This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Warn on unescaped /[]}]/ under re strict
[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;
3d2bd50a 185 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 186 within pattern */
b1603ef8
DM
187 int num_code_blocks; /* size of code_blocks[] */
188 int code_index; /* next code_blocks[] slot */
ee273784 189 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
190 scan_frame *frame_head;
191 scan_frame *frame_last;
192 U32 frame_count;
7eec73eb 193 AV *warn_text;
dc6d7f5c 194#ifdef ADD_TO_REGEXEC
830247a4
IZ
195 char *starttry; /* -Dr: where regtry was called. */
196#define RExC_starttry (pRExC_state->starttry)
197#endif
d24ca0c5 198 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 199#ifdef DEBUGGING
be8e71aa 200 const char *lastparse;
3dab1dad 201 I32 lastnum;
1f1031fe 202 AV *paren_name_list; /* idx -> name */
d9a72fcc 203 U32 study_chunk_recursed_count;
c9f0d54c
YO
204 SV *mysv1;
205 SV *mysv2;
3dab1dad
YO
206#define RExC_lastparse (pRExC_state->lastparse)
207#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 208#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 209#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
210#define RExC_mysv (pRExC_state->mysv1)
211#define RExC_mysv1 (pRExC_state->mysv1)
212#define RExC_mysv2 (pRExC_state->mysv2)
213
3dab1dad 214#endif
512c0f5a 215 bool seen_unfolded_sharp_s;
911bd04e 216 bool strict;
da7cf1cc 217 bool study_started;
09b2b2e6 218};
830247a4 219
e2509266 220#define RExC_flags (pRExC_state->flags)
514a91f1 221#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 222#define RExC_precomp (pRExC_state->precomp)
285b5ca0
KW
223#define RExC_precomp_adj (pRExC_state->precomp_adj)
224#define RExC_adjusted_start (pRExC_state->adjusted_start)
711b303b 225#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 226#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 227#define RExC_rx (pRExC_state->rx)
f8fc2ecf 228#define RExC_rxi (pRExC_state->rxi)
fac92740 229#define RExC_start (pRExC_state->start)
830247a4
IZ
230#define RExC_end (pRExC_state->end)
231#define RExC_parse (pRExC_state->parse)
232#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
233
234/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
243
7122b237 244#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
245#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 others */
7122b237 247#endif
830247a4 248#define RExC_emit (pRExC_state->emit)
f7c7e32a 249#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 250#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 251#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
252#define RExC_sawback (pRExC_state->sawback)
253#define RExC_seen (pRExC_state->seen)
254#define RExC_size (pRExC_state->size)
ee273784 255#define RExC_maxlen (pRExC_state->maxlen)
830247a4 256#define RExC_npar (pRExC_state->npar)
e2e6a0f1 257#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
258#define RExC_extralen (pRExC_state->extralen)
259#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 260#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 261#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 262#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
263#define RExC_open_parens (pRExC_state->open_parens)
264#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 265#define RExC_end_op (pRExC_state->end_op)
81714fb9 266#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
267#define RExC_recurse (pRExC_state->recurse)
268#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 269#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
270#define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 272#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 273#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
274#ifdef EBCDIC
275# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276#endif
9d53c457 277#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
278#define RExC_frame_head (pRExC_state->frame_head)
279#define RExC_frame_last (pRExC_state->frame_last)
280#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 281#define RExC_strict (pRExC_state->strict)
da7cf1cc 282#define RExC_study_started (pRExC_state->study_started)
7eec73eb 283#define RExC_warn_text (pRExC_state->warn_text)
830247a4 284
99807a43
HS
285/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
286 * a flag to disable back-off on the fixed/floating substrings - if it's
287 * a high complexity pattern we assume the benefit of avoiding a full match
288 * is worth the cost of checking for the substrings even if they rarely help.
289 */
290#define RExC_naughty (pRExC_state->naughty)
291#define TOO_NAUGHTY (10)
292#define MARK_NAUGHTY(add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += (add)
295#define MARK_NAUGHTY_EXP(exp, add) \
296 if (RExC_naughty < TOO_NAUGHTY) \
297 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 298
a687059c
LW
299#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
300#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 301 ((*s) == '{' && regcurly(s)))
a687059c
LW
302
303/*
304 * Flags to be passed up and down.
305 */
a687059c 306#define WORST 0 /* Worst case. */
a3b492c3 307#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 308
e64f369d 309/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
310 * character. (There needs to be a case: in the switch statement in regexec.c
311 * for any node marked SIMPLE.) Note that this is not the same thing as
312 * REGNODE_SIMPLE */
fda99bee 313#define SIMPLE 0x02
e64f369d 314#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
315#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
316#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
317#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
318#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
319 calcuate sizes as UTF-8 */
a687059c 320
3dab1dad
YO
321#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
322
07be1b83
YO
323/* whether trie related optimizations are enabled */
324#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
325#define TRIE_STUDY_OPT
786e8c11 326#define FULL_TRIE_STUDY
07be1b83
YO
327#define TRIE_STCLASS
328#endif
1de06328
YO
329
330
40d049e4
YO
331
332#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
333#define PBITVAL(paren) (1 << ((paren) & 7))
334#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
335#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
336#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
337
82a6ada4 338#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 339 if (!UTF) { \
82a6ada4
KW
340 assert(PASS1); \
341 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
342 return NULL; \
343 } \
82a6ada4 344 } STMT_END
40d049e4 345
512c0f5a
KW
346/* Change from /d into /u rules, and restart the parse if we've already seen
347 * something whose size would increase as a result, by setting *flagp and
348 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
349 * we've change to /u during the parse. */
350#define REQUIRE_UNI_RULES(flagp, restart_retval) \
351 STMT_START { \
352 if (DEPENDS_SEMANTICS) { \
353 assert(PASS1); \
354 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
355 RExC_uni_semantics = 1; \
356 if (RExC_seen_unfolded_sharp_s) { \
357 *flagp |= RESTART_PASS1; \
358 return restart_retval; \
359 } \
360 } \
361 } STMT_END
362
f19b1a63
KW
363/* This converts the named class defined in regcomp.h to its equivalent class
364 * number defined in handy.h. */
365#define namedclass_to_classnum(class) ((int) ((class) / 2))
366#define classnum_to_namedclass(classnum) ((classnum) * 2)
367
de92f5e6
KW
368#define _invlist_union_complement_2nd(a, b, output) \
369 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
370#define _invlist_intersection_complement_2nd(a, b, output) \
371 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
372
1de06328
YO
373/* About scan_data_t.
374
375 During optimisation we recurse through the regexp program performing
376 various inplace (keyhole style) optimisations. In addition study_chunk
377 and scan_commit populate this data structure with information about
538e84ed 378 what strings MUST appear in the pattern. We look for the longest
3b753521 379 string that must appear at a fixed location, and we look for the
1de06328
YO
380 longest string that may appear at a floating location. So for instance
381 in the pattern:
538e84ed 382
1de06328 383 /FOO[xX]A.*B[xX]BAR/
538e84ed 384
1de06328
YO
385 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
386 strings (because they follow a .* construct). study_chunk will identify
387 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 388
1de06328 389 The strings can be composites, for instance
538e84ed 390
1de06328 391 /(f)(o)(o)/
538e84ed 392
1de06328 393 will result in a composite fixed substring 'foo'.
538e84ed 394
1de06328 395 For each string some basic information is maintained:
538e84ed 396
1de06328
YO
397 - offset or min_offset
398 This is the position the string must appear at, or not before.
399 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
400 characters must match before the string we are searching for.
401 Likewise when combined with minlenp and the length of the string it
538e84ed 402 tells us how many characters must appear after the string we have
1de06328 403 found.
538e84ed 404
1de06328
YO
405 - max_offset
406 Only used for floating strings. This is the rightmost point that
ea3daa5d 407 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 408 string can occur infinitely far to the right.
538e84ed 409
1de06328 410 - minlenp
2d608413
KW
411 A pointer to the minimum number of characters of the pattern that the
412 string was found inside. This is important as in the case of positive
538e84ed 413 lookahead or positive lookbehind we can have multiple patterns
1de06328 414 involved. Consider
538e84ed 415
1de06328 416 /(?=FOO).*F/
538e84ed 417
1de06328
YO
418 The minimum length of the pattern overall is 3, the minimum length
419 of the lookahead part is 3, but the minimum length of the part that
538e84ed 420 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 421 minimum length for the F is 1. This is important as the minimum length
538e84ed 422 is used to determine offsets in front of and behind the string being
1de06328 423 looked for. Since strings can be composites this is the length of the
486ec47a 424 pattern at the time it was committed with a scan_commit. Note that
1de06328 425 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 426 are not known until the full pattern has been compiled, thus the
1de06328 427 pointer to the value.
538e84ed 428
1de06328 429 - lookbehind
538e84ed 430
1de06328 431 In the case of lookbehind the string being searched for can be
538e84ed 432 offset past the start point of the final matching string.
1de06328
YO
433 If this value was just blithely removed from the min_offset it would
434 invalidate some of the calculations for how many chars must match
435 before or after (as they are derived from min_offset and minlen and
538e84ed 436 the length of the string being searched for).
1de06328
YO
437 When the final pattern is compiled and the data is moved from the
438 scan_data_t structure into the regexp structure the information
538e84ed
KW
439 about lookbehind is factored in, with the information that would
440 have been lost precalculated in the end_shift field for the
1de06328
YO
441 associated string.
442
443 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 444 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
445
446*/
2c2d71f5
JH
447
448typedef struct scan_data_t {
1de06328
YO
449 /*I32 len_min; unused */
450 /*I32 len_delta; unused */
49f55535 451 SSize_t pos_min;
ea3daa5d 452 SSize_t pos_delta;
2c2d71f5 453 SV *last_found;
ea3daa5d 454 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 455 SSize_t last_start_min;
ea3daa5d 456 SSize_t last_start_max;
1de06328
YO
457 SV **longest; /* Either &l_fixed, or &l_float. */
458 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 459 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 460 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
461 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
462 SV *longest_float; /* longest floating string found in pattern */
49f55535 463 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
464 SSize_t offset_float_max; /* latest point in string it can appear */
465 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 466 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
467 I32 flags;
468 I32 whilem_c;
49f55535 469 SSize_t *last_closep;
b8f7bb16 470 regnode_ssc *start_class;
2c2d71f5
JH
471} scan_data_t;
472
a687059c 473/*
e50aee73 474 * Forward declarations for pregcomp()'s friends.
a687059c 475 */
a0d0e21e 476
27da23d5 477static const scan_data_t zero_scan_data =
1de06328 478 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
479
480#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
481#define SF_BEFORE_SEOL 0x0001
482#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
483#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
484#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
485
44e3dfd2
BF
486#define SF_FIX_SHIFT_EOL (+2)
487#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
488
489#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
490#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
491
492#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
493#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
494#define SF_IS_INF 0x0040
495#define SF_HAS_PAR 0x0080
496#define SF_IN_PAR 0x0100
497#define SF_HAS_EVAL 0x0200
bc604ad8
KW
498
499
500/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
501 * longest substring in the pattern. When it is not set the optimiser keeps
502 * track of position, but does not keep track of the actual strings seen,
503 *
504 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
505 * /foo/i will not.
506 *
507 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
508 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
509 * turned off because of the alternation (BRANCH). */
07be1b83 510#define SCF_DO_SUBSTR 0x0400
bc604ad8 511
653099ff
GS
512#define SCF_DO_STCLASS_AND 0x0800
513#define SCF_DO_STCLASS_OR 0x1000
514#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 515#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 516
786e8c11 517#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 518#define SCF_SEEN_ACCEPT 0x8000
688e0391 519#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
520#define SCF_IN_DEFINE 0x20000
521
522
523
07be1b83 524
43fead97 525#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
526
527/* The enums for all these are ordered so things work out correctly */
a62b1201 528#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
529#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
530 == REGEX_DEPENDS_CHARSET)
00b27cfc 531#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
532#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
533 >= REGEX_UNICODE_CHARSET)
534#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
535 == REGEX_ASCII_RESTRICTED_CHARSET)
536#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
537 >= REGEX_ASCII_RESTRICTED_CHARSET)
538#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
539 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 540
43fead97 541#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 542
f2c2a6ab
KW
543/* For programs that want to be strictly Unicode compatible by dying if any
544 * attempt is made to match a non-Unicode code point against a Unicode
545 * property. */
546#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
547
93733859 548#define OOB_NAMEDCLASS -1
b8c5462f 549
8e661ac5
KW
550/* There is no code point that is out-of-bounds, so this is problematic. But
551 * its only current use is to initialize a variable that is always set before
552 * looked at. */
553#define OOB_UNICODE 0xDEADBEEF
554
a0ed51b3 555#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 556
8615cb43 557
b45f050a
JF
558/* length of regex to show in messages that don't mark a position within */
559#define RegexLengthToShowInErrorMessages 127
560
561/*
562 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
563 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
564 * op/pragma/warn/regcomp.
565 */
7253e4e3
RK
566#define MARKER1 "<-- HERE" /* marker as it appears in the description */
567#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 568
538e84ed 569#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 570 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 571
285b5ca0
KW
572/* The code in this file in places uses one level of recursion with parsing
573 * rebased to an alternate string constructed by us in memory. This can take
574 * the form of something that is completely different from the input, or
575 * something that uses the input as part of the alternate. In the first case,
576 * there should be no possibility of an error, as we are in complete control of
577 * the alternate string. But in the second case we don't control the input
578 * portion, so there may be errors in that. Here's an example:
579 * /[abc\x{DF}def]/ui
580 * is handled specially because \x{df} folds to a sequence of more than one
581 * character, 'ss'. What is done is to create and parse an alternate string,
582 * which looks like this:
583 * /(?:\x{DF}|[abc\x{DF}def])/ui
584 * where it uses the input unchanged in the middle of something it constructs,
585 * which is a branch for the DF outside the character class, and clustering
586 * parens around the whole thing. (It knows enough to skip the DF inside the
587 * class while in this substitute parse.) 'abc' and 'def' may have errors that
588 * need to be reported. The general situation looks like this:
589 *
590 * sI tI xI eI
591 * Input: ----------------------------------------------------
592 * Constructed: ---------------------------------------------------
593 * sC tC xC eC EC
594 *
595 * The input string sI..eI is the input pattern. The string sC..EC is the
596 * constructed substitute parse string. The portions sC..tC and eC..EC are
597 * constructed by us. The portion tC..eC is an exact duplicate of the input
598 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
599 * while parsing, we find an error at xC. We want to display a message showing
600 * the real input string. Thus we need to find the point xI in it which
601 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
602 * been constructed by us, and so shouldn't have errors. We get:
603 *
604 * xI = sI + (tI - sI) + (xC - tC)
605 *
606 * and, the offset into sI is:
607 *
608 * (xI - sI) = (tI - sI) + (xC - tC)
609 *
610 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
611 * and we save tC as RExC_adjusted_start.
903c858a
KW
612 *
613 * During normal processing of the input pattern, everything points to that,
614 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
285b5ca0
KW
615 */
616
617#define tI_sI RExC_precomp_adj
618#define tC RExC_adjusted_start
619#define sC RExC_precomp
620#define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
621#define xI(xC) (sC + xI_offset(xC))
622#define eC RExC_precomp_end
623
624#define REPORT_LOCATION_ARGS(xC) \
625 UTF8fARG(UTF, \
626 (xI(xC) > eC) /* Don't run off end */ \
627 ? eC - sC /* Length before the <--HERE */ \
628 : xI_offset(xC), \
629 sC), /* The input pattern printed up to the <--HERE */ \
630 UTF8fARG(UTF, \
631 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
632 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
c1d900c3 633
8a6d8ec6
HS
634/* Used to point after bad bytes for an error message, but avoid skipping
635 * past a nul byte. */
636#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
637
b45f050a
JF
638/*
639 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
640 * arg. Show regex, up to a maximum length. If it's too long, chop and add
641 * "...".
642 */
58e23c8d 643#define _FAIL(code) STMT_START { \
bfed75c6 644 const char *ellipses = ""; \
711b303b 645 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
646 \
647 if (!SIZE_ONLY) \
a5e7bc51 648 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
649 if (len > RegexLengthToShowInErrorMessages) { \
650 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
651 len = RegexLengthToShowInErrorMessages - 10; \
652 ellipses = "..."; \
653 } \
58e23c8d 654 code; \
ccb2c380 655} STMT_END
8615cb43 656
58e23c8d 657#define FAIL(msg) _FAIL( \
147e3846 658 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 659 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
660
661#define FAIL2(msg,arg) _FAIL( \
147e3846 662 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 663 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 664
b45f050a 665/*
b45f050a
JF
666 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
667 */
ccb2c380 668#define Simple_vFAIL(m) STMT_START { \
ccb2c380 669 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 670 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 671} STMT_END
b45f050a
JF
672
673/*
674 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
675 */
ccb2c380
MP
676#define vFAIL(m) STMT_START { \
677 if (!SIZE_ONLY) \
a5e7bc51 678 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
679 Simple_vFAIL(m); \
680} STMT_END
b45f050a
JF
681
682/*
683 * Like Simple_vFAIL(), but accepts two arguments.
684 */
ccb2c380 685#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
686 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
687 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 688} STMT_END
b45f050a
JF
689
690/*
691 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
692 */
ccb2c380
MP
693#define vFAIL2(m,a1) STMT_START { \
694 if (!SIZE_ONLY) \
a5e7bc51 695 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
696 Simple_vFAIL2(m, a1); \
697} STMT_END
b45f050a
JF
698
699
700/*
701 * Like Simple_vFAIL(), but accepts three arguments.
702 */
ccb2c380 703#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 704 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 705 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 706} STMT_END
b45f050a
JF
707
708/*
709 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
710 */
ccb2c380
MP
711#define vFAIL3(m,a1,a2) STMT_START { \
712 if (!SIZE_ONLY) \
a5e7bc51 713 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
714 Simple_vFAIL3(m, a1, a2); \
715} STMT_END
b45f050a
JF
716
717/*
718 * Like Simple_vFAIL(), but accepts four arguments.
719 */
ccb2c380 720#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
721 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
722 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 723} STMT_END
b45f050a 724
95db3ffa
KW
725#define vFAIL4(m,a1,a2,a3) STMT_START { \
726 if (!SIZE_ONLY) \
727 SAVEFREESV(RExC_rx_sv); \
728 Simple_vFAIL4(m, a1, a2, a3); \
729} STMT_END
730
946095af 731/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
732#define vFAIL2utf8f(m, a1) STMT_START { \
733 if (!SIZE_ONLY) \
734 SAVEFREESV(RExC_rx_sv); \
735 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
736 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
737} STMT_END
738
3ba22297 739#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
740 if (!SIZE_ONLY) \
741 SAVEFREESV(RExC_rx_sv); \
742 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 743 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
744} STMT_END
745
499333dc
KW
746/* These have asserts in them because of [perl #122671] Many warnings in
747 * regcomp.c can occur twice. If they get output in pass1 and later in that
748 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
749 * would get output again. So they should be output in pass2, and these
750 * asserts make sure new warnings follow that paradigm. */
946095af 751
5e0a247b
KW
752/* m is not necessarily a "literal string", in this macro */
753#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
754 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
755 "%s" REPORT_LOCATION, \
756 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
757} STMT_END
758
668c081a 759#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
760 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
761 m REPORT_LOCATION, \
762 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
763} STMT_END
764
b927b7e9 765#define vWARN(loc, m) STMT_START { \
d528642a
KW
766 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
767 m REPORT_LOCATION, \
768 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
769} STMT_END
770
0d6106aa 771#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
772 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
773 m REPORT_LOCATION, \
774 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
775} STMT_END
776
147508a2 777#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
778 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
779 m REPORT_LOCATION, \
780 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
781} STMT_END
782
d528642a
KW
783#define ckWARNregdep(loc,m) STMT_START { \
784 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
785 WARN_REGEXP), \
786 m REPORT_LOCATION, \
787 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
788} STMT_END
789
d528642a
KW
790#define ckWARN2reg_d(loc,m, a1) STMT_START { \
791 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
792 m REPORT_LOCATION, \
793 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
794} STMT_END
795
d528642a
KW
796#define ckWARN2reg(loc, m, a1) STMT_START { \
797 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
798 m REPORT_LOCATION, \
799 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
800} STMT_END
801
d528642a
KW
802#define vWARN3(loc, m, a1, a2) STMT_START { \
803 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
804 m REPORT_LOCATION, \
805 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
806} STMT_END
807
d528642a
KW
808#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
809 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
810 m REPORT_LOCATION, \
811 a1, a2, \
812 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
813} STMT_END
814
ccb2c380 815#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
816 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
817 m REPORT_LOCATION, \
818 a1, a2, a3, \
819 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
820} STMT_END
821
668c081a 822#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
823 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
824 m REPORT_LOCATION, \
825 a1, a2, a3, \
826 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
827} STMT_END
828
ccb2c380 829#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
830 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
831 m REPORT_LOCATION, \
832 a1, a2, a3, a4, \
833 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 834} STMT_END
9d1d55b5 835
538e84ed 836/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
837 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
838 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
839 * Element 0 holds the number n.
07be1b83 840 * Position is 1 indexed.
fac92740 841 */
7122b237
YO
842#ifndef RE_TRACK_PATTERN_OFFSETS
843#define Set_Node_Offset_To_R(node,byte)
844#define Set_Node_Offset(node,byte)
845#define Set_Cur_Node_Offset
846#define Set_Node_Length_To_R(node,len)
847#define Set_Node_Length(node,len)
6a86c6ad 848#define Set_Node_Cur_Length(node,start)
538e84ed
KW
849#define Node_Offset(n)
850#define Node_Length(n)
7122b237
YO
851#define Set_Node_Offset_Length(node,offset,len)
852#define ProgLen(ri) ri->u.proglen
853#define SetProgLen(ri,x) ri->u.proglen = x
854#else
855#define ProgLen(ri) ri->u.offsets[0]
856#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
857#define Set_Node_Offset_To_R(node,byte) STMT_START { \
858 if (! SIZE_ONLY) { \
859 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 860 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 861 if((node) < 0) { \
538e84ed
KW
862 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
863 (int)(node)); \
ccb2c380
MP
864 } else { \
865 RExC_offsets[2*(node)-1] = (byte); \
866 } \
867 } \
868} STMT_END
869
870#define Set_Node_Offset(node,byte) \
871 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
872#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
873
874#define Set_Node_Length_To_R(node,len) STMT_START { \
875 if (! SIZE_ONLY) { \
876 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 877 __LINE__, (int)(node), (int)(len))); \
ccb2c380 878 if((node) < 0) { \
538e84ed
KW
879 Perl_croak(aTHX_ "value of node is %d in Length macro", \
880 (int)(node)); \
ccb2c380
MP
881 } else { \
882 RExC_offsets[2*(node)] = (len); \
883 } \
884 } \
885} STMT_END
886
887#define Set_Node_Length(node,len) \
888 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
889#define Set_Node_Cur_Length(node, start) \
890 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
891
892/* Get offsets and lengths */
893#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
894#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
895
07be1b83
YO
896#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
897 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
898 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
899} STMT_END
7122b237 900#endif
07be1b83
YO
901
902#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
903#define EXPERIMENTAL_INPLACESCAN
f427392e 904#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 905
cb41e5d6
YO
906#ifdef DEBUGGING
907int
6ad9a8ab 908Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
909{
910 va_list ap;
911 int result;
912 PerlIO *f= Perl_debug_log;
913 PERL_ARGS_ASSERT_RE_PRINTF;
914 va_start(ap, fmt);
915 result = PerlIO_vprintf(f, fmt, ap);
916 va_end(ap);
917 return result;
918}
919
920int
7b031478 921Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
922{
923 va_list ap;
924 int result;
925 PerlIO *f= Perl_debug_log;
926 PERL_ARGS_ASSERT_RE_INDENTF;
927 va_start(ap, depth);
daeb874b 928 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
929 result = PerlIO_vprintf(f, fmt, ap);
930 va_end(ap);
931 return result;
932}
933#endif /* DEBUGGING */
934
7b031478 935#define DEBUG_RExC_seen() \
538e84ed 936 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 937 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 938 \
e384d5c1 939 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 940 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 941 \
e384d5c1 942 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 943 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 944 \
e384d5c1 945 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 946 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 947 \
e384d5c1 948 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 949 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 950 \
7b031478 951 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 952 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 953 \
e384d5c1 954 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 955 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 956 \
e384d5c1 957 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 958 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 959 \
e384d5c1 960 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 961 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 962 \
e384d5c1 963 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 964 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 965 \
7b031478 966 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 967 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 968 \
6ad9a8ab 969 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
970 });
971
fdfb4f21 972#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 973 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21
YO
974
975#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
976 if ( ( flags ) ) { \
6ad9a8ab 977 Perl_re_printf( aTHX_ "%s", open_str); \
fdfb4f21
YO
978 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
991 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
992 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
6ad9a8ab 993 Perl_re_printf( aTHX_ "%s", close_str); \
fdfb4f21
YO
994 }
995
996
304ee84b
YO
997#define DEBUG_STUDYDATA(str,data,depth) \
998DEBUG_OPTIMISE_MORE_r(if(data){ \
147e3846
KW
999 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
1000 " Flags: 0x%" UVXf, \
7b031478 1001 depth, \
1de06328
YO
1002 (IV)((data)->pos_min), \
1003 (IV)((data)->pos_delta), \
fdfb4f21
YO
1004 (UV)((data)->flags) \
1005 ); \
1006 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
147e3846
KW
1007 Perl_re_printf( aTHX_ \
1008 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1de06328 1009 (IV)((data)->whilem_c), \
304ee84b
YO
1010 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1011 is_inf ? "INF " : "" \
1de06328
YO
1012 ); \
1013 if ((data)->last_found) \
147e3846
KW
1014 Perl_re_printf( aTHX_ \
1015 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1016 " %sFixed:'%s' @ %" IVdf \
1017 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1de06328
YO
1018 SvPVX_const((data)->last_found), \
1019 (IV)((data)->last_end), \
1020 (IV)((data)->last_start_min), \
1021 (IV)((data)->last_start_max), \
1022 ((data)->longest && \
1023 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1024 SvPVX_const((data)->longest_fixed), \
1025 (IV)((data)->offset_fixed), \
1026 ((data)->longest && \
1027 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1028 SvPVX_const((data)->longest_float), \
1029 (IV)((data)->offset_float_min), \
1030 (IV)((data)->offset_float_max) \
1031 ); \
6ad9a8ab 1032 Perl_re_printf( aTHX_ "\n"); \
1de06328
YO
1033});
1034
cb41e5d6 1035
c6871b76
KW
1036/* =========================================================
1037 * BEGIN edit_distance stuff.
1038 *
1039 * This calculates how many single character changes of any type are needed to
1040 * transform a string into another one. It is taken from version 3.1 of
1041 *
1042 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1043 */
1044
1045/* Our unsorted dictionary linked list. */
1046/* Note we use UVs, not chars. */
1047
1048struct dictionary{
1049 UV key;
1050 UV value;
1051 struct dictionary* next;
1052};
1053typedef struct dictionary item;
1054
1055
1056PERL_STATIC_INLINE item*
1057push(UV key,item* curr)
1058{
1059 item* head;
1060 Newxz(head, 1, item);
1061 head->key = key;
1062 head->value = 0;
1063 head->next = curr;
1064 return head;
1065}
1066
1067
1068PERL_STATIC_INLINE item*
1069find(item* head, UV key)
1070{
1071 item* iterator = head;
1072 while (iterator){
1073 if (iterator->key == key){
1074 return iterator;
1075 }
1076 iterator = iterator->next;
1077 }
1078
1079 return NULL;
1080}
1081
1082PERL_STATIC_INLINE item*
1083uniquePush(item* head,UV key)
1084{
1085 item* iterator = head;
1086
1087 while (iterator){
1088 if (iterator->key == key) {
1089 return head;
1090 }
1091 iterator = iterator->next;
1092 }
1093
1094 return push(key,head);
1095}
1096
1097PERL_STATIC_INLINE void
1098dict_free(item* head)
1099{
1100 item* iterator = head;
1101
1102 while (iterator) {
1103 item* temp = iterator;
1104 iterator = iterator->next;
1105 Safefree(temp);
1106 }
1107
1108 head = NULL;
1109}
1110
1111/* End of Dictionary Stuff */
1112
1113/* All calculations/work are done here */
1114STATIC int
1115S_edit_distance(const UV* src,
1116 const UV* tgt,
1117 const STRLEN x, /* length of src[] */
1118 const STRLEN y, /* length of tgt[] */
1119 const SSize_t maxDistance
1120)
1121{
1122 item *head = NULL;
1123 UV swapCount,swapScore,targetCharCount,i,j;
1124 UV *scores;
1125 UV score_ceil = x + y;
1126
1127 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1128
1129 /* intialize matrix start values */
1130 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1131 scores[0] = score_ceil;
1132 scores[1 * (y + 2) + 0] = score_ceil;
1133 scores[0 * (y + 2) + 1] = score_ceil;
1134 scores[1 * (y + 2) + 1] = 0;
1135 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1136
1137 /* work loops */
1138 /* i = src index */
1139 /* j = tgt index */
1140 for (i=1;i<=x;i++) {
1141 if (i < x)
1142 head = uniquePush(head,src[i]);
1143 scores[(i+1) * (y + 2) + 1] = i;
1144 scores[(i+1) * (y + 2) + 0] = score_ceil;
1145 swapCount = 0;
1146
1147 for (j=1;j<=y;j++) {
1148 if (i == 1) {
1149 if(j < y)
1150 head = uniquePush(head,tgt[j]);
1151 scores[1 * (y + 2) + (j + 1)] = j;
1152 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1153 }
1154
1155 targetCharCount = find(head,tgt[j-1])->value;
1156 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1157
1158 if (src[i-1] != tgt[j-1]){
1159 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));
1160 }
1161 else {
1162 swapCount = j;
1163 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1164 }
1165 }
1166
1167 find(head,src[i-1])->value = i;
1168 }
1169
1170 {
1171 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1172 dict_free(head);
1173 Safefree(scores);
1174 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1175 }
1176}
1177
1178/* END of edit_distance() stuff
1179 * ========================================================= */
1180
8e35b056
KW
1181/* is c a control character for which we have a mnemonic? */
1182#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1183
549b4e78
KW
1184STATIC const char *
1185S_cntrl_to_mnemonic(const U8 c)
1186{
1187 /* Returns the mnemonic string that represents character 'c', if one
1188 * exists; NULL otherwise. The only ones that exist for the purposes of
1189 * this routine are a few control characters */
1190
1191 switch (c) {
1192 case '\a': return "\\a";
1193 case '\b': return "\\b";
1194 case ESC_NATIVE: return "\\e";
1195 case '\f': return "\\f";
1196 case '\n': return "\\n";
1197 case '\r': return "\\r";
1198 case '\t': return "\\t";
1199 }
1200
1201 return NULL;
1202}
1203
653099ff 1204/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 1205 Update the longest found anchored substring and the longest found
653099ff
GS
1206 floating substrings if needed. */
1207
4327152a 1208STATIC void
ea3daa5d
FC
1209S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1210 SSize_t *minlenp, int is_inf)
c277df42 1211{
e1ec3a88
AL
1212 const STRLEN l = CHR_SVLEN(data->last_found);
1213 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 1214 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1215
7918f24d
NC
1216 PERL_ARGS_ASSERT_SCAN_COMMIT;
1217
c277df42 1218 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 1219 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
1220 if (*data->longest == data->longest_fixed) {
1221 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1222 if (data->flags & SF_BEFORE_EOL)
b81d288d 1223 data->flags
c277df42
IZ
1224 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1225 else
1226 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 1227 data->minlen_fixed=minlenp;
1de06328 1228 data->lookbehind_fixed=0;
a0ed51b3 1229 }
304ee84b 1230 else { /* *data->longest == data->longest_float */
c277df42 1231 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 1232 data->offset_float_max = (l
646e8787
DM
1233 ? data->last_start_max
1234 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1235 ? SSize_t_MAX
1236 : data->pos_min + data->pos_delta));
1237 if (is_inf
1238 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1239 data->offset_float_max = SSize_t_MAX;
c277df42 1240 if (data->flags & SF_BEFORE_EOL)
b81d288d 1241 data->flags
c277df42
IZ
1242 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1243 else
1244 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
1245 data->minlen_float=minlenp;
1246 data->lookbehind_float=0;
c277df42
IZ
1247 }
1248 }
1249 SvCUR_set(data->last_found, 0);
0eda9292 1250 {
a28509cc 1251 SV * const sv = data->last_found;
097eb12c
AL
1252 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1253 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1254 if (mg)
1255 mg->mg_len = 0;
1256 }
0eda9292 1257 }
c277df42
IZ
1258 data->last_end = -1;
1259 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1260 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1261}
1262
cdd87c1d
KW
1263/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1264 * list that describes which code points it matches */
1265
653099ff 1266STATIC void
3420edd7 1267S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1268{
cdd87c1d
KW
1269 /* Set the SSC 'ssc' to match an empty string or any code point */
1270
557bd3fb 1271 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1272
71068078 1273 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1274
0854ea0b
KW
1275 /* mortalize so won't leak */
1276 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1277 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1278}
1279
653099ff 1280STATIC int
dc3bf405 1281S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1282{
c144baaa
KW
1283 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1284 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1285 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1286 * in any way, so there's no point in using it */
cdd87c1d
KW
1287
1288 UV start, end;
1289 bool ret;
653099ff 1290
557bd3fb 1291 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1292
71068078 1293 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1294
93e92956 1295 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1296 return FALSE;
1297 }
1298
1299 /* See if the list consists solely of the range 0 - Infinity */
1300 invlist_iterinit(ssc->invlist);
1301 ret = invlist_iternext(ssc->invlist, &start, &end)
1302 && start == 0
1303 && end == UV_MAX;
1304
1305 invlist_iterfinish(ssc->invlist);
1306
1307 if (ret) {
1308 return TRUE;
1309 }
1310
1311 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1312 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1313 int i;
1314 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1315 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1316 return TRUE;
1317 }
1318 }
1319 }
1320
1321 return FALSE;
653099ff
GS
1322}
1323
653099ff 1324STATIC void
cdd87c1d 1325S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1326{
cdd87c1d
KW
1327 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1328 * string, any code point, or any posix class under locale */
1329
557bd3fb 1330 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1331
557bd3fb 1332 Zero(ssc, 1, regnode_ssc);
71068078 1333 set_ANYOF_SYNTHETIC(ssc);
93e92956 1334 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1335 ssc_anything(ssc);
cdd87c1d 1336
2f306ab9
KW
1337 /* If any portion of the regex is to operate under locale rules that aren't
1338 * fully known at compile time, initialization includes it. The reason
1339 * this isn't done for all regexes is that the optimizer was written under
1340 * the assumption that locale was all-or-nothing. Given the complexity and
1341 * lack of documentation in the optimizer, and that there are inadequate
1342 * test cases for locale, many parts of it may not work properly, it is
1343 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1344 if (RExC_contains_locale) {
1345 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1346 }
1347 else {
1348 ANYOF_POSIXL_ZERO(ssc);
1349 }
653099ff
GS
1350}
1351
b423522f 1352STATIC int
dc3bf405
BF
1353S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1354 const regnode_ssc *ssc)
b423522f
KW
1355{
1356 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1357 * to the list of code points matched, and locale posix classes; hence does
1358 * not check its flags) */
1359
1360 UV start, end;
1361 bool ret;
1362
1363 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1364
71068078 1365 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1366
1367 invlist_iterinit(ssc->invlist);
1368 ret = invlist_iternext(ssc->invlist, &start, &end)
1369 && start == 0
1370 && end == UV_MAX;
1371
1372 invlist_iterfinish(ssc->invlist);
1373
1374 if (! ret) {
1375 return FALSE;
1376 }
1377
e0e1be5f 1378 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1379 return FALSE;
b423522f
KW
1380 }
1381
1382 return TRUE;
1383}
1384
1385STATIC SV*
1386S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1387 const regnode_charclass* const node)
b423522f
KW
1388{
1389 /* Returns a mortal inversion list defining which code points are matched
1390 * by 'node', which is of type ANYOF. Handles complementing the result if
1391 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1392 * returned list must, and will, contain every code point that is a
1393 * possibility. */
b423522f 1394
e2506fa7 1395 SV* invlist = NULL;
1ee208c4 1396 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1397 unsigned int i;
1398 const U32 n = ARG(node);
31f05a37 1399 bool new_node_has_latin1 = FALSE;
b423522f
KW
1400
1401 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1402
1403 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1404 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1405 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1406 AV * const av = MUTABLE_AV(SvRV(rv));
1407 SV **const ary = AvARRAY(av);
1408 assert(RExC_rxi->data->what[n] == 's');
1409
1410 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1411 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1412 }
1413 else if (ary[0] && ary[0] != &PL_sv_undef) {
1414
1415 /* Here, no compile-time swash, and there are things that won't be
1416 * known until runtime -- we have to assume it could be anything */
e2506fa7 1417 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1418 return _add_range_to_invlist(invlist, 0, UV_MAX);
1419 }
1ee208c4 1420 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1421
1422 /* Here no compile-time swash, and no run-time only data. Use the
1423 * node's inversion list */
1ee208c4
KW
1424 invlist = sv_2mortal(invlist_clone(ary[3]));
1425 }
1426
1427 /* Get the code points valid only under UTF-8 locales */
037715a6 1428 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1429 && ary[2] && ary[2] != &PL_sv_undef)
1430 {
1431 only_utf8_locale_invlist = ary[2];
b423522f
KW
1432 }
1433 }
1434
e2506fa7
KW
1435 if (! invlist) {
1436 invlist = sv_2mortal(_new_invlist(0));
1437 }
1438
dcb20b36
KW
1439 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1440 * code points, and an inversion list for the others, but if there are code
1441 * points that should match only conditionally on the target string being
1442 * UTF-8, those are placed in the inversion list, and not the bitmap.
1443 * Since there are circumstances under which they could match, they are
1444 * included in the SSC. But if the ANYOF node is to be inverted, we have
1445 * to exclude them here, so that when we invert below, the end result
1446 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1447 * have to do this here before we add the unconditionally matched code
1448 * points */
b423522f
KW
1449 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1450 _invlist_intersection_complement_2nd(invlist,
1451 PL_UpperLatin1,
1452 &invlist);
1453 }
1454
1455 /* Add in the points from the bit map */
dcb20b36 1456 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1457 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1458 unsigned int start = i++;
1459
1460 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1461 /* empty */
1462 }
1463 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1464 new_node_has_latin1 = TRUE;
b423522f
KW
1465 }
1466 }
1467
1468 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1469 * as well. But don't add them if inverting, as when that gets done below,
1470 * it would exclude all these characters, including the ones it shouldn't
1471 * that were added just above */
1472 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1473 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1474 {
b423522f
KW
1475 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1476 }
1477
1478 /* Similarly for these */
93e92956 1479 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1480 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1481 }
1482
1483 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1484 _invlist_invert(invlist);
1485 }
037715a6 1486 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1487
1488 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1489 * locale. We can skip this if there are no 0-255 at all. */
1490 _invlist_union(invlist, PL_Latin1, &invlist);
1491 }
1492
1ee208c4
KW
1493 /* Similarly add the UTF-8 locale possible matches. These have to be
1494 * deferred until after the non-UTF-8 locale ones are taken care of just
1495 * above, or it leads to wrong results under ANYOF_INVERT */
1496 if (only_utf8_locale_invlist) {
31f05a37 1497 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1498 only_utf8_locale_invlist,
31f05a37
KW
1499 ANYOF_FLAGS(node) & ANYOF_INVERT,
1500 &invlist);
1501 }
b423522f
KW
1502
1503 return invlist;
1504}
1505
1051e1c4 1506/* These two functions currently do the exact same thing */
557bd3fb 1507#define ssc_init_zero ssc_init
653099ff 1508
cdd87c1d
KW
1509#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1510#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1511
557bd3fb 1512/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1513 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1514 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1515
653099ff 1516STATIC void
b423522f 1517S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1518 const regnode_charclass *and_with)
653099ff 1519{
cdd87c1d
KW
1520 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1521 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1522
a0dd4231
KW
1523 SV* anded_cp_list;
1524 U8 anded_flags;
1e6ade67 1525
cdd87c1d 1526 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1527
71068078 1528 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1529
1530 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1531 * the code point inversion list and just the relevant flags */
71068078 1532 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1533 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1534 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1535
1536 /* XXX This is a kludge around what appears to be deficiencies in the
1537 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1538 * there are paths through the optimizer where it doesn't get weeded
1539 * out when it should. And if we don't make some extra provision for
1540 * it like the code just below, it doesn't get added when it should.
1541 * This solution is to add it only when AND'ing, which is here, and
1542 * only when what is being AND'ed is the pristine, original node
1543 * matching anything. Thus it is like adding it to ssc_anything() but
1544 * only when the result is to be AND'ed. Probably the same solution
1545 * could be adopted for the same problem we have with /l matching,
1546 * which is solved differently in S_ssc_init(), and that would lead to
1547 * fewer false positives than that solution has. But if this solution
1548 * creates bugs, the consequences are only that a warning isn't raised
1549 * that should be; while the consequences for having /l bugs is
1550 * incorrect matches */
7dcac5f6 1551 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1552 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1553 }
a0dd4231
KW
1554 }
1555 else {
5c0f85ef 1556 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1557 if (OP(and_with) == ANYOFD) {
1558 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1559 }
1560 else {
1561 anded_flags = ANYOF_FLAGS(and_with)
1562 &( ANYOF_COMMON_FLAGS
108316fb
KW
1563 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1564 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1565 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1566 anded_flags &=
1567 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1568 }
f240c685 1569 }
a0dd4231
KW
1570 }
1571
1572 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1573
1574 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1575 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1576 * 'and_with' may be inverted. When not inverted, we have the situation of
1577 * computing:
1578 * (C1 | P1) & (C2 | P2)
1579 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1580 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1581 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1582 * <= ((C1 & C2) | P1 | P2)
1583 * Alternatively, the last few steps could be:
1584 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1585 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1586 * <= (C1 | C2 | (P1 & P2))
1587 * We favor the second approach if either P1 or P2 is non-empty. This is
1588 * because these components are a barrier to doing optimizations, as what
1589 * they match cannot be known until the moment of matching as they are
1590 * dependent on the current locale, 'AND"ing them likely will reduce or
1591 * eliminate them.
1592 * But we can do better if we know that C1,P1 are in their initial state (a
1593 * frequent occurrence), each matching everything:
1594 * (<everything>) & (C2 | P2) = C2 | P2
1595 * Similarly, if C2,P2 are in their initial state (again a frequent
1596 * occurrence), the result is a no-op
1597 * (C1 | P1) & (<everything>) = C1 | P1
1598 *
1599 * Inverted, we have
1600 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1601 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1602 * <= (C1 & ~C2) | (P1 & ~P2)
1603 * */
1aa99e6b 1604
a0dd4231 1605 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1606 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1607 {
cdd87c1d 1608 unsigned int i;
8951c461 1609
cdd87c1d
KW
1610 ssc_intersection(ssc,
1611 anded_cp_list,
1612 FALSE /* Has already been inverted */
1613 );
c6b76537 1614
cdd87c1d
KW
1615 /* If either P1 or P2 is empty, the intersection will be also; can skip
1616 * the loop */
93e92956 1617 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1618 ANYOF_POSIXL_ZERO(ssc);
1619 }
e0e1be5f 1620 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1621
1622 /* Note that the Posix class component P from 'and_with' actually
1623 * looks like:
1624 * P = Pa | Pb | ... | Pn
1625 * where each component is one posix class, such as in [\w\s].
1626 * Thus
1627 * ~P = ~(Pa | Pb | ... | Pn)
1628 * = ~Pa & ~Pb & ... & ~Pn
1629 * <= ~Pa | ~Pb | ... | ~Pn
1630 * The last is something we can easily calculate, but unfortunately
1631 * is likely to have many false positives. We could do better
1632 * in some (but certainly not all) instances if two classes in
1633 * P have known relationships. For example
1634 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1635 * So
1636 * :lower: & :print: = :lower:
1637 * And similarly for classes that must be disjoint. For example,
1638 * since \s and \w can have no elements in common based on rules in
1639 * the POSIX standard,
1640 * \w & ^\S = nothing
1641 * Unfortunately, some vendor locales do not meet the Posix
1642 * standard, in particular almost everything by Microsoft.
1643 * The loop below just changes e.g., \w into \W and vice versa */
1644
1ee208c4 1645 regnode_charclass_posixl temp;
cdd87c1d
KW
1646 int add = 1; /* To calculate the index of the complement */
1647
1648 ANYOF_POSIXL_ZERO(&temp);
1649 for (i = 0; i < ANYOF_MAX; i++) {
1650 assert(i % 2 != 0
7dcac5f6
KW
1651 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1652 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1653
7dcac5f6 1654 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1655 ANYOF_POSIXL_SET(&temp, i + add);
1656 }
1657 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1658 }
1659 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1660
cdd87c1d
KW
1661 } /* else ssc already has no posixes */
1662 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1663 in its initial state */
71068078 1664 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1665 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1666 {
1667 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1668 * copy it over 'ssc' */
1669 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1670 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1671 StructCopy(and_with, ssc, regnode_ssc);
1672 }
1673 else {
1674 ssc->invlist = anded_cp_list;
1675 ANYOF_POSIXL_ZERO(ssc);
93e92956 1676 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1677 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1678 }
1679 }
1680 }
e0e1be5f 1681 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1682 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1683 {
1684 /* One or the other of P1, P2 is non-empty. */
93e92956 1685 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1686 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1687 }
cdd87c1d
KW
1688 ssc_union(ssc, anded_cp_list, FALSE);
1689 }
1690 else { /* P1 = P2 = empty */
1691 ssc_intersection(ssc, anded_cp_list, FALSE);
1692 }
137165a6 1693 }
653099ff
GS
1694}
1695
653099ff 1696STATIC void
cdd87c1d 1697S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1698 const regnode_charclass *or_with)
653099ff 1699{
cdd87c1d
KW
1700 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1701 * another SSC or a regular ANYOF class. Can create false positives if
1702 * 'or_with' is to be inverted. */
7918f24d 1703
a0dd4231
KW
1704 SV* ored_cp_list;
1705 U8 ored_flags;
c6b76537 1706
cdd87c1d 1707 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1708
71068078 1709 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1710
1711 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1712 * the code point inversion list and just the relevant flags */
71068078 1713 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1714 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1715 ored_flags = ANYOF_FLAGS(or_with);
1716 }
1717 else {
5c0f85ef 1718 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1719 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1720 if (OP(or_with) != ANYOFD) {
1721 ored_flags
1722 |= ANYOF_FLAGS(or_with)
108316fb
KW
1723 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1724 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1725 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1726 ored_flags |=
1727 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1728 }
f240c685 1729 }
a0dd4231
KW
1730 }
1731
1732 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1733
1734 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1735 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1736 * 'or_with' may be inverted. When not inverted, we have the simple
1737 * situation of computing:
1738 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1739 * If P1|P2 yields a situation with both a class and its complement are
1740 * set, like having both \w and \W, this matches all code points, and we
1741 * can delete these from the P component of the ssc going forward. XXX We
1742 * might be able to delete all the P components, but I (khw) am not certain
1743 * about this, and it is better to be safe.
1744 *
1745 * Inverted, we have
1746 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1747 * <= (C1 | P1) | ~C2
1748 * <= (C1 | ~C2) | P1
1749 * (which results in actually simpler code than the non-inverted case)
1750 * */
9826f543 1751
a0dd4231 1752 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1753 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1754 {
cdd87c1d 1755 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1756 } /* else Not inverted */
93e92956 1757 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1758 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1759 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1760 unsigned int i;
1761 for (i = 0; i < ANYOF_MAX; i += 2) {
1762 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1763 {
1764 ssc_match_all_cp(ssc);
1765 ANYOF_POSIXL_CLEAR(ssc, i);
1766 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1767 }
1768 }
1769 }
1aa99e6b 1770 }
cdd87c1d
KW
1771
1772 ssc_union(ssc,
1773 ored_cp_list,
1774 FALSE /* Already has been inverted */
1775 );
653099ff
GS
1776}
1777
b423522f
KW
1778PERL_STATIC_INLINE void
1779S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1780{
1781 PERL_ARGS_ASSERT_SSC_UNION;
1782
71068078 1783 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1784
1785 _invlist_union_maybe_complement_2nd(ssc->invlist,
1786 invlist,
1787 invert2nd,
1788 &ssc->invlist);
1789}
1790
1791PERL_STATIC_INLINE void
1792S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1793 SV* const invlist,
1794 const bool invert2nd)
1795{
1796 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1797
71068078 1798 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1799
1800 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1801 invlist,
1802 invert2nd,
1803 &ssc->invlist);
1804}
1805
1806PERL_STATIC_INLINE void
1807S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1808{
1809 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1810
71068078 1811 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1812
1813 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1814}
1815
1816PERL_STATIC_INLINE void
1817S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1818{
1819 /* AND just the single code point 'cp' into the SSC 'ssc' */
1820
1821 SV* cp_list = _new_invlist(2);
1822
1823 PERL_ARGS_ASSERT_SSC_CP_AND;
1824
71068078 1825 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1826
1827 cp_list = add_cp_to_invlist(cp_list, cp);
1828 ssc_intersection(ssc, cp_list,
1829 FALSE /* Not inverted */
1830 );
1831 SvREFCNT_dec_NN(cp_list);
1832}
1833
1834PERL_STATIC_INLINE void
dc3bf405 1835S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1836{
1837 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1838 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1839
71068078 1840 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1841
1842 ANYOF_POSIXL_ZERO(ssc);
1843 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1844}
1845
b35552de
KW
1846#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1847
1848STATIC bool
1849S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1850{
1851 /* The synthetic start class is used to hopefully quickly winnow down
1852 * places where a pattern could start a match in the target string. If it
1853 * doesn't really narrow things down that much, there isn't much point to
1854 * having the overhead of using it. This function uses some very crude
1855 * heuristics to decide if to use the ssc or not.
1856 *
1857 * It returns TRUE if 'ssc' rules out more than half what it considers to
1858 * be the "likely" possible matches, but of course it doesn't know what the
1859 * actual things being matched are going to be; these are only guesses
1860 *
1861 * For /l matches, it assumes that the only likely matches are going to be
1862 * in the 0-255 range, uniformly distributed, so half of that is 127
1863 * For /a and /d matches, it assumes that the likely matches will be just
1864 * the ASCII range, so half of that is 63
1865 * For /u and there isn't anything matching above the Latin1 range, it
1866 * assumes that that is the only range likely to be matched, and uses
1867 * half that as the cut-off: 127. If anything matches above Latin1,
1868 * it assumes that all of Unicode could match (uniformly), except for
1869 * non-Unicode code points and things in the General Category "Other"
1870 * (unassigned, private use, surrogates, controls and formats). This
1871 * is a much large number. */
1872
b35552de
KW
1873 U32 count = 0; /* Running total of number of code points matched by
1874 'ssc' */
1875 UV start, end; /* Start and end points of current range in inversion
1876 list */
72400949
KW
1877 const U32 max_code_points = (LOC)
1878 ? 256
1879 : (( ! UNI_SEMANTICS
1880 || invlist_highest(ssc->invlist) < 256)
1881 ? 128
1882 : NON_OTHER_COUNT);
1883 const U32 max_match = max_code_points / 2;
b35552de
KW
1884
1885 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1886
1887 invlist_iterinit(ssc->invlist);
1888 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1889 if (start >= max_code_points) {
1890 break;
b35552de 1891 }
72400949 1892 end = MIN(end, max_code_points - 1);
b35552de 1893 count += end - start + 1;
72400949 1894 if (count >= max_match) {
b35552de
KW
1895 invlist_iterfinish(ssc->invlist);
1896 return FALSE;
1897 }
1898 }
1899
1900 return TRUE;
1901}
1902
1903
b423522f
KW
1904STATIC void
1905S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1906{
1907 /* The inversion list in the SSC is marked mortal; now we need a more
1908 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1909 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1910 * map */
b423522f
KW
1911
1912 SV* invlist = invlist_clone(ssc->invlist);
1913
1914 PERL_ARGS_ASSERT_SSC_FINALIZE;
1915
71068078 1916 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1917
a0dd4231 1918 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1919 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1920 * by the time we reach here */
f240c685
KW
1921 assert(! (ANYOF_FLAGS(ssc)
1922 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
1923 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1924 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 1925
b423522f
KW
1926 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1927
1ee208c4
KW
1928 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1929 NULL, NULL, NULL, FALSE);
b423522f 1930
85c8e306
KW
1931 /* Make sure is clone-safe */
1932 ssc->invlist = NULL;
1933
e0e1be5f 1934 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1935 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1936 }
1462525b 1937
b2e90ddf
KW
1938 if (RExC_contains_locale) {
1939 OP(ssc) = ANYOFL;
1940 }
1941
1462525b 1942 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1943}
1944
a3621e74
YO
1945#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1946#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1947#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1948#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1949 ? (TRIE_LIST_CUR( idx ) - 1) \
1950 : 0 )
a3621e74 1951
3dab1dad
YO
1952
1953#ifdef DEBUGGING
07be1b83 1954/*
2b8b4781
NC
1955 dump_trie(trie,widecharmap,revcharmap)
1956 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1957 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1958
1959 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1960 The _interim_ variants are used for debugging the interim
1961 tables that are used to generate the final compressed
1962 representation which is what dump_trie expects.
1963
486ec47a 1964 Part of the reason for their existence is to provide a form
3dab1dad 1965 of documentation as to how the different representations function.
07be1b83
YO
1966
1967*/
3dab1dad
YO
1968
1969/*
3dab1dad
YO
1970 Dumps the final compressed table form of the trie to Perl_debug_log.
1971 Used for debugging make_trie().
1972*/
b9a59e08 1973
3dab1dad 1974STATIC void
2b8b4781
NC
1975S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1976 AV *revcharmap, U32 depth)
3dab1dad
YO
1977{
1978 U32 state;
ab3bbdeb 1979 SV *sv=sv_newmortal();
55eed653 1980 int colwidth= widecharmap ? 6 : 4;
2e64971a 1981 U16 word;
3dab1dad
YO
1982 GET_RE_DEBUG_FLAGS_DECL;
1983
7918f24d 1984 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1985
6ad9a8ab 1986 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 1987 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
1988
1989 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1990 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1991 if ( tmp ) {
6ad9a8ab 1992 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 1993 colwidth,
538e84ed 1994 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1995 PL_colors[0], PL_colors[1],
1996 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1997 PERL_PV_ESCAPE_FIRSTCHAR
1998 )
ab3bbdeb 1999 );
3dab1dad
YO
2000 }
2001 }
1e37780e
YO
2002 Perl_re_printf( aTHX_ "\n");
2003 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2004
2005 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2006 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2007 Perl_re_printf( aTHX_ "\n");
3dab1dad 2008
1e2e3d02 2009 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2010 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2011
147e3846 2012 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2013
2014 if ( trie->states[ state ].wordnum ) {
1e37780e 2015 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2016 } else {
6ad9a8ab 2017 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2018 }
2019
147e3846 2020 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2021
2022 if ( base ) {
2023 U32 ofs = 0;
2024
2025 while( ( base + ofs < trie->uniquecharcount ) ||
2026 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2027 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2028 != state))
3dab1dad
YO
2029 ofs++;
2030
147e3846 2031 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2032
2033 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2034 if ( ( base + ofs >= trie->uniquecharcount )
2035 && ( base + ofs - trie->uniquecharcount
2036 < trie->lasttrans )
2037 && trie->trans[ base + ofs
2038 - trie->uniquecharcount ].check == state )
3dab1dad 2039 {
147e3846 2040 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2041 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2042 );
3dab1dad 2043 } else {
6ad9a8ab 2044 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2045 }
2046 }
2047
6ad9a8ab 2048 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2049
2050 }
6ad9a8ab 2051 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2052 }
6ad9a8ab 2053 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2054 depth);
2e64971a 2055 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2056 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2057 (int)word, (int)(trie->wordinfo[word].prev),
2058 (int)(trie->wordinfo[word].len));
2059 }
6ad9a8ab 2060 Perl_re_printf( aTHX_ "\n" );
538e84ed 2061}
3dab1dad 2062/*
3dab1dad 2063 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2064 List tries normally only are used for construction when the number of
3dab1dad
YO
2065 possible chars (trie->uniquecharcount) is very high.
2066 Used for debugging make_trie().
2067*/
2068STATIC void
55eed653 2069S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2070 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2071 U32 depth)
3dab1dad
YO
2072{
2073 U32 state;
ab3bbdeb 2074 SV *sv=sv_newmortal();
55eed653 2075 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2076 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2077
2078 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2079
3dab1dad 2080 /* print out the table precompression. */
6ad9a8ab 2081 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2082 depth+1 );
6ad9a8ab 2083 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2084 depth+1, "------:-----+-----------------\n" );
538e84ed 2085
3dab1dad
YO
2086 for( state=1 ; state < next_alloc ; state ++ ) {
2087 U16 charid;
538e84ed 2088
147e3846 2089 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2090 depth+1, (UV)state );
3dab1dad 2091 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2092 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2093 } else {
6ad9a8ab 2094 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2095 trie->states[ state ].wordnum
2096 );
2097 }
2098 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2099 SV ** const tmp = av_fetch( revcharmap,
2100 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2101 if ( tmp ) {
147e3846 2102 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2103 colwidth,
538e84ed
KW
2104 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2105 colwidth,
2106 PL_colors[0], PL_colors[1],
2107 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2108 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2109 ) ,
1e2e3d02
YO
2110 TRIE_LIST_ITEM(state,charid).forid,
2111 (UV)TRIE_LIST_ITEM(state,charid).newstate
2112 );
538e84ed 2113 if (!(charid % 10))
6ad9a8ab 2114 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2115 (int)((depth * 2) + 14), "");
1e2e3d02 2116 }
ab3bbdeb 2117 }
6ad9a8ab 2118 Perl_re_printf( aTHX_ "\n");
3dab1dad 2119 }
538e84ed 2120}
3dab1dad
YO
2121
2122/*
3dab1dad 2123 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2124 This is the normal DFA style state transition table, with a few
2125 twists to facilitate compression later.
3dab1dad
YO
2126 Used for debugging make_trie().
2127*/
2128STATIC void
55eed653 2129S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2130 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2131 U32 depth)
3dab1dad
YO
2132{
2133 U32 state;
2134 U16 charid;
ab3bbdeb 2135 SV *sv=sv_newmortal();
55eed653 2136 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2137 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2138
2139 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2140
3dab1dad
YO
2141 /*
2142 print out the table precompression so that we can do a visual check
2143 that they are identical.
2144 */
538e84ed 2145
6ad9a8ab 2146 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2147
2148 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2149 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2150 if ( tmp ) {
6ad9a8ab 2151 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2152 colwidth,
538e84ed 2153 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2154 PL_colors[0], PL_colors[1],
2155 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2156 PERL_PV_ESCAPE_FIRSTCHAR
2157 )
ab3bbdeb 2158 );
3dab1dad
YO
2159 }
2160 }
2161
4aaafc03
YO
2162 Perl_re_printf( aTHX_ "\n");
2163 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2164
2165 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2166 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2167 }
2168
6ad9a8ab 2169 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2170
2171 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2172
147e3846 2173 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2174 depth+1,
3dab1dad
YO
2175 (UV)TRIE_NODENUM( state ) );
2176
2177 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2178 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2179 if (v)
147e3846 2180 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2181 else
6ad9a8ab 2182 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2183 }
2184 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2185 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2186 (UV)trie->trans[ state ].check );
3dab1dad 2187 } else {
147e3846 2188 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2189 (UV)trie->trans[ state ].check,
3dab1dad
YO
2190 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2191 }
2192 }
07be1b83 2193}
3dab1dad
YO
2194
2195#endif
2196
2e64971a 2197
786e8c11
YO
2198/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2199 startbranch: the first branch in the whole branch sequence
2200 first : start branch of sequence of branch-exact nodes.
2201 May be the same as startbranch
2202 last : Thing following the last branch.
2203 May be the same as tail.
2204 tail : item following the branch sequence
2205 count : words in the sequence
a4525e78 2206 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2207 depth : indent depth
3dab1dad 2208
786e8c11 2209Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2210
786e8c11
YO
2211A trie is an N'ary tree where the branches are determined by digital
2212decomposition of the key. IE, at the root node you look up the 1st character and
2213follow that branch repeat until you find the end of the branches. Nodes can be
2214marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2215
786e8c11 2216 /he|she|his|hers/
72f13be8 2217
786e8c11
YO
2218would convert into the following structure. Numbers represent states, letters
2219following numbers represent valid transitions on the letter from that state, if
2220the number is in square brackets it represents an accepting state, otherwise it
2221will be in parenthesis.
07be1b83 2222
786e8c11
YO
2223 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2224 | |
2225 | (2)
2226 | |
2227 (1) +-i->(6)-+-s->[7]
2228 |
2229 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2230
786e8c11
YO
2231 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2232
2233This shows that when matching against the string 'hers' we will begin at state 1
2234read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2235then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2236is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2237single traverse. We store a mapping from accepting to state to which word was
2238matched, and then when we have multiple possibilities we try to complete the
b8fda935 2239rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2240
2241The only prior NFA like behaviour that would be changed by the TRIE support is
2242the silent ignoring of duplicate alternations which are of the form:
2243
2244 / (DUPE|DUPE) X? (?{ ... }) Y /x
2245
4b714af6 2246Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2247and without the optimisation. With the optimisations dupes will be silently
486ec47a 2248ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2249the following demonstrates:
2250
2251 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2252
2253which prints out 'word' three times, but
2254
2255 'words'=~/(word|word|word)(?{ print $1 })S/
2256
2257which doesnt print it out at all. This is due to other optimisations kicking in.
2258
2259Example of what happens on a structural level:
2260
486ec47a 2261The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2262
2263 1: CURLYM[1] {1,32767}(18)
2264 5: BRANCH(8)
2265 6: EXACT <ac>(16)
2266 8: BRANCH(11)
2267 9: EXACT <ad>(16)
2268 11: BRANCH(14)
2269 12: EXACT <ab>(16)
2270 16: SUCCEED(0)
2271 17: NOTHING(18)
2272 18: END(0)
2273
2274This would be optimizable with startbranch=5, first=5, last=16, tail=16
2275and should turn into:
2276
2277 1: CURLYM[1] {1,32767}(18)
2278 5: TRIE(16)
2279 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2280 <ac>
2281 <ad>
2282 <ab>
2283 16: SUCCEED(0)
2284 17: NOTHING(18)
2285 18: END(0)
2286
2287Cases where tail != last would be like /(?foo|bar)baz/:
2288
2289 1: BRANCH(4)
2290 2: EXACT <foo>(8)
2291 4: BRANCH(7)
2292 5: EXACT <bar>(8)
2293 7: TAIL(8)
2294 8: EXACT <baz>(10)
2295 10: END(0)
2296
2297which would be optimizable with startbranch=1, first=1, last=7, tail=8
2298and would end up looking like:
2299
2300 1: TRIE(8)
2301 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2302 <foo>
2303 <bar>
2304 7: TAIL(8)
2305 8: EXACT <baz>(10)
2306 10: END(0)
2307
c80e42f3 2308 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2309
2310is the recommended Unicode-aware way of saying
2311
2312 *(d++) = uv;
2313*/
2314
fab2782b 2315#define TRIE_STORE_REVCHAR(val) \
786e8c11 2316 STMT_START { \
73031816 2317 if (UTF) { \
668fcfea 2318 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2319 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2320 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2321 SvCUR_set(zlopp, kapow - flrbbbbb); \
2322 SvPOK_on(zlopp); \
2323 SvUTF8_on(zlopp); \
2324 av_push(revcharmap, zlopp); \
2325 } else { \
fab2782b 2326 char ooooff = (char)val; \
73031816
NC
2327 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2328 } \
2329 } STMT_END
786e8c11 2330
914a25d5
KW
2331/* This gets the next character from the input, folding it if not already
2332 * folded. */
2333#define TRIE_READ_CHAR STMT_START { \
2334 wordlen++; \
2335 if ( UTF ) { \
2336 /* if it is UTF then it is either already folded, or does not need \
2337 * folding */ \
1c1d615a 2338 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2339 } \
2340 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2341 /* This folder implies Unicode rules, which in the range expressible \
2342 * by not UTF is the lower case, with the two exceptions, one of \
2343 * which should have been taken care of before calling this */ \
2344 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2345 uvc = toLOWER_L1(*uc); \
2346 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2347 len = 1; \
914a25d5
KW
2348 } else { \
2349 /* raw data, will be folded later if needed */ \
2350 uvc = (U32)*uc; \
2351 len = 1; \
2352 } \
786e8c11
YO
2353} STMT_END
2354
2355
2356
2357#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2358 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
2359 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2360 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
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
3476 /* and now dump out the compressed format */