (y + 2) + 1] = score_ceil;
1133 scores[1 * (y + 2) + 1] = 0;
1134 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1135
1136 /* work loops */
1137 /* i = src index */
1138 /* j = tgt index */
1139 for (i=1;i<=x;i++) {
1140 if (i < x)
1141 head = uniquePush(head,src[i]);
1142 scores[(i+1) * (y + 2) + 1] = i;
1143 scores[(i+1) * (y + 2) + 0] = score_ceil;
1144 swapCount = 0;
1145
1146 for (j=1;j<=y;j++) {
1147 if (i == 1) {
1148 if(j < y)
1149 head = uniquePush(head,tgt[j]);
1150 scores[1 * (y + 2) + (j + 1)] = j;
1151 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1152 }
1153
1154 targetCharCount = find(head,tgt[j-1])->value;
1155 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1156
1157 if (src[i-1] != tgt[j-1]){
1158 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));
1159 }
1160 else {
1161 swapCount = j;
1162 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1163 }
1164 }
1165
1166 find(head,src[i-1])->value = i;
1167 }
1168
1169 {
1170 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1171 dict_free(head);
1172 Safefree(scores);
1173 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1174 }
1175}
1176
1177/* END of edit_distance() stuff
1178 * ========================================================= */
1179
8e35b056
KW
1180/* is c a control character for which we have a mnemonic? */
1181#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1182
549b4e78
KW
1183STATIC const char *
1184S_cntrl_to_mnemonic(const U8 c)
1185{
1186 /* Returns the mnemonic string that represents character 'c', if one
1187 * exists; NULL otherwise. The only ones that exist for the purposes of
1188 * this routine are a few control characters */
1189
1190 switch (c) {
1191 case '\a': return "\\a";
1192 case '\b': return "\\b";
1193 case ESC_NATIVE: return "\\e";
1194 case '\f': return "\\f";
1195 case '\n': return "\\n";
1196 case '\r': return "\\r";
1197 case '\t': return "\\t";
1198 }
1199
1200 return NULL;
1201}
1202
653099ff 1203/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 1204 Update the longest found anchored substring and the longest found
653099ff
GS
1205 floating substrings if needed. */
1206
4327152a 1207STATIC void
ea3daa5d
FC
1208S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1209 SSize_t *minlenp, int is_inf)
c277df42 1210{
e1ec3a88
AL
1211 const STRLEN l = CHR_SVLEN(data->last_found);
1212 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 1213 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1214
7918f24d
NC
1215 PERL_ARGS_ASSERT_SCAN_COMMIT;
1216
c277df42 1217 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 1218 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
1219 if (*data->longest == data->longest_fixed) {
1220 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1221 if (data->flags & SF_BEFORE_EOL)
b81d288d 1222 data->flags
c277df42
IZ
1223 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1224 else
1225 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 1226 data->minlen_fixed=minlenp;
1de06328 1227 data->lookbehind_fixed=0;
a0ed51b3 1228 }
304ee84b 1229 else { /* *data->longest == data->longest_float */
c277df42 1230 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 1231 data->offset_float_max = (l
646e8787
DM
1232 ? data->last_start_max
1233 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1234 ? SSize_t_MAX
1235 : data->pos_min + data->pos_delta));
1236 if (is_inf
1237 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1238 data->offset_float_max = SSize_t_MAX;
c277df42 1239 if (data->flags & SF_BEFORE_EOL)
b81d288d 1240 data->flags
c277df42
IZ
1241 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1242 else
1243 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
1244 data->minlen_float=minlenp;
1245 data->lookbehind_float=0;
c277df42
IZ
1246 }
1247 }
1248 SvCUR_set(data->last_found, 0);
0eda9292 1249 {
a28509cc 1250 SV * const sv = data->last_found;
097eb12c
AL
1251 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1252 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1253 if (mg)
1254 mg->mg_len = 0;
1255 }
0eda9292 1256 }
c277df42
IZ
1257 data->last_end = -1;
1258 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1259 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1260}
1261
cdd87c1d
KW
1262/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1263 * list that describes which code points it matches */
1264
653099ff 1265STATIC void
3420edd7 1266S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1267{
cdd87c1d
KW
1268 /* Set the SSC 'ssc' to match an empty string or any code point */
1269
557bd3fb 1270 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1271
71068078 1272 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1273
0854ea0b
KW
1274 /* mortalize so won't leak */
1275 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1276 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1277}
1278
653099ff 1279STATIC int
dc3bf405 1280S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1281{
c144baaa
KW
1282 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1283 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1284 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1285 * in any way, so there's no point in using it */
cdd87c1d
KW
1286
1287 UV start, end;
1288 bool ret;
653099ff 1289
557bd3fb 1290 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1291
71068078 1292 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1293
93e92956 1294 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1295 return FALSE;
1296 }
1297
1298 /* See if the list consists solely of the range 0 - Infinity */
1299 invlist_iterinit(ssc->invlist);
1300 ret = invlist_iternext(ssc->invlist, &start, &end)
1301 && start == 0
1302 && end == UV_MAX;
1303
1304 invlist_iterfinish(ssc->invlist);
1305
1306 if (ret) {
1307 return TRUE;
1308 }
1309
1310 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1311 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1312 int i;
1313 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1314 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1315 return TRUE;
1316 }
1317 }
1318 }
1319
1320 return FALSE;
653099ff
GS
1321}
1322
653099ff 1323STATIC void
cdd87c1d 1324S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1325{
cdd87c1d
KW
1326 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1327 * string, any code point, or any posix class under locale */
1328
557bd3fb 1329 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1330
557bd3fb 1331 Zero(ssc, 1, regnode_ssc);
71068078 1332 set_ANYOF_SYNTHETIC(ssc);
93e92956 1333 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1334 ssc_anything(ssc);
cdd87c1d 1335
2f306ab9
KW
1336 /* If any portion of the regex is to operate under locale rules that aren't
1337 * fully known at compile time, initialization includes it. The reason
1338 * this isn't done for all regexes is that the optimizer was written under
1339 * the assumption that locale was all-or-nothing. Given the complexity and
1340 * lack of documentation in the optimizer, and that there are inadequate
1341 * test cases for locale, many parts of it may not work properly, it is
1342 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1343 if (RExC_contains_locale) {
1344 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1345 }
1346 else {
1347 ANYOF_POSIXL_ZERO(ssc);
1348 }
653099ff
GS
1349}
1350
b423522f 1351STATIC int
dc3bf405
BF
1352S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1353 const regnode_ssc *ssc)
b423522f
KW
1354{
1355 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1356 * to the list of code points matched, and locale posix classes; hence does
1357 * not check its flags) */
1358
1359 UV start, end;
1360 bool ret;
1361
1362 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1363
71068078 1364 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1365
1366 invlist_iterinit(ssc->invlist);
1367 ret = invlist_iternext(ssc->invlist, &start, &end)
1368 && start == 0
1369 && end == UV_MAX;
1370
1371 invlist_iterfinish(ssc->invlist);
1372
1373 if (! ret) {
1374 return FALSE;
1375 }
1376
e0e1be5f 1377 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1378 return FALSE;
b423522f
KW
1379 }
1380
1381 return TRUE;
1382}
1383
1384STATIC SV*
1385S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1386 const regnode_charclass* const node)
b423522f
KW
1387{
1388 /* Returns a mortal inversion list defining which code points are matched
1389 * by 'node', which is of type ANYOF. Handles complementing the result if
1390 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1391 * returned list must, and will, contain every code point that is a
1392 * possibility. */
b423522f 1393
e2506fa7 1394 SV* invlist = NULL;
1ee208c4 1395 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1396 unsigned int i;
1397 const U32 n = ARG(node);
31f05a37 1398 bool new_node_has_latin1 = FALSE;
b423522f
KW
1399
1400 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1401
1402 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1403 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1404 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1405 AV * const av = MUTABLE_AV(SvRV(rv));
1406 SV **const ary = AvARRAY(av);
1407 assert(RExC_rxi->data->what[n] == 's');
1408
1409 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1410 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1411 }
1412 else if (ary[0] && ary[0] != &PL_sv_undef) {
1413
1414 /* Here, no compile-time swash, and there are things that won't be
1415 * known until runtime -- we have to assume it could be anything */
e2506fa7 1416 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1417 return _add_range_to_invlist(invlist, 0, UV_MAX);
1418 }
1ee208c4 1419 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1420
1421 /* Here no compile-time swash, and no run-time only data. Use the
1422 * node's inversion list */
1ee208c4
KW
1423 invlist = sv_2mortal(invlist_clone(ary[3]));
1424 }
1425
1426 /* Get the code points valid only under UTF-8 locales */
037715a6 1427 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1428 && ary[2] && ary[2] != &PL_sv_undef)
1429 {
1430 only_utf8_locale_invlist = ary[2];
b423522f
KW
1431 }
1432 }
1433
e2506fa7
KW
1434 if (! invlist) {
1435 invlist = sv_2mortal(_new_invlist(0));
1436 }
1437
dcb20b36
KW
1438 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1439 * code points, and an inversion list for the others, but if there are code
1440 * points that should match only conditionally on the target string being
1441 * UTF-8, those are placed in the inversion list, and not the bitmap.
1442 * Since there are circumstances under which they could match, they are
1443 * included in the SSC. But if the ANYOF node is to be inverted, we have
1444 * to exclude them here, so that when we invert below, the end result
1445 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1446 * have to do this here before we add the unconditionally matched code
1447 * points */
b423522f
KW
1448 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1449 _invlist_intersection_complement_2nd(invlist,
1450 PL_UpperLatin1,
1451 &invlist);
1452 }
1453
1454 /* Add in the points from the bit map */
dcb20b36 1455 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1456 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1457 unsigned int start = i++;
1458
1459 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1460 /* empty */
1461 }
1462 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1463 new_node_has_latin1 = TRUE;
b423522f
KW
1464 }
1465 }
1466
1467 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1468 * as well. But don't add them if inverting, as when that gets done below,
1469 * it would exclude all these characters, including the ones it shouldn't
1470 * that were added just above */
1471 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1472 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1473 {
b423522f
KW
1474 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1475 }
1476
1477 /* Similarly for these */
93e92956 1478 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1479 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1480 }
1481
1482 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1483 _invlist_invert(invlist);
1484 }
037715a6 1485 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1486
1487 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1488 * locale. We can skip this if there are no 0-255 at all. */
1489 _invlist_union(invlist, PL_Latin1, &invlist);
1490 }
1491
1ee208c4
KW
1492 /* Similarly add the UTF-8 locale possible matches. These have to be
1493 * deferred until after the non-UTF-8 locale ones are taken care of just
1494 * above, or it leads to wrong results under ANYOF_INVERT */
1495 if (only_utf8_locale_invlist) {
31f05a37 1496 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1497 only_utf8_locale_invlist,
31f05a37
KW
1498 ANYOF_FLAGS(node) & ANYOF_INVERT,
1499 &invlist);
1500 }
b423522f
KW
1501
1502 return invlist;
1503}
1504
1051e1c4 1505/* These two functions currently do the exact same thing */
557bd3fb 1506#define ssc_init_zero ssc_init
653099ff 1507
cdd87c1d
KW
1508#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1509#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1510
557bd3fb 1511/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1512 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1513 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1514
653099ff 1515STATIC void
b423522f 1516S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1517 const regnode_charclass *and_with)
653099ff 1518{
cdd87c1d
KW
1519 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1520 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1521
a0dd4231
KW
1522 SV* anded_cp_list;
1523 U8 anded_flags;
1e6ade67 1524
cdd87c1d 1525 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1526
71068078 1527 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1528
1529 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1530 * the code point inversion list and just the relevant flags */
71068078 1531 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1532 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1533 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1534
1535 /* XXX This is a kludge around what appears to be deficiencies in the
1536 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1537 * there are paths through the optimizer where it doesn't get weeded
1538 * out when it should. And if we don't make some extra provision for
1539 * it like the code just below, it doesn't get added when it should.
1540 * This solution is to add it only when AND'ing, which is here, and
1541 * only when what is being AND'ed is the pristine, original node
1542 * matching anything. Thus it is like adding it to ssc_anything() but
1543 * only when the result is to be AND'ed. Probably the same solution
1544 * could be adopted for the same problem we have with /l matching,
1545 * which is solved differently in S_ssc_init(), and that would lead to
1546 * fewer false positives than that solution has. But if this solution
1547 * creates bugs, the consequences are only that a warning isn't raised
1548 * that should be; while the consequences for having /l bugs is
1549 * incorrect matches */
7dcac5f6 1550 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1551 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1552 }
a0dd4231
KW
1553 }
1554 else {
5c0f85ef 1555 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1556 if (OP(and_with) == ANYOFD) {
1557 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1558 }
1559 else {
1560 anded_flags = ANYOF_FLAGS(and_with)
1561 &( ANYOF_COMMON_FLAGS
108316fb
KW
1562 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1563 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1564 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1565 anded_flags &=
1566 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1567 }
f240c685 1568 }
a0dd4231
KW
1569 }
1570
1571 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1572
1573 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1574 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1575 * 'and_with' may be inverted. When not inverted, we have the situation of
1576 * computing:
1577 * (C1 | P1) & (C2 | P2)
1578 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1579 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1580 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1581 * <= ((C1 & C2) | P1 | P2)
1582 * Alternatively, the last few steps could be:
1583 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1584 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1585 * <= (C1 | C2 | (P1 & P2))
1586 * We favor the second approach if either P1 or P2 is non-empty. This is
1587 * because these components are a barrier to doing optimizations, as what
1588 * they match cannot be known until the moment of matching as they are
1589 * dependent on the current locale, 'AND"ing them likely will reduce or
1590 * eliminate them.
1591 * But we can do better if we know that C1,P1 are in their initial state (a
1592 * frequent occurrence), each matching everything:
1593 * (<everything>) & (C2 | P2) = C2 | P2
1594 * Similarly, if C2,P2 are in their initial state (again a frequent
1595 * occurrence), the result is a no-op
1596 * (C1 | P1) & (<everything>) = C1 | P1
1597 *
1598 * Inverted, we have
1599 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1600 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1601 * <= (C1 & ~C2) | (P1 & ~P2)
1602 * */
1aa99e6b 1603
a0dd4231 1604 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1605 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1606 {
cdd87c1d 1607 unsigned int i;
8951c461 1608
cdd87c1d
KW
1609 ssc_intersection(ssc,
1610 anded_cp_list,
1611 FALSE /* Has already been inverted */
1612 );
c6b76537 1613
cdd87c1d
KW
1614 /* If either P1 or P2 is empty, the intersection will be also; can skip
1615 * the loop */
93e92956 1616 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1617 ANYOF_POSIXL_ZERO(ssc);
1618 }
e0e1be5f 1619 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1620
1621 /* Note that the Posix class component P from 'and_with' actually
1622 * looks like:
1623 * P = Pa | Pb | ... | Pn
1624 * where each component is one posix class, such as in [\w\s].
1625 * Thus
1626 * ~P = ~(Pa | Pb | ... | Pn)
1627 * = ~Pa & ~Pb & ... & ~Pn
1628 * <= ~Pa | ~Pb | ... | ~Pn
1629 * The last is something we can easily calculate, but unfortunately
1630 * is likely to have many false positives. We could do better
1631 * in some (but certainly not all) instances if two classes in
1632 * P have known relationships. For example
1633 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1634 * So
1635 * :lower: & :print: = :lower:
1636 * And similarly for classes that must be disjoint. For example,
1637 * since \s and \w can have no elements in common based on rules in
1638 * the POSIX standard,
1639 * \w & ^\S = nothing
1640 * Unfortunately, some vendor locales do not meet the Posix
1641 * standard, in particular almost everything by Microsoft.
1642 * The loop below just changes e.g., \w into \W and vice versa */
1643
1ee208c4 1644 regnode_charclass_posixl temp;
cdd87c1d
KW
1645 int add = 1; /* To calculate the index of the complement */
1646
1647 ANYOF_POSIXL_ZERO(&temp);
1648 for (i = 0; i < ANYOF_MAX; i++) {
1649 assert(i % 2 != 0
7dcac5f6
KW
1650 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1651 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1652
7dcac5f6 1653 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1654 ANYOF_POSIXL_SET(&temp, i + add);
1655 }
1656 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1657 }
1658 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1659
cdd87c1d
KW
1660 } /* else ssc already has no posixes */
1661 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1662 in its initial state */
71068078 1663 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1664 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1665 {
1666 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1667 * copy it over 'ssc' */
1668 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1669 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1670 StructCopy(and_with, ssc, regnode_ssc);
1671 }
1672 else {
1673 ssc->invlist = anded_cp_list;
1674 ANYOF_POSIXL_ZERO(ssc);
93e92956 1675 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1676 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1677 }
1678 }
1679 }
e0e1be5f 1680 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1681 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1682 {
1683 /* One or the other of P1, P2 is non-empty. */
93e92956 1684 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1685 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1686 }
cdd87c1d
KW
1687 ssc_union(ssc, anded_cp_list, FALSE);
1688 }
1689 else { /* P1 = P2 = empty */
1690 ssc_intersection(ssc, anded_cp_list, FALSE);
1691 }
137165a6 1692 }
653099ff
GS
1693}
1694
653099ff 1695STATIC void
cdd87c1d 1696S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1697 const regnode_charclass *or_with)
653099ff 1698{
cdd87c1d
KW
1699 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1700 * another SSC or a regular ANYOF class. Can create false positives if
1701 * 'or_with' is to be inverted. */
7918f24d 1702
a0dd4231
KW
1703 SV* ored_cp_list;
1704 U8 ored_flags;
c6b76537 1705
cdd87c1d 1706 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1707
71068078 1708 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1709
1710 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1711 * the code point inversion list and just the relevant flags */
71068078 1712 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1713 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1714 ored_flags = ANYOF_FLAGS(or_with);
1715 }
1716 else {
5c0f85ef 1717 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1718 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1719 if (OP(or_with) != ANYOFD) {
1720 ored_flags
1721 |= ANYOF_FLAGS(or_with)
108316fb
KW
1722 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1723 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1724 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1725 ored_flags |=
1726 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1727 }
f240c685 1728 }
a0dd4231
KW
1729 }
1730
1731 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1732
1733 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1734 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1735 * 'or_with' may be inverted. When not inverted, we have the simple
1736 * situation of computing:
1737 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1738 * If P1|P2 yields a situation with both a class and its complement are
1739 * set, like having both \w and \W, this matches all code points, and we
1740 * can delete these from the P component of the ssc going forward. XXX We
1741 * might be able to delete all the P components, but I (khw) am not certain
1742 * about this, and it is better to be safe.
1743 *
1744 * Inverted, we have
1745 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1746 * <= (C1 | P1) | ~C2
1747 * <= (C1 | ~C2) | P1
1748 * (which results in actually simpler code than the non-inverted case)
1749 * */
9826f543 1750
a0dd4231 1751 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1752 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1753 {
cdd87c1d 1754 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1755 } /* else Not inverted */
93e92956 1756 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1757 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1758 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1759 unsigned int i;
1760 for (i = 0; i < ANYOF_MAX; i += 2) {
1761 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1762 {
1763 ssc_match_all_cp(ssc);
1764 ANYOF_POSIXL_CLEAR(ssc, i);
1765 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1766 }
1767 }
1768 }
1aa99e6b 1769 }
cdd87c1d
KW
1770
1771 ssc_union(ssc,
1772 ored_cp_list,
1773 FALSE /* Already has been inverted */
1774 );
653099ff
GS
1775}
1776
b423522f
KW
1777PERL_STATIC_INLINE void
1778S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1779{
1780 PERL_ARGS_ASSERT_SSC_UNION;
1781
71068078 1782 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1783
1784 _invlist_union_maybe_complement_2nd(ssc->invlist,
1785 invlist,
1786 invert2nd,
1787 &ssc->invlist);
1788}
1789
1790PERL_STATIC_INLINE void
1791S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1792 SV* const invlist,
1793 const bool invert2nd)
1794{
1795 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1796
71068078 1797 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1798
1799 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1800 invlist,
1801 invert2nd,
1802 &ssc->invlist);
1803}
1804
1805PERL_STATIC_INLINE void
1806S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1807{
1808 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1809
71068078 1810 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1811
1812 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1813}
1814
1815PERL_STATIC_INLINE void
1816S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1817{
1818 /* AND just the single code point 'cp' into the SSC 'ssc' */
1819
1820 SV* cp_list = _new_invlist(2);
1821
1822 PERL_ARGS_ASSERT_SSC_CP_AND;
1823
71068078 1824 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1825
1826 cp_list = add_cp_to_invlist(cp_list, cp);
1827 ssc_intersection(ssc, cp_list,
1828 FALSE /* Not inverted */
1829 );
1830 SvREFCNT_dec_NN(cp_list);
1831}
1832
1833PERL_STATIC_INLINE void
dc3bf405 1834S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1835{
1836 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1837 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1838
71068078 1839 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1840
1841 ANYOF_POSIXL_ZERO(ssc);
1842 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1843}
1844
b35552de
KW
1845#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1846
1847STATIC bool
1848S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1849{
1850 /* The synthetic start class is used to hopefully quickly winnow down
1851 * places where a pattern could start a match in the target string. If it
1852 * doesn't really narrow things down that much, there isn't much point to
1853 * having the overhead of using it. This function uses some very crude
1854 * heuristics to decide if to use the ssc or not.
1855 *
1856 * It returns TRUE if 'ssc' rules out more than half what it considers to
1857 * be the "likely" possible matches, but of course it doesn't know what the
1858 * actual things being matched are going to be; these are only guesses
1859 *
1860 * For /l matches, it assumes that the only likely matches are going to be
1861 * in the 0-255 range, uniformly distributed, so half of that is 127
1862 * For /a and /d matches, it assumes that the likely matches will be just
1863 * the ASCII range, so half of that is 63
1864 * For /u and there isn't anything matching above the Latin1 range, it
1865 * assumes that that is the only range likely to be matched, and uses
1866 * half that as the cut-off: 127. If anything matches above Latin1,
1867 * it assumes that all of Unicode could match (uniformly), except for
1868 * non-Unicode code points and things in the General Category "Other"
1869 * (unassigned, private use, surrogates, controls and formats). This
1870 * is a much large number. */
1871
b35552de
KW
1872 U32 count = 0; /* Running total of number of code points matched by
1873 'ssc' */
1874 UV start, end; /* Start and end points of current range in inversion
1875 list */
72400949
KW
1876 const U32 max_code_points = (LOC)
1877 ? 256
1878 : (( ! UNI_SEMANTICS
1879 || invlist_highest(ssc->invlist) < 256)
1880 ? 128
1881 : NON_OTHER_COUNT);
1882 const U32 max_match = max_code_points / 2;
b35552de
KW
1883
1884 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1885
1886 invlist_iterinit(ssc->invlist);
1887 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1888 if (start >= max_code_points) {
1889 break;
b35552de 1890 }
72400949 1891 end = MIN(end, max_code_points - 1);
b35552de 1892 count += end - start + 1;
72400949 1893 if (count >= max_match) {
b35552de
KW
1894 invlist_iterfinish(ssc->invlist);
1895 return FALSE;
1896 }
1897 }
1898
1899 return TRUE;
1900}
1901
1902
b423522f
KW
1903STATIC void
1904S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1905{
1906 /* The inversion list in the SSC is marked mortal; now we need a more
1907 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1908 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1909 * map */
b423522f
KW
1910
1911 SV* invlist = invlist_clone(ssc->invlist);
1912
1913 PERL_ARGS_ASSERT_SSC_FINALIZE;
1914
71068078 1915 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1916
a0dd4231 1917 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1918 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1919 * by the time we reach here */
f240c685
KW
1920 assert(! (ANYOF_FLAGS(ssc)
1921 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
1922 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1923 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 1924
b423522f
KW
1925 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1926
1ee208c4
KW
1927 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1928 NULL, NULL, NULL, FALSE);
b423522f 1929
85c8e306
KW
1930 /* Make sure is clone-safe */
1931 ssc->invlist = NULL;
1932
e0e1be5f 1933 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1934 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1935 }
1462525b 1936
b2e90ddf
KW
1937 if (RExC_contains_locale) {
1938 OP(ssc) = ANYOFL;
1939 }
1940
1462525b 1941 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1942}
1943
a3621e74
YO
1944#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1945#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1946#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1947#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1948 ? (TRIE_LIST_CUR( idx ) - 1) \
1949 : 0 )
a3621e74 1950
3dab1dad
YO
1951
1952#ifdef DEBUGGING
07be1b83 1953/*
2b8b4781
NC
1954 dump_trie(trie,widecharmap,revcharmap)
1955 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1956 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1957
1958 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1959 The _interim_ variants are used for debugging the interim
1960 tables that are used to generate the final compressed
1961 representation which is what dump_trie expects.
1962
486ec47a 1963 Part of the reason for their existence is to provide a form
3dab1dad 1964 of documentation as to how the different representations function.
07be1b83
YO
1965
1966*/
3dab1dad
YO
1967
1968/*
3dab1dad
YO
1969 Dumps the final compressed table form of the trie to Perl_debug_log.
1970 Used for debugging make_trie().
1971*/
b9a59e08 1972
3dab1dad 1973STATIC void
2b8b4781
NC
1974S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1975 AV *revcharmap, U32 depth)
3dab1dad
YO
1976{
1977 U32 state;
ab3bbdeb 1978 SV *sv=sv_newmortal();
55eed653 1979 int colwidth= widecharmap ? 6 : 4;
2e64971a 1980 U16 word;
3dab1dad
YO
1981 GET_RE_DEBUG_FLAGS_DECL;
1982
7918f24d 1983 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1984
6ad9a8ab 1985 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 1986 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
1987
1988 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1989 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1990 if ( tmp ) {
6ad9a8ab 1991 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 1992 colwidth,
538e84ed 1993 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1994 PL_colors[0], PL_colors[1],
1995 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1996 PERL_PV_ESCAPE_FIRSTCHAR
1997 )
ab3bbdeb 1998 );
3dab1dad
YO
1999 }
2000 }
1e37780e
YO
2001 Perl_re_printf( aTHX_ "\n");
2002 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2003
2004 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2005 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2006 Perl_re_printf( aTHX_ "\n");
3dab1dad 2007
1e2e3d02 2008 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2009 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2010
147e3846 2011 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2012
2013 if ( trie->states[ state ].wordnum ) {
1e37780e 2014 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2015 } else {
6ad9a8ab 2016 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2017 }
2018
147e3846 2019 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2020
2021 if ( base ) {
2022 U32 ofs = 0;
2023
2024 while( ( base + ofs < trie->uniquecharcount ) ||
2025 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2026 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2027 != state))
3dab1dad
YO
2028 ofs++;
2029
147e3846 2030 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2031
2032 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2033 if ( ( base + ofs >= trie->uniquecharcount )
2034 && ( base + ofs - trie->uniquecharcount
2035 < trie->lasttrans )
2036 && trie->trans[ base + ofs
2037 - trie->uniquecharcount ].check == state )
3dab1dad 2038 {
147e3846 2039 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2040 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2041 );
3dab1dad 2042 } else {
6ad9a8ab 2043 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2044 }
2045 }
2046
6ad9a8ab 2047 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2048
2049 }
6ad9a8ab 2050 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2051 }
6ad9a8ab 2052 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2053 depth);
2e64971a 2054 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2055 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2056 (int)word, (int)(trie->wordinfo[word].prev),
2057 (int)(trie->wordinfo[word].len));
2058 }
6ad9a8ab 2059 Perl_re_printf( aTHX_ "\n" );
538e84ed 2060}
3dab1dad 2061/*
3dab1dad 2062 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2063 List tries normally only are used for construction when the number of
3dab1dad
YO
2064 possible chars (trie->uniquecharcount) is very high.
2065 Used for debugging make_trie().
2066*/
2067STATIC void
55eed653 2068S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2069 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2070 U32 depth)
3dab1dad
YO
2071{
2072 U32 state;
ab3bbdeb 2073 SV *sv=sv_newmortal();
55eed653 2074 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2075 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2076
2077 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2078
3dab1dad 2079 /* print out the table precompression. */
6ad9a8ab 2080 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2081 depth+1 );
6ad9a8ab 2082 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2083 depth+1, "------:-----+-----------------\n" );
538e84ed 2084
3dab1dad
YO
2085 for( state=1 ; state < next_alloc ; state ++ ) {
2086 U16 charid;
538e84ed 2087
147e3846 2088 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2089 depth+1, (UV)state );
3dab1dad 2090 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2091 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2092 } else {
6ad9a8ab 2093 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2094 trie->states[ state ].wordnum
2095 );
2096 }
2097 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2098 SV ** const tmp = av_fetch( revcharmap,
2099 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2100 if ( tmp ) {
147e3846 2101 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2102 colwidth,
538e84ed
KW
2103 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2104 colwidth,
2105 PL_colors[0], PL_colors[1],
2106 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2107 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2108 ) ,
1e2e3d02
YO
2109 TRIE_LIST_ITEM(state,charid).forid,
2110 (UV)TRIE_LIST_ITEM(state,charid).newstate
2111 );
538e84ed 2112 if (!(charid % 10))
6ad9a8ab 2113 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2114 (int)((depth * 2) + 14), "");
1e2e3d02 2115 }
ab3bbdeb 2116 }
6ad9a8ab 2117 Perl_re_printf( aTHX_ "\n");
3dab1dad 2118 }
538e84ed 2119}
3dab1dad
YO
2120
2121/*
3dab1dad 2122 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2123 This is the normal DFA style state transition table, with a few
2124 twists to facilitate compression later.
3dab1dad
YO
2125 Used for debugging make_trie().
2126*/
2127STATIC void
55eed653 2128S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2129 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2130 U32 depth)
3dab1dad
YO
2131{
2132 U32 state;
2133 U16 charid;
ab3bbdeb 2134 SV *sv=sv_newmortal();
55eed653 2135 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2136 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2137
2138 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2139
3dab1dad
YO
2140 /*
2141 print out the table precompression so that we can do a visual check
2142 that they are identical.
2143 */
538e84ed 2144
6ad9a8ab 2145 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2146
2147 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2148 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2149 if ( tmp ) {
6ad9a8ab 2150 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2151 colwidth,
538e84ed 2152 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2153 PL_colors[0], PL_colors[1],
2154 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2155 PERL_PV_ESCAPE_FIRSTCHAR
2156 )
ab3bbdeb 2157 );
3dab1dad
YO
2158 }
2159 }
2160
4aaafc03
YO
2161 Perl_re_printf( aTHX_ "\n");
2162 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2163
2164 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2165 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2166 }
2167
6ad9a8ab 2168 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2169
2170 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2171
147e3846 2172 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2173 depth+1,
3dab1dad
YO
2174 (UV)TRIE_NODENUM( state ) );
2175
2176 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2177 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2178 if (v)
147e3846 2179 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2180 else
6ad9a8ab 2181 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2182 }
2183 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2184 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2185 (UV)trie->trans[ state ].check );
3dab1dad 2186 } else {
147e3846 2187 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2188 (UV)trie->trans[ state ].check,
3dab1dad
YO
2189 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2190 }
2191 }
07be1b83 2192}
3dab1dad
YO
2193
2194#endif
2195
2e64971a 2196
786e8c11
YO
2197/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2198 startbranch: the first branch in the whole branch sequence
2199 first : start branch of sequence of branch-exact nodes.
2200 May be the same as startbranch
2201 last : Thing following the last branch.
2202 May be the same as tail.
2203 tail : item following the branch sequence
2204 count : words in the sequence
a4525e78 2205 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2206 depth : indent depth
3dab1dad 2207
786e8c11 2208Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2209
786e8c11
YO
2210A trie is an N'ary tree where the branches are determined by digital
2211decomposition of the key. IE, at the root node you look up the 1st character and
2212follow that branch repeat until you find the end of the branches. Nodes can be
2213marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2214
786e8c11 2215 /he|she|his|hers/
72f13be8 2216
786e8c11
YO
2217would convert into the following structure. Numbers represent states, letters
2218following numbers represent valid transitions on the letter from that state, if
2219the number is in square brackets it represents an accepting state, otherwise it
2220will be in parenthesis.
07be1b83 2221
786e8c11
YO
2222 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2223 | |
2224 | (2)
2225 | |
2226 (1) +-i->(6)-+-s->[7]
2227 |
2228 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2229
786e8c11
YO
2230 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2231
2232This shows that when matching against the string 'hers' we will begin at state 1
2233read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2234then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2235is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2236single traverse. We store a mapping from accepting to state to which word was
2237matched, and then when we have multiple possibilities we try to complete the
b8fda935 2238rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2239
2240The only prior NFA like behaviour that would be changed by the TRIE support is
2241the silent ignoring of duplicate alternations which are of the form:
2242
2243 / (DUPE|DUPE) X? (?{ ... }) Y /x
2244
4b714af6 2245Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2246and without the optimisation. With the optimisations dupes will be silently
486ec47a 2247ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2248the following demonstrates:
2249
2250 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2251
2252which prints out 'word' three times, but
2253
2254 'words'=~/(word|word|word)(?{ print $1 })S/
2255
2256which doesnt print it out at all. This is due to other optimisations kicking in.
2257
2258Example of what happens on a structural level:
2259
486ec47a 2260The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2261
2262 1: CURLYM[1] {1,32767}(18)
2263 5: BRANCH(8)
2264 6: EXACT <ac>(16)
2265 8: BRANCH(11)
2266 9: EXACT <ad>(16)
2267 11: BRANCH(14)
2268 12: EXACT <ab>(16)
2269 16: SUCCEED(0)
2270 17: NOTHING(18)
2271 18: END(0)
2272
2273This would be optimizable with startbranch=5, first=5, last=16, tail=16
2274and should turn into:
2275
2276 1: CURLYM[1] {1,32767}(18)
2277 5: TRIE(16)
2278 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2279 <ac>
2280 <ad>
2281 <ab>
2282 16: SUCCEED(0)
2283 17: NOTHING(18)
2284 18: END(0)
2285
2286Cases where tail != last would be like /(?foo|bar)baz/:
2287
2288 1: BRANCH(4)
2289 2: EXACT <foo>(8)
2290 4: BRANCH(7)
2291 5: EXACT <bar>(8)
2292 7: TAIL(8)
2293 8: EXACT <baz>(10)
2294 10: END(0)
2295
2296which would be optimizable with startbranch=1, first=1, last=7, tail=8
2297and would end up looking like:
2298
2299 1: TRIE(8)
2300 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2301 <foo>
2302 <bar>
2303 7: TAIL(8)
2304 8: EXACT <baz>(10)
2305 10: END(0)
2306
c80e42f3 2307 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2308
2309is the recommended Unicode-aware way of saying
2310
2311 *(d++) = uv;
2312*/
2313
fab2782b 2314#define TRIE_STORE_REVCHAR(val) \
786e8c11 2315 STMT_START { \
73031816 2316 if (UTF) { \
668fcfea 2317 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2318 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2319 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2320 SvCUR_set(zlopp, kapow - flrbbbbb); \
2321 SvPOK_on(zlopp); \
2322 SvUTF8_on(zlopp); \
2323 av_push(revcharmap, zlopp); \
2324 } else { \
fab2782b 2325 char ooooff = (char)val; \
73031816
NC
2326 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2327 } \
2328 } STMT_END
786e8c11 2329
914a25d5
KW
2330/* This gets the next character from the input, folding it if not already
2331 * folded. */
2332#define TRIE_READ_CHAR STMT_START { \
2333 wordlen++; \
2334 if ( UTF ) { \
2335 /* if it is UTF then it is either already folded, or does not need \
2336 * folding */ \
1c1d615a 2337 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2338 } \
2339 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2340 /* This folder implies Unicode rules, which in the range expressible \
2341 * by not UTF is the lower case, with the two exceptions, one of \
2342 * which should have been taken care of before calling this */ \
2343 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2344 uvc = toLOWER_L1(*uc); \
2345 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2346 len = 1; \
914a25d5
KW
2347 } else { \
2348 /* raw data, will be folded later if needed */ \
2349 uvc = (U32)*uc; \
2350 len = 1; \
2351 } \
786e8c11
YO
2352} STMT_END
2353
2354
2355
2356#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2357 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2358 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2359 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2360 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2361 } \
2362 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2363 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2364 TRIE_LIST_CUR( state )++; \
2365} STMT_END
07be1b83 2366
786e8c11
YO
2367#define TRIE_LIST_NEW(state) STMT_START { \
2368 Newxz( trie->states[ state ].trans.list, \
2369 4, reg_trie_trans_le ); \
2370 TRIE_LIST_CUR( state ) = 1; \
2371 TRIE_LIST_LEN( state ) = 4; \
2372} STMT_END
07be1b83 2373
786e8c11
YO
2374#define TRIE_HANDLE_WORD(state) STMT_START { \
2375 U16 dupe= trie->states[ state ].wordnum; \
2376 regnode * const noper_next = regnext( noper ); \
2377 \
786e8c11
YO
2378 DEBUG_r({ \
2379 /* store the word for dumping */ \
2380 SV* tmp; \
2381 if (OP(noper) != NOTHING) \
740cce10 2382 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2383 else \
740cce10 2384 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2385 av_push( trie_words, tmp ); \
786e8c11
YO
2386 }); \
2387 \
2388 curword++; \
2e64971a
DM
2389 trie->wordinfo[curword].prev = 0; \
2390 trie->wordinfo[curword].len = wordlen; \
2391 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2392 \
2393 if ( noper_next < tail ) { \
2394 if (!trie->jump) \
538e84ed
KW
2395 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2396 sizeof(U16) ); \
7f69552c 2397 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2398 if (!jumper) \
2399 jumper = noper_next; \
2400 if (!nextbranch) \
2401 nextbranch= regnext(cur); \
2402 } \
2403 \
2404 if ( dupe ) { \
2e64971a
DM
2405 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2406 /* chain, so that when the bits of chain are later */\
2407 /* linked together, the dups appear in the chain */\
2408 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2409 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2410 } else { \
2411 /* we haven't inserted this word yet. */ \
2412 trie->states[ state ].wordnum = curword; \
2413 } \
2414} STMT_END
07be1b83 2415
3dab1dad 2416
786e8c11
YO
2417#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2418 ( ( base + charid >= ucharcount \
2419 && base + charid < ubound \
2420 && state == trie->trans[ base - ucharcount + charid ].check \
2421 && trie->trans[ base - ucharcount + charid ].next ) \
2422 ? trie->trans[ base - ucharcount + charid ].next \
2423 : ( state==1 ? special : 0 ) \
2424 )
3dab1dad 2425
8bcafbf4
YO
2426#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2427STMT_START { \
2428 TRIE_BITMAP_SET(trie, uvc); \
2429 /* store the folded codepoint */ \
2430 if ( folder ) \
2431 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2432 \
2433 if ( !UTF ) { \
2434 /* store first byte of utf8 representation of */ \
2435 /* variant codepoints */ \
2436 if (! UVCHR_IS_INVARIANT(uvc)) { \
2437 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2438 } \
2439 } \
2440} STMT_END
786e8c11
YO
2441#define MADE_TRIE 1
2442#define MADE_JUMP_TRIE 2
2443#define MADE_EXACT_TRIE 4
3dab1dad 2444
a3621e74 2445STATIC I32
538e84ed
KW
2446S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2447 regnode *first, regnode *last, regnode *tail,
2448 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2449{
2450 /* first pass, loop through and scan words */
2451 reg_trie_data *trie;
55eed653 2452 HV *widecharmap = NULL;
2b8b4781 2453 AV *revcharmap = newAV();
a3621e74 2454 regnode *cur;
a3621e74
YO
2455 STRLEN len = 0;
2456 UV uvc = 0;
2457 U16 curword = 0;
2458 U32 next_alloc = 0;
786e8c11
YO
2459 regnode *jumper = NULL;
2460 regnode *nextbranch = NULL;
7f69552c 2461 regnode *convert = NULL;
2e64971a 2462 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2463 /* we just use folder as a flag in utf8 */
1e696034 2464 const U8 * folder = NULL;
a3621e74 2465
2b8b4781 2466#ifdef DEBUGGING
cf78de0b 2467 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2468 AV *trie_words = NULL;
2469 /* along with revcharmap, this only used during construction but both are
2470 * useful during debugging so we store them in the struct when debugging.
8e11feef 2471 */
2b8b4781 2472#else
cf78de0b 2473 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2474 STRLEN trie_charcount=0;
3dab1dad 2475#endif
2b8b4781 2476 SV *re_trie_maxbuff;
a3621e74 2477 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2478
2479 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2480#ifndef DEBUGGING
2481 PERL_UNUSED_ARG(depth);
2482#endif
a3621e74 2483
1e696034 2484 switch (flags) {
a4525e78 2485 case EXACT: case EXACTL: break;
2f7f8cb1 2486 case EXACTFA:
fab2782b 2487 case EXACTFU_SS:
a4525e78
KW
2488 case EXACTFU:
2489 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2490 case EXACTF: folder = PL_fold; break;
fab2782b 2491 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2492 }
2493
c944940b 2494 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2495 trie->refcount = 1;
3dab1dad 2496 trie->startstate = 1;
786e8c11 2497 trie->wordcount = word_count;
f8fc2ecf 2498 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2499 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2500 if (flags == EXACT || flags == EXACTL)
c944940b 2501 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2502 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2503 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2504
a3621e74 2505 DEBUG_r({
2b8b4781 2506 trie_words = newAV();
a3621e74 2507 });
a3621e74 2508
0111c4fd 2509 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2510 assert(re_trie_maxbuff);
a3621e74 2511 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2512 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2513 }
df826430 2514 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2515 Perl_re_indentf( aTHX_
cb41e5d6
YO
2516 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2517 depth+1,
538e84ed
KW
2518 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2519 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2520 });
538e84ed 2521
7f69552c
YO
2522 /* Find the node we are going to overwrite */
2523 if ( first == startbranch && OP( last ) != BRANCH ) {
2524 /* whole branch chain */
2525 convert = first;
2526 } else {
2527 /* branch sub-chain */
2528 convert = NEXTOPER( first );
2529 }
538e84ed 2530
a3621e74
YO
2531 /* -- First loop and Setup --
2532
2533 We first traverse the branches and scan each word to determine if it
2534 contains widechars, and how many unique chars there are, this is
2535 important as we have to build a table with at least as many columns as we
2536 have unique chars.
2537
2538 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2539 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2540 the native representation of the character value as the key and IV's for
2541 the coded index.
a3621e74
YO
2542
2543 *TODO* If we keep track of how many times each character is used we can
2544 remap the columns so that the table compression later on is more
3b753521 2545 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2546 middle and the least common are on the outside. IMO this would be better
2547 than a most to least common mapping as theres a decent chance the most
2548 common letter will share a node with the least common, meaning the node
486ec47a 2549 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2550 case is when we have the least common nodes twice.
2551
2552 */
2553
a3621e74 2554 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2555 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2556 const U8 *uc;
2557 const U8 *e;
bc031a7d 2558 int foldlen = 0;
07be1b83 2559 U32 wordlen = 0; /* required init */
bc031a7d
KW
2560 STRLEN minchars = 0;
2561 STRLEN maxchars = 0;
538e84ed
KW
2562 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2563 bitmap?*/
a3621e74 2564
3dab1dad 2565 if (OP(noper) == NOTHING) {
20ed8c88
YO
2566 /* skip past a NOTHING at the start of an alternation
2567 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2568 */
df826430 2569 regnode *noper_next= regnext(noper);
944e05e3
YO
2570 if (noper_next < tail)
2571 noper= noper_next;
2572 }
2573
dca5fc4c
YO
2574 if ( noper < tail &&
2575 (
2576 OP(noper) == flags ||
2577 (
2578 flags == EXACTFU &&
2579 OP(noper) == EXACTFU_SS
2580 )
2581 )
2582 ) {
944e05e3
YO
2583 uc= (U8*)STRING(noper);
2584 e= uc + STR_LEN(noper);
2585 } else {
2586 trie->minlen= 0;
2587 continue;
3dab1dad 2588 }
df826430 2589
944e05e3 2590
fab2782b 2591 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2592 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2593 regardless of encoding */
fab2782b
YO
2594 if (OP( noper ) == EXACTFU_SS) {
2595 /* false positives are ok, so just set this */
0dc4a61d 2596 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2597 }
2598 }
dca5fc4c 2599
bc031a7d
KW
2600 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2601 branch */
3dab1dad 2602 TRIE_CHARCOUNT(trie)++;
a3621e74 2603 TRIE_READ_CHAR;
645de4ce 2604
bc031a7d
KW
2605 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2606 * is in effect. Under /i, this character can match itself, or
2607 * anything that folds to it. If not under /i, it can match just
2608 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2609 * all fold to k, and all are single characters. But some folds
2610 * expand to more than one character, so for example LATIN SMALL
2611 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2612 * the string beginning at 'uc' is 'ffi', it could be matched by
2613 * three characters, or just by the one ligature character. (It
2614 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2615 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2616 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2617 * match.) The trie needs to know the minimum and maximum number
2618 * of characters that could match so that it can use size alone to
2619 * quickly reject many match attempts. The max is simple: it is
2620 * the number of folded characters in this branch (since a fold is
2621 * never shorter than what folds to it. */
2622
2623 maxchars++;
2624
2625 /* And the min is equal to the max if not under /i (indicated by
2626 * 'folder' being NULL), or there are no multi-character folds. If
2627 * there is a multi-character fold, the min is incremented just
2628 * once, for the character that folds to the sequence. Each
2629 * character in the sequence needs to be added to the list below of
2630 * characters in the trie, but we count only the first towards the
2631 * min number of characters needed. This is done through the
2632 * variable 'foldlen', which is returned by the macros that look
2633 * for these sequences as the number of bytes the sequence
2634 * occupies. Each time through the loop, we decrement 'foldlen' by
2635 * how many bytes the current char occupies. Only when it reaches
2636 * 0 do we increment 'minchars' or look for another multi-character
2637 * sequence. */
2638 if (folder == NULL) {
2639 minchars++;
2640 }
2641 else if (foldlen > 0) {
2642 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2643 }
2644 else {
bc031a7d
KW
2645 minchars++;
2646
2647 /* See if *uc is the beginning of a multi-character fold. If
2648 * so, we decrement the length remaining to look at, to account
2649 * for the current character this iteration. (We can use 'uc'
2650 * instead of the fold returned by TRIE_READ_CHAR because for
2651 * non-UTF, the latin1_safe macro is smart enough to account
2652 * for all the unfolded characters, and because for UTF, the
2653 * string will already have been folded earlier in the
2654 * compilation process */
2655 if (UTF) {
2656 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2657 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2658 }
2659 }
bc031a7d
KW
2660 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2661 foldlen--;
2662 }
645de4ce 2663 }
bc031a7d
KW
2664
2665 /* The current character (and any potential folds) should be added
2666 * to the possible matching characters for this position in this
2667 * branch */
a3621e74 2668 if ( uvc < 256 ) {
fab2782b
YO
2669 if ( folder ) {
2670 U8 folded= folder[ (U8) uvc ];
2671 if ( !trie->charmap[ folded ] ) {
2672 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2673 TRIE_STORE_REVCHAR( folded );
2674 }
2675 }
a3621e74
YO
2676 if ( !trie->charmap[ uvc ] ) {
2677 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2678 TRIE_STORE_REVCHAR( uvc );
a3621e74 2679 }
02daf0ab 2680 if ( set_bit ) {
62012aee
KW
2681 /* store the codepoint in the bitmap, and its folded
2682 * equivalent. */
8bcafbf4 2683 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2684 set_bit = 0; /* We've done our bit :-) */
2685 }
a3621e74 2686 } else {
bc031a7d
KW
2687
2688 /* XXX We could come up with the list of code points that fold
2689 * to this using PL_utf8_foldclosures, except not for
2690 * multi-char folds, as there may be multiple combinations
2691 * there that could work, which needs to wait until runtime to
2692 * resolve (The comment about LIGATURE FFI above is such an
2693 * example */
2694
a3621e74 2695 SV** svpp;
55eed653
NC
2696 if ( !widecharmap )
2697 widecharmap = newHV();
a3621e74 2698
55eed653 2699 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2700
2701 if ( !svpp )
147e3846 2702 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2703
2704 if ( !SvTRUE( *svpp ) ) {
2705 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2706 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2707 }
2708 }
bc031a7d
KW
2709 } /* end loop through characters in this branch of the trie */
2710
2711 /* We take the min and max for this branch and combine to find the min
2712 * and max for all branches processed so far */
3dab1dad 2713 if( cur == first ) {
bc031a7d
KW
2714 trie->minlen = minchars;
2715 trie->maxlen = maxchars;
2716 } else if (minchars < trie->minlen) {
2717 trie->minlen = minchars;
2718 } else if (maxchars > trie->maxlen) {
2719 trie->maxlen = maxchars;
fab2782b 2720 }
a3621e74
YO
2721 } /* end first pass */
2722 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2723 Perl_re_indentf( aTHX_
cb41e5d6
YO
2724 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2725 depth+1,
55eed653 2726 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2727 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2728 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2729 );
a3621e74
YO
2730
2731 /*
2732 We now know what we are dealing with in terms of unique chars and
2733 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2734 representation using a flat table will take. If it's over a reasonable
2735 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2736 conservative but potentially much slower representation using an array
2737 of lists.
2738
2739 At the end we convert both representations into the same compressed
2740 form that will be used in regexec.c for matching with. The latter
2741 is a form that cannot be used to construct with but has memory
2742 properties similar to the list form and access properties similar
2743 to the table form making it both suitable for fast searches and
2744 small enough that its feasable to store for the duration of a program.
2745
2746 See the comment in the code where the compressed table is produced
2747 inplace from the flat tabe representation for an explanation of how
2748 the compression works.
2749
2750 */
2751
2752
2e64971a
DM
2753 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2754 prev_states[1] = 0;
2755
538e84ed
KW
2756 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2757 > SvIV(re_trie_maxbuff) )
2758 {
a3621e74
YO
2759 /*
2760 Second Pass -- Array Of Lists Representation
2761
2762 Each state will be represented by a list of charid:state records
2763 (reg_trie_trans_le) the first such element holds the CUR and LEN
2764 points of the allocated array. (See defines above).
2765
2766 We build the initial structure using the lists, and then convert
2767 it into the compressed table form which allows faster lookups
2768 (but cant be modified once converted).
a3621e74
YO
2769 */
2770
a3621e74
YO
2771 STRLEN transcount = 1;
2772
6ad9a8ab 2773 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2774 depth+1));
686b73d4 2775
c944940b
JH
2776 trie->states = (reg_trie_state *)
2777 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2778 sizeof(reg_trie_state) );
a3621e74
YO
2779 TRIE_LIST_NEW(1);
2780 next_alloc = 2;
2781
2782 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2783
df826430 2784 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2785 U32 state = 1; /* required init */
2786 U16 charid = 0; /* sanity init */
07be1b83 2787 U32 wordlen = 0; /* required init */
c445ea15 2788
df826430
YO
2789 if (OP(noper) == NOTHING) {
2790 regnode *noper_next= regnext(noper);
944e05e3
YO
2791 if (noper_next < tail)
2792 noper= noper_next;
df826430
YO
2793 }
2794
944e05e3
YO
2795 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2796 const U8 *uc= (U8*)STRING(noper);
2797 const U8 *e= uc + STR_LEN(noper);
2798
786e8c11 2799 for ( ; uc < e ; uc += len ) {
c445ea15 2800
786e8c11 2801 TRIE_READ_CHAR;
c445ea15 2802
786e8c11
YO
2803 if ( uvc < 256 ) {
2804 charid = trie->charmap[ uvc ];
c445ea15 2805 } else {
538e84ed
KW
2806 SV** const svpp = hv_fetch( widecharmap,
2807 (char*)&uvc,
2808 sizeof( UV ),
2809 0);
786e8c11
YO
2810 if ( !svpp ) {
2811 charid = 0;
2812 } else {
2813 charid=(U16)SvIV( *svpp );
2814 }
c445ea15 2815 }
538e84ed
KW
2816 /* charid is now 0 if we dont know the char read, or
2817 * nonzero if we do */
786e8c11 2818 if ( charid ) {
a3621e74 2819
786e8c11
YO
2820 U16 check;
2821 U32 newstate = 0;
a3621e74 2822
786e8c11
YO
2823 charid--;
2824 if ( !trie->states[ state ].trans.list ) {
2825 TRIE_LIST_NEW( state );
c445ea15 2826 }
538e84ed
KW
2827 for ( check = 1;
2828 check <= TRIE_LIST_USED( state );
2829 check++ )
2830 {
2831 if ( TRIE_LIST_ITEM( state, check ).forid
2832 == charid )
2833 {
786e8c11
YO
2834 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2835 break;
2836 }
2837 }
2838 if ( ! newstate ) {
2839 newstate = next_alloc++;
2e64971a 2840 prev_states[newstate] = state;
786e8c11
YO
2841 TRIE_LIST_PUSH( state, charid, newstate );
2842 transcount++;
2843 }
2844 state = newstate;
2845 } else {
147e3846 2846 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 2847 }
a28509cc 2848 }
c445ea15 2849 }
3dab1dad 2850 TRIE_HANDLE_WORD(state);
a3621e74
YO
2851
2852 } /* end second pass */
2853
1e2e3d02 2854 /* next alloc is the NEXT state to be allocated */
538e84ed 2855 trie->statecount = next_alloc;
c944940b
JH
2856 trie->states = (reg_trie_state *)
2857 PerlMemShared_realloc( trie->states,
2858 next_alloc
2859 * sizeof(reg_trie_state) );
a3621e74 2860
3dab1dad 2861 /* and now dump it out before we compress it */
2b8b4781
NC
2862 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2863 revcharmap, next_alloc,
2864 depth+1)
1e2e3d02 2865 );
a3621e74 2866
c944940b
JH
2867 trie->trans = (reg_trie_trans *)
2868 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2869 {
2870 U32 state;
a3621e74
YO
2871 U32 tp = 0;
2872 U32 zp = 0;
2873
2874
2875 for( state=1 ; state < next_alloc ; state ++ ) {
2876 U32 base=0;
2877
2878 /*
2879 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2880 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
2881 );
2882 */
2883
2884 if (trie->states[state].trans.list) {
2885 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2886 U16 maxid=minid;
a28509cc 2887 U16 idx;
a3621e74
YO
2888
2889 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2890 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2891 if ( forid < minid ) {
2892 minid=forid;
2893 } else if ( forid > maxid ) {
2894 maxid=forid;
2895 }
a3621e74
YO
2896 }
2897 if ( transcount < tp + maxid - minid + 1) {
2898 transcount *= 2;
c944940b
JH
2899 trie->trans = (reg_trie_trans *)
2900 PerlMemShared_realloc( trie->trans,
446bd890
NC
2901 transcount
2902 * sizeof(reg_trie_trans) );
538e84ed
KW
2903 Zero( trie->trans + (transcount / 2),
2904 transcount / 2,
2905 reg_trie_trans );
a3621e74
YO
2906 }
2907 base = trie->uniquecharcount + tp - minid;
2908 if ( maxid == minid ) {
2909 U32 set = 0;
2910 for ( ; zp < tp ; zp++ ) {
2911 if ( ! trie->trans[ zp ].next ) {
2912 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2913 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2914 1).newstate;
a3621e74
YO
2915 trie->trans[ zp ].check = state;
2916 set = 1;
2917 break;
2918 }
2919 }
2920 if ( !set ) {
538e84ed
KW
2921 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2922 1).newstate;
a3621e74
YO
2923 trie->trans[ tp ].check = state;
2924 tp++;
2925 zp = tp;
2926 }
2927 } else {
2928 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2929 const U32 tid = base
2930 - trie->uniquecharcount
2931 + TRIE_LIST_ITEM( state, idx ).forid;
2932 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2933 idx ).newstate;
a3621e74
YO
2934 trie->trans[ tid ].check = state;
2935 }
2936 tp += ( maxid - minid + 1 );
2937 }
2938 Safefree(trie->states[ state ].trans.list);
2939 }
2940 /*
2941 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2942 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
2943 );
2944 */
2945 trie->states[ state ].trans.base=base;
2946 }
cc601c31 2947 trie->lasttrans = tp + 1;
a3621e74
YO
2948 }
2949 } else {
2950 /*
2951 Second Pass -- Flat Table Representation.
2952
b423522f
KW
2953 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2954 each. We know that we will need Charcount+1 trans at most to store
2955 the data (one row per char at worst case) So we preallocate both
2956 structures assuming worst case.
a3621e74
YO
2957
2958 We then construct the trie using only the .next slots of the entry
2959 structs.
2960
b423522f
KW
2961 We use the .check field of the first entry of the node temporarily
2962 to make compression both faster and easier by keeping track of how
2963 many non zero fields are in the node.
a3621e74
YO
2964
2965 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2966 transition.
2967
b423522f
KW
2968 There are two terms at use here: state as a TRIE_NODEIDX() which is
2969 a number representing the first entry of the node, and state as a
2970 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2971 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2972 if there are 2 entrys per node. eg:
a3621e74
YO
2973
2974 A B A B
2975 1. 2 4 1. 3 7
2976 2. 0 3 3. 0 5
2977 3. 0 0 5. 0 0
2978 4. 0 0 7. 0 0
2979
b423522f
KW
2980 The table is internally in the right hand, idx form. However as we
2981 also have to deal with the states array which is indexed by nodenum
2982 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2983
2984 */
6ad9a8ab 2985 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 2986 depth+1));
3dab1dad 2987
c944940b
JH
2988 trie->trans = (reg_trie_trans *)
2989 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2990 * trie->uniquecharcount + 1,
2991 sizeof(reg_trie_trans) );
2992 trie->states = (reg_trie_state *)
2993 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2994 sizeof(reg_trie_state) );
a3621e74
YO
2995 next_alloc = trie->uniquecharcount + 1;
2996
3dab1dad 2997
a3621e74
YO
2998 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2999
df826430 3000 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3001
3002 U32 state = 1; /* required init */
3003
3004 U16 charid = 0; /* sanity init */
3005 U32 accept_state = 0; /* sanity init */
a3621e74 3006
07be1b83 3007 U32 wordlen = 0; /* required init */
a3621e74 3008
df826430
YO
3009 if (OP(noper) == NOTHING) {
3010 regnode *noper_next= regnext(noper);
944e05e3
YO
3011 if (noper_next < tail)
3012 noper= noper_next;
df826430 3013 }
fab2782b 3014
944e05e3
YO
3015 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3016 const U8 *uc= (U8*)STRING(noper);
3017 const U8 *e= uc + STR_LEN(noper);
3018
786e8c11 3019 for ( ; uc < e ; uc += len ) {
a3621e74 3020
786e8c11 3021 TRIE_READ_CHAR;
a3621e74 3022
786e8c11
YO
3023 if ( uvc < 256 ) {
3024 charid = trie->charmap[ uvc ];
3025 } else {
538e84ed
KW
3026 SV* const * const svpp = hv_fetch( widecharmap,
3027 (char*)&uvc,
3028 sizeof( UV ),
3029 0);
786e8c11 3030 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3031 }
786e8c11
YO
3032 if ( charid ) {
3033 charid--;
3034 if ( !trie->trans[ state + charid ].next ) {
3035 trie->trans[ state + charid ].next = next_alloc;
3036 trie->trans[ state ].check++;
2e64971a
DM
3037 prev_states[TRIE_NODENUM(next_alloc)]
3038 = TRIE_NODENUM(state);
786e8c11
YO
3039 next_alloc += trie->uniquecharcount;
3040 }
3041 state = trie->trans[ state + charid ].next;
3042 } else {
147e3846 3043 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3044 }
538e84ed
KW
3045 /* charid is now 0 if we dont know the char read, or
3046 * nonzero if we do */
a3621e74 3047 }
a3621e74 3048 }
3dab1dad
YO
3049 accept_state = TRIE_NODENUM( state );
3050 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3051
3052 } /* end second pass */
3053
3dab1dad 3054 /* and now dump it out before we compress it */
2b8b4781
NC
3055 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3056 revcharmap,
3057 next_alloc, depth+1));
a3621e74 3058
a3621e74
YO
3059 {
3060 /*
3061 * Inplace compress the table.*
3062
3063 For sparse data sets the table constructed by the trie algorithm will
3064 be mostly 0/FAIL transitions or to put it another way mostly empty.
3065 (Note that leaf nodes will not contain any transitions.)
3066
3067 This algorithm compresses the tables by eliminating most such
3068 transitions, at the cost of a modest bit of extra work during lookup:
3069
3070 - Each states[] entry contains a .base field which indicates the
3071 index in the state[] array wheres its transition data is stored.
3072
3b753521 3073 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3074
3075 - If .base is nonzero then charid is added to it to find an entry in
3076 the trans array.
3077
3078 -If trans[states[state].base+charid].check!=state then the
3079 transition is taken to be a 0/Fail transition. Thus if there are fail
3080 transitions at the front of the node then the .base offset will point
3081 somewhere inside the previous nodes data (or maybe even into a node
3082 even earlier), but the .check field determines if the transition is
3083 valid.
3084
786e8c11 3085 XXX - wrong maybe?
a3621e74 3086 The following process inplace converts the table to the compressed
3b753521 3087 table: We first do not compress the root node 1,and mark all its
a3621e74 3088 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3089 allows us to do a DFA construction from the compressed table later,
3090 and ensures that any .base pointers we calculate later are greater
3091 than 0.
a3621e74
YO
3092
3093 - We set 'pos' to indicate the first entry of the second node.
3094
3095 - We then iterate over the columns of the node, finding the first and
3096 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3097 and set the .check pointers accordingly, and advance pos
3098 appropriately and repreat for the next node. Note that when we copy
3099 the next pointers we have to convert them from the original
3100 NODEIDX form to NODENUM form as the former is not valid post
3101 compression.
3102
3103 - If a node has no transitions used we mark its base as 0 and do not
3104 advance the pos pointer.
3105
3106 - If a node only has one transition we use a second pointer into the
3107 structure to fill in allocated fail transitions from other states.
3108 This pointer is independent of the main pointer and scans forward
3109 looking for null transitions that are allocated to a state. When it
3110 finds one it writes the single transition into the "hole". If the
786e8c11 3111 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3112
3113 - Once compressed we can Renew/realloc the structures to release the
3114 excess space.
3115
3116 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3117 specifically Fig 3.47 and the associated pseudocode.
3118
3119 demq
3120 */
a3b680e6 3121 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3122 U32 state, charid;
a3621e74 3123 U32 pos = 0, zp=0;
1e2e3d02 3124 trie->statecount = laststate;
a3621e74
YO
3125
3126 for ( state = 1 ; state < laststate ; state++ ) {
3127 U8 flag = 0;
a28509cc
AL
3128 const U32 stateidx = TRIE_NODEIDX( state );
3129 const U32 o_used = trie->trans[ stateidx ].check;
3130 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3131 trie->trans[ stateidx ].check = 0;
3132
538e84ed
KW
3133 for ( charid = 0;
3134 used && charid < trie->uniquecharcount;
3135 charid++ )
3136 {
a3621e74
YO
3137 if ( flag || trie->trans[ stateidx + charid ].next ) {
3138 if ( trie->trans[ stateidx + charid ].next ) {
3139 if (o_used == 1) {
3140 for ( ; zp < pos ; zp++ ) {
3141 if ( ! trie->trans[ zp ].next ) {
3142 break;
3143 }
3144 }
538e84ed
KW
3145 trie->states[ state ].trans.base
3146 = zp
3147 + trie->uniquecharcount
3148 - charid ;
3149 trie->trans[ zp ].next
3150 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3151 + charid ].next );
a3621e74
YO
3152 trie->trans[ zp ].check = state;
3153 if ( ++zp > pos ) pos = zp;
3154 break;
3155 }
3156 used--;
3157 }
3158 if ( !flag ) {
3159 flag = 1;
538e84ed
KW
3160 trie->states[ state ].trans.base
3161 = pos + trie->uniquecharcount - charid ;
a3621e74 3162 }
538e84ed
KW
3163 trie->trans[ pos ].next
3164 = SAFE_TRIE_NODENUM(
3165 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3166 trie->trans[ pos ].check = state;
3167 pos++;
3168 }
3169 }
3170 }
cc601c31 3171 trie->lasttrans = pos + 1;
c944940b
JH
3172 trie->states = (reg_trie_state *)
3173 PerlMemShared_realloc( trie->states, laststate
3174 * sizeof(reg_trie_state) );
a3621e74 3175 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3176 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
cb41e5d6 3177 depth+1,
538e84ed
KW
3178 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3179 + 1 ),
3180 (IV)next_alloc,
3181 (IV)pos,
3182 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3183 );
3184
3185 } /* end table compress */
3186 }
1e2e3d02 3187 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3188 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
cb41e5d6 3189 depth+1,
1e2e3d02
YO
3190 (UV)trie->statecount,
3191 (UV)trie->lasttrans)
3192 );
cc601c31 3193 /* resize the trans array to remove unused space */
c944940b
JH
3194 trie->trans = (reg_trie_trans *)
3195 PerlMemShared_realloc( trie->trans, trie->lasttrans
3196 * sizeof(reg_trie_trans) );
a3621e74 3197
538e84ed 3198 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3199 U8 nodetype =(U8)(flags & 0xFF);
3200 char *str=NULL;
538e84ed 3201
07be1b83 3202#ifdef DEBUGGING
e62cc96a 3203 regnode *optimize = NULL;
7122b237
YO
3204#ifdef RE_TRACK_PATTERN_OFFSETS
3205
b57a0404
JH
3206 U32 mjd_offset = 0;
3207 U32 mjd_nodelen = 0;
7122b237
YO
3208#endif /* RE_TRACK_PATTERN_OFFSETS */
3209#endif /* DEBUGGING */
a3621e74 3210 /*
3dab1dad
YO
3211 This means we convert either the first branch or the first Exact,
3212 depending on whether the thing following (in 'last') is a branch
3213 or not and whther first is the startbranch (ie is it a sub part of
3214 the alternation or is it the whole thing.)
3b753521 3215 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3216 the whole branch sequence, including the first.
a3621e74 3217 */
3dab1dad 3218 /* Find the node we are going to overwrite */
7f69552c 3219 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3220 /* branch sub-chain */
3dab1dad 3221 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3222#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3223 DEBUG_r({
3224 mjd_offset= Node_Offset((convert));
3225 mjd_nodelen= Node_Length((convert));
3226 });
7122b237 3227#endif
7f69552c 3228 /* whole branch chain */
7122b237
YO
3229 }
3230#ifdef RE_TRACK_PATTERN_OFFSETS
3231 else {
7f69552c
YO
3232 DEBUG_r({
3233 const regnode *nop = NEXTOPER( convert );
3234 mjd_offset= Node_Offset((nop));
3235 mjd_nodelen= Node_Length((nop));
3236 });
07be1b83
YO
3237 }
3238 DEBUG_OPTIMISE_r(
147e3846 3239 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
cb41e5d6 3240 depth+1,
786e8c11 3241 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3242 );
7122b237 3243#endif
538e84ed 3244 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3245 split out as an EXACT and put in front of the TRIE node. */
3246 trie->startstate= 1;
55eed653 3247 if ( trie->bitmap && !widecharmap && !trie->jump ) {
5ee57374
YO
3248 /* we want to find the first state that has more than
3249 * one transition, if that state is not the first state
3250 * then we have a common prefix which we can remove.
3251 */
3dab1dad 3252 U32 state;
1e2e3d02 3253 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3254 U32 ofs = 0;
ee0dfd0b
YO
3255 I32 first_ofs = -1; /* keeps track of the ofs of the first
3256 transition, -1 means none */
8e11feef
RGS
3257 U32 count = 0;
3258 const U32 base = trie->states[ state ].trans.base;
a3621e74 3259
5ee57374 3260 /* does this state terminate an alternation? */
3dab1dad 3261 if ( trie->states[state].wordnum )
8e11feef 3262 count = 1;
a3621e74 3263
8e11feef 3264 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3265 if ( ( base + ofs >= trie->uniquecharcount ) &&
3266 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3267 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3268 {
3dab1dad 3269 if ( ++count > 1 ) {
5ee57374 3270 /* we have more than one transition */
d3d91c7c
YO
3271 SV **tmp;
3272 U8 *ch;
5ee57374
YO
3273 /* if this is the first state there is no common prefix
3274 * to extract, so we can exit */
8e11feef 3275 if ( state == 1 ) break;
d3d91c7c
YO
3276 tmp = av_fetch( revcharmap, ofs, 0);
3277 ch = (U8*)SvPV_nolen_const( *tmp );
3278
5ee57374
YO
3279 /* if we are on count 2 then we need to initialize the
3280 * bitmap, and store the previous char if there was one
3281 * in it*/
3dab1dad 3282 if ( count == 2 ) {
5ee57374 3283 /* clear the bitmap */
3dab1dad
YO
3284 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3285 DEBUG_OPTIMISE_r(
147e3846 3286 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
cb41e5d6 3287 depth+1,
786e8c11 3288 (UV)state));
ee0dfd0b
YO
3289 if (first_ofs >= 0) {
3290 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
be8e71aa 3291 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3292
8bcafbf4 3293 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3dab1dad 3294 DEBUG_OPTIMISE_r(
6ad9a8ab 3295 Perl_re_printf( aTHX_ "%s", (char*)ch)
3dab1dad 3296 );
8e11feef
RGS
3297 }
3298 }
8bcafbf4
YO
3299 /* store the current firstchar in the bitmap */
3300 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
6ad9a8ab 3301 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
8e11feef 3302 }
ee0dfd0b 3303 first_ofs = ofs;
8e11feef 3304 }
3dab1dad
YO
3305 }
3306 if ( count == 1 ) {
5ee57374
YO
3307 /* This state has only one transition, its transition is part
3308 * of a common prefix - we need to concatenate the char it
3309 * represents to what we have so far. */
ee0dfd0b 3310 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
c490c714
YO
3311 STRLEN len;
3312 char *ch = SvPV( *tmp, len );
de734bd5
A
3313 DEBUG_OPTIMISE_r({
3314 SV *sv=sv_newmortal();
147e3846 3315 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
cb41e5d6 3316 depth+1,
ee0dfd0b 3317 (UV)state, (UV)first_ofs,
538e84ed 3318 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3319 PL_colors[0], PL_colors[1],
3320 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3321 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3322 )
3323 );
3324 });
3dab1dad
YO
3325 if ( state==1 ) {
3326 OP( convert ) = nodetype;
3327 str=STRING(convert);
3328 STR_LEN(convert)=0;
3329 }
c490c714
YO
3330 STR_LEN(convert) += len;
3331 while (len--)
de734bd5 3332 *str++ = *ch++;
8e11feef 3333 } else {
538e84ed 3334#ifdef DEBUGGING
8e11feef 3335 if (state>1)
6ad9a8ab 3336 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
f9049ba1 3337#endif
8e11feef
RGS
3338 break;
3339 }
3340 }
2e64971a 3341 trie->prefixlen = (state-1);
3dab1dad 3342 if (str) {
8e11feef 3343 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3344 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3345 trie->startstate = state;
07be1b83
YO
3346 trie->minlen -= (state - 1);
3347 trie->maxlen -= (state - 1);
33809eae
JH
3348#ifdef DEBUGGING
3349 /* At least the UNICOS C compiler choked on this
3350 * being argument to DEBUG_r(), so let's just have
3351 * it right here. */
3352 if (
3353#ifdef PERL_EXT_RE_BUILD
3354 1
3355#else
3356 DEBUG_r_TEST
3357#endif
3358 ) {
3359 regnode *fix = convert;
3360 U32 word = trie->wordcount;
3361 mjd_nodelen++;
3362 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3363 while( ++fix < n ) {
3364 Set_Node_Offset_Length(fix, 0, 0);
3365 }
3366 while (word--) {
3367 SV ** const tmp = av_fetch( trie_words, word, 0 );
3368 if (tmp) {
3369 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3370 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3371 else
3372 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3373 }
3374 }
3375 }
3376#endif
8e11feef
RGS
3377 if (trie->maxlen) {
3378 convert = n;
3379 } else {
3dab1dad 3380 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3381 DEBUG_r(optimize= n);
3dab1dad
YO
3382 }
3383 }
3384 }
538e84ed
KW
3385 if (!jumper)
3386 jumper = last;
3dab1dad 3387 if ( trie->maxlen ) {
8e11feef
RGS
3388 NEXT_OFF( convert ) = (U16)(tail - convert);
3389 ARG_SET( convert, data_slot );
538e84ed
KW
3390 /* Store the offset to the first unabsorbed branch in
3391 jump[0], which is otherwise unused by the jump logic.
786e8c11 3392 We use this when dumping a trie and during optimisation. */
538e84ed 3393 if (trie->jump)
7f69552c 3394 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3395
6c48061a
YO
3396 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3397 * and there is a bitmap
3398 * and the first "jump target" node we found leaves enough room
3399 * then convert the TRIE node into a TRIEC node, with the bitmap
3400 * embedded inline in the opcode - this is hypothetically faster.
3401 */
3402 if ( !trie->states[trie->startstate].wordnum
3403 && trie->bitmap
3404 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3405 {
3406 OP( convert ) = TRIEC;
3407 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3408 PerlMemShared_free(trie->bitmap);
786e8c11 3409 trie->bitmap= NULL;
538e84ed 3410 } else
786e8c11 3411 OP( convert ) = TRIE;
a3621e74 3412
3dab1dad
YO
3413 /* store the type in the flags */
3414 convert->flags = nodetype;
a5ca303d 3415 DEBUG_r({
538e84ed
KW
3416 optimize = convert
3417 + NODE_STEP_REGNODE
a5ca303d
YO
3418 + regarglen[ OP( convert ) ];
3419 });
538e84ed 3420 /* XXX We really should free up the resource in trie now,
a5ca303d 3421 as we won't use them - (which resources?) dmq */
3dab1dad 3422 }
a3621e74 3423 /* needed for dumping*/
e62cc96a 3424 DEBUG_r(if (optimize) {
07be1b83 3425 regnode *opt = convert;
bcdf7404 3426
e62cc96a 3427 while ( ++opt < optimize) {
07be1b83
YO
3428 Set_Node_Offset_Length(opt,0,0);
3429 }
538e84ed
KW
3430 /*
3431 Try to clean up some of the debris left after the
786e8c11 3432 optimisation.
a3621e74 3433 */
786e8c11 3434 while( optimize < jumper ) {
07be1b83 3435 mjd_nodelen += Node_Length((optimize));
a3621e74 3436 OP( optimize ) = OPTIMIZED;
07be1b83 3437 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3438 optimize++;
3439 }
07be1b83 3440 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3441 });
3442 } /* end node insert */
2e64971a
DM
3443
3444 /* Finish populating the prev field of the wordinfo array. Walk back
3445 * from each accept state until we find another accept state, and if
3446 * so, point the first word's .prev field at the second word. If the
3447 * second already has a .prev field set, stop now. This will be the
3448 * case either if we've already processed that word's accept state,
3b753521
FN
3449 * or that state had multiple words, and the overspill words were
3450 * already linked up earlier.
2e64971a
DM
3451 */
3452 {
3453 U16 word;
3454 U32 state;
3455 U16 prev;
3456
3457 for (word=1; word <= trie->wordcount; word++) {
3458 prev = 0;
3459 if (trie->wordinfo[word].prev)
3460 continue;
3461 state = trie->wordinfo[word].accept;
3462 while (state) {
3463 state = prev_states[state];
3464 if (!state)
3465 break;
3466 prev = trie->states[state].wordnum;
3467 if (prev)
3468 break;
3469 }
3470 trie->wordinfo[word].prev = prev;
3471 }
3472 Safefree(prev_states);
3473 }
3474
3475