This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Simplify expression
[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. */
eb7486de 122#define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
f4ae5a27
KW
123
124
09b2b2e6 125struct RExC_state_t {
514a91f1
DM
126 U32 flags; /* RXf_* are we folding, multilining? */
127 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 128 char *precomp; /* uncompiled string. */
711b303b 129 char *precomp_end; /* pointer to end of uncompiled string. */
288b8c02 130 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 131 regexp *rx; /* perl core regexp structure */
538e84ed
KW
132 regexp_internal *rxi; /* internal data for regexp object
133 pprivate field */
fac92740 134 char *start; /* Start of input for compile */
830247a4
IZ
135 char *end; /* End of input for compile */
136 char *parse; /* Input-scan pointer. */
285b5ca0
KW
137 char *adjusted_start; /* 'start', adjusted. See code use */
138 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
ea3daa5d 139 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 140 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
141 regnode *emit_bound; /* First regnode outside of the
142 allocated space */
f7c7e32a
DM
143 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
144 implies compiling, so don't emit */
9a81a976
KW
145 regnode_ssc emit_dummy; /* placeholder for emit to point to;
146 large enough for the largest
147 non-EXACTish node, so can use it as
148 scratch in pass1 */
830247a4
IZ
149 I32 naughty; /* How bad is this pattern? */
150 I32 sawback; /* Did we see \1, ...? */
151 U32 seen;
ea3daa5d 152 SSize_t size; /* Code size. */
538e84ed
KW
153 I32 npar; /* Capture buffer count, (OPEN) plus
154 one. ("par" 0 is the whole
155 pattern)*/
156 I32 nestroot; /* root parens we are in - used by
157 accept */
830247a4
IZ
158 I32 extralen;
159 I32 seen_zerolen;
40d049e4
YO
160 regnode **open_parens; /* pointers to open parens */
161 regnode **close_parens; /* pointers to close parens */
d5a00e4a 162 regnode *end_op; /* END node in program */
02daf0ab
YO
163 I32 utf8; /* whether the pattern is utf8 or not */
164 I32 orig_utf8; /* whether the pattern was originally in utf8 */
165 /* XXX use this for future optimisation of case
166 * where pattern must be upgraded to utf8. */
e40e74fe
KW
167 I32 uni_semantics; /* If a d charset modifier should use unicode
168 rules, even if the pattern is not in
169 utf8 */
81714fb9 170 HV *paren_names; /* Paren names */
538e84ed 171
40d049e4 172 regnode **recurse; /* Recurse regops */
d5a00e4a 173 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 174 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 175 through */
09a65838 176 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 177 I32 in_lookbehind;
4624b182 178 I32 contains_locale;
bb3f3ed2 179 I32 override_recoding;
b6d67071
KW
180#ifdef EBCDIC
181 I32 recode_x_to_native;
182#endif
9d53c457 183 I32 in_multi_char_class;
1acab4c5 184 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 185 within pattern */
b1603ef8 186 int code_index; /* next code_blocks[] slot */
ee273784 187 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
188 scan_frame *frame_head;
189 scan_frame *frame_last;
190 U32 frame_count;
7eec73eb 191 AV *warn_text;
dc6d7f5c 192#ifdef ADD_TO_REGEXEC
830247a4
IZ
193 char *starttry; /* -Dr: where regtry was called. */
194#define RExC_starttry (pRExC_state->starttry)
195#endif
d24ca0c5 196 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 197#ifdef DEBUGGING
be8e71aa 198 const char *lastparse;
3dab1dad 199 I32 lastnum;
1f1031fe 200 AV *paren_name_list; /* idx -> name */
d9a72fcc 201 U32 study_chunk_recursed_count;
c9f0d54c
YO
202 SV *mysv1;
203 SV *mysv2;
3dab1dad
YO
204#define RExC_lastparse (pRExC_state->lastparse)
205#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 206#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 207#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
208#define RExC_mysv (pRExC_state->mysv1)
209#define RExC_mysv1 (pRExC_state->mysv1)
210#define RExC_mysv2 (pRExC_state->mysv2)
211
3dab1dad 212#endif
512c0f5a 213 bool seen_unfolded_sharp_s;
911bd04e 214 bool strict;
da7cf1cc 215 bool study_started;
09b2b2e6 216};
830247a4 217
e2509266 218#define RExC_flags (pRExC_state->flags)
514a91f1 219#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 220#define RExC_precomp (pRExC_state->precomp)
285b5ca0
KW
221#define RExC_precomp_adj (pRExC_state->precomp_adj)
222#define RExC_adjusted_start (pRExC_state->adjusted_start)
711b303b 223#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 224#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 225#define RExC_rx (pRExC_state->rx)
f8fc2ecf 226#define RExC_rxi (pRExC_state->rxi)
fac92740 227#define RExC_start (pRExC_state->start)
830247a4
IZ
228#define RExC_end (pRExC_state->end)
229#define RExC_parse (pRExC_state->parse)
230#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
231
232/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
241
7122b237 242#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
243#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
244 others */
7122b237 245#endif
830247a4 246#define RExC_emit (pRExC_state->emit)
f7c7e32a 247#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 248#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 249#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
250#define RExC_sawback (pRExC_state->sawback)
251#define RExC_seen (pRExC_state->seen)
252#define RExC_size (pRExC_state->size)
ee273784 253#define RExC_maxlen (pRExC_state->maxlen)
830247a4 254#define RExC_npar (pRExC_state->npar)
e2e6a0f1 255#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
256#define RExC_extralen (pRExC_state->extralen)
257#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 258#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 259#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 260#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
261#define RExC_open_parens (pRExC_state->open_parens)
262#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 263#define RExC_end_op (pRExC_state->end_op)
81714fb9 264#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
265#define RExC_recurse (pRExC_state->recurse)
266#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 267#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
268#define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 270#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 271#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
272#ifdef EBCDIC
273# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
274#endif
9d53c457 275#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
276#define RExC_frame_head (pRExC_state->frame_head)
277#define RExC_frame_last (pRExC_state->frame_last)
278#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 279#define RExC_strict (pRExC_state->strict)
da7cf1cc 280#define RExC_study_started (pRExC_state->study_started)
7eec73eb 281#define RExC_warn_text (pRExC_state->warn_text)
830247a4 282
99807a43
HS
283/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
284 * a flag to disable back-off on the fixed/floating substrings - if it's
285 * a high complexity pattern we assume the benefit of avoiding a full match
286 * is worth the cost of checking for the substrings even if they rarely help.
287 */
288#define RExC_naughty (pRExC_state->naughty)
289#define TOO_NAUGHTY (10)
290#define MARK_NAUGHTY(add) \
291 if (RExC_naughty < TOO_NAUGHTY) \
292 RExC_naughty += (add)
293#define MARK_NAUGHTY_EXP(exp, add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 296
a687059c
LW
297#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
298#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 299 ((*s) == '{' && regcurly(s)))
a687059c
LW
300
301/*
302 * Flags to be passed up and down.
303 */
a687059c 304#define WORST 0 /* Worst case. */
a3b492c3 305#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 306
e64f369d 307/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
308 * character. (There needs to be a case: in the switch statement in regexec.c
309 * for any node marked SIMPLE.) Note that this is not the same thing as
310 * REGNODE_SIMPLE */
fda99bee 311#define SIMPLE 0x02
e64f369d 312#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
313#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
314#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
315#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
316#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
317 calcuate sizes as UTF-8 */
a687059c 318
3dab1dad
YO
319#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
320
07be1b83
YO
321/* whether trie related optimizations are enabled */
322#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
323#define TRIE_STUDY_OPT
786e8c11 324#define FULL_TRIE_STUDY
07be1b83
YO
325#define TRIE_STCLASS
326#endif
1de06328
YO
327
328
40d049e4
YO
329
330#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
331#define PBITVAL(paren) (1 << ((paren) & 7))
332#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
333#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
334#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
335
82a6ada4 336#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 337 if (!UTF) { \
82a6ada4
KW
338 assert(PASS1); \
339 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
340 return NULL; \
341 } \
82a6ada4 342 } STMT_END
40d049e4 343
512c0f5a
KW
344/* Change from /d into /u rules, and restart the parse if we've already seen
345 * something whose size would increase as a result, by setting *flagp and
346 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
347 * we've change to /u during the parse. */
348#define REQUIRE_UNI_RULES(flagp, restart_retval) \
349 STMT_START { \
350 if (DEPENDS_SEMANTICS) { \
351 assert(PASS1); \
352 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
353 RExC_uni_semantics = 1; \
354 if (RExC_seen_unfolded_sharp_s) { \
355 *flagp |= RESTART_PASS1; \
356 return restart_retval; \
357 } \
358 } \
359 } STMT_END
360
f19b1a63
KW
361/* This converts the named class defined in regcomp.h to its equivalent class
362 * number defined in handy.h. */
363#define namedclass_to_classnum(class) ((int) ((class) / 2))
364#define classnum_to_namedclass(classnum) ((classnum) * 2)
365
de92f5e6
KW
366#define _invlist_union_complement_2nd(a, b, output) \
367 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
368#define _invlist_intersection_complement_2nd(a, b, output) \
369 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
370
1de06328
YO
371/* About scan_data_t.
372
373 During optimisation we recurse through the regexp program performing
374 various inplace (keyhole style) optimisations. In addition study_chunk
375 and scan_commit populate this data structure with information about
538e84ed 376 what strings MUST appear in the pattern. We look for the longest
3b753521 377 string that must appear at a fixed location, and we look for the
1de06328
YO
378 longest string that may appear at a floating location. So for instance
379 in the pattern:
538e84ed 380
1de06328 381 /FOO[xX]A.*B[xX]BAR/
538e84ed 382
1de06328
YO
383 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
384 strings (because they follow a .* construct). study_chunk will identify
385 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 386
1de06328 387 The strings can be composites, for instance
538e84ed 388
1de06328 389 /(f)(o)(o)/
538e84ed 390
1de06328 391 will result in a composite fixed substring 'foo'.
538e84ed 392
1de06328 393 For each string some basic information is maintained:
538e84ed 394
1de06328
YO
395 - offset or min_offset
396 This is the position the string must appear at, or not before.
397 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
398 characters must match before the string we are searching for.
399 Likewise when combined with minlenp and the length of the string it
538e84ed 400 tells us how many characters must appear after the string we have
1de06328 401 found.
538e84ed 402
1de06328
YO
403 - max_offset
404 Only used for floating strings. This is the rightmost point that
ea3daa5d 405 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 406 string can occur infinitely far to the right.
538e84ed 407
1de06328 408 - minlenp
2d608413
KW
409 A pointer to the minimum number of characters of the pattern that the
410 string was found inside. This is important as in the case of positive
538e84ed 411 lookahead or positive lookbehind we can have multiple patterns
1de06328 412 involved. Consider
538e84ed 413
1de06328 414 /(?=FOO).*F/
538e84ed 415
1de06328
YO
416 The minimum length of the pattern overall is 3, the minimum length
417 of the lookahead part is 3, but the minimum length of the part that
538e84ed 418 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 419 minimum length for the F is 1. This is important as the minimum length
538e84ed 420 is used to determine offsets in front of and behind the string being
1de06328 421 looked for. Since strings can be composites this is the length of the
486ec47a 422 pattern at the time it was committed with a scan_commit. Note that
1de06328 423 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 424 are not known until the full pattern has been compiled, thus the
1de06328 425 pointer to the value.
538e84ed 426
1de06328 427 - lookbehind
538e84ed 428
1de06328 429 In the case of lookbehind the string being searched for can be
538e84ed 430 offset past the start point of the final matching string.
1de06328
YO
431 If this value was just blithely removed from the min_offset it would
432 invalidate some of the calculations for how many chars must match
433 before or after (as they are derived from min_offset and minlen and
538e84ed 434 the length of the string being searched for).
1de06328
YO
435 When the final pattern is compiled and the data is moved from the
436 scan_data_t structure into the regexp structure the information
538e84ed
KW
437 about lookbehind is factored in, with the information that would
438 have been lost precalculated in the end_shift field for the
1de06328
YO
439 associated string.
440
441 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 442 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
443
444*/
2c2d71f5
JH
445
446typedef struct scan_data_t {
1de06328
YO
447 /*I32 len_min; unused */
448 /*I32 len_delta; unused */
49f55535 449 SSize_t pos_min;
ea3daa5d 450 SSize_t pos_delta;
2c2d71f5 451 SV *last_found;
ea3daa5d 452 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 453 SSize_t last_start_min;
ea3daa5d 454 SSize_t last_start_max;
1de06328
YO
455 SV **longest; /* Either &l_fixed, or &l_float. */
456 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 457 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 458 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
459 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
460 SV *longest_float; /* longest floating string found in pattern */
49f55535 461 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
462 SSize_t offset_float_max; /* latest point in string it can appear */
463 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 464 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
465 I32 flags;
466 I32 whilem_c;
49f55535 467 SSize_t *last_closep;
b8f7bb16 468 regnode_ssc *start_class;
2c2d71f5
JH
469} scan_data_t;
470
a687059c 471/*
e50aee73 472 * Forward declarations for pregcomp()'s friends.
a687059c 473 */
a0d0e21e 474
27da23d5 475static const scan_data_t zero_scan_data =
1de06328 476 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
477
478#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
479#define SF_BEFORE_SEOL 0x0001
480#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
481#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
482#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
483
44e3dfd2
BF
484#define SF_FIX_SHIFT_EOL (+2)
485#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
486
487#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
488#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
489
490#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
491#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
492#define SF_IS_INF 0x0040
493#define SF_HAS_PAR 0x0080
494#define SF_IN_PAR 0x0100
495#define SF_HAS_EVAL 0x0200
bc604ad8
KW
496
497
498/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
499 * longest substring in the pattern. When it is not set the optimiser keeps
500 * track of position, but does not keep track of the actual strings seen,
501 *
502 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
503 * /foo/i will not.
504 *
505 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
506 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
507 * turned off because of the alternation (BRANCH). */
07be1b83 508#define SCF_DO_SUBSTR 0x0400
bc604ad8 509
653099ff
GS
510#define SCF_DO_STCLASS_AND 0x0800
511#define SCF_DO_STCLASS_OR 0x1000
512#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 513#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 514
786e8c11 515#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 516#define SCF_SEEN_ACCEPT 0x8000
688e0391 517#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
518#define SCF_IN_DEFINE 0x20000
519
520
521
07be1b83 522
43fead97 523#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
524
525/* The enums for all these are ordered so things work out correctly */
a62b1201 526#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
527#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
528 == REGEX_DEPENDS_CHARSET)
00b27cfc 529#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
530#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
531 >= REGEX_UNICODE_CHARSET)
532#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
533 == REGEX_ASCII_RESTRICTED_CHARSET)
534#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
535 >= REGEX_ASCII_RESTRICTED_CHARSET)
536#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
537 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 538
43fead97 539#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 540
f2c2a6ab
KW
541/* For programs that want to be strictly Unicode compatible by dying if any
542 * attempt is made to match a non-Unicode code point against a Unicode
543 * property. */
544#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
545
93733859 546#define OOB_NAMEDCLASS -1
b8c5462f 547
8e661ac5
KW
548/* There is no code point that is out-of-bounds, so this is problematic. But
549 * its only current use is to initialize a variable that is always set before
550 * looked at. */
551#define OOB_UNICODE 0xDEADBEEF
552
a0ed51b3 553#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 554
8615cb43 555
b45f050a
JF
556/* length of regex to show in messages that don't mark a position within */
557#define RegexLengthToShowInErrorMessages 127
558
559/*
560 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
561 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
562 * op/pragma/warn/regcomp.
563 */
7253e4e3
RK
564#define MARKER1 "<-- HERE" /* marker as it appears in the description */
565#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 566
538e84ed 567#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 568 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 569
285b5ca0
KW
570/* The code in this file in places uses one level of recursion with parsing
571 * rebased to an alternate string constructed by us in memory. This can take
572 * the form of something that is completely different from the input, or
573 * something that uses the input as part of the alternate. In the first case,
574 * there should be no possibility of an error, as we are in complete control of
575 * the alternate string. But in the second case we don't control the input
576 * portion, so there may be errors in that. Here's an example:
577 * /[abc\x{DF}def]/ui
578 * is handled specially because \x{df} folds to a sequence of more than one
579 * character, 'ss'. What is done is to create and parse an alternate string,
580 * which looks like this:
581 * /(?:\x{DF}|[abc\x{DF}def])/ui
582 * where it uses the input unchanged in the middle of something it constructs,
583 * which is a branch for the DF outside the character class, and clustering
584 * parens around the whole thing. (It knows enough to skip the DF inside the
585 * class while in this substitute parse.) 'abc' and 'def' may have errors that
586 * need to be reported. The general situation looks like this:
587 *
588 * sI tI xI eI
589 * Input: ----------------------------------------------------
590 * Constructed: ---------------------------------------------------
591 * sC tC xC eC EC
592 *
593 * The input string sI..eI is the input pattern. The string sC..EC is the
594 * constructed substitute parse string. The portions sC..tC and eC..EC are
595 * constructed by us. The portion tC..eC is an exact duplicate of the input
596 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
597 * while parsing, we find an error at xC. We want to display a message showing
598 * the real input string. Thus we need to find the point xI in it which
599 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
600 * been constructed by us, and so shouldn't have errors. We get:
601 *
602 * xI = sI + (tI - sI) + (xC - tC)
603 *
604 * and, the offset into sI is:
605 *
606 * (xI - sI) = (tI - sI) + (xC - tC)
607 *
608 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
609 * and we save tC as RExC_adjusted_start.
903c858a
KW
610 *
611 * During normal processing of the input pattern, everything points to that,
612 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
285b5ca0
KW
613 */
614
615#define tI_sI RExC_precomp_adj
616#define tC RExC_adjusted_start
617#define sC RExC_precomp
618#define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
619#define xI(xC) (sC + xI_offset(xC))
620#define eC RExC_precomp_end
621
622#define REPORT_LOCATION_ARGS(xC) \
623 UTF8fARG(UTF, \
624 (xI(xC) > eC) /* Don't run off end */ \
625 ? eC - sC /* Length before the <--HERE */ \
626 : xI_offset(xC), \
627 sC), /* The input pattern printed up to the <--HERE */ \
628 UTF8fARG(UTF, \
629 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
630 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
c1d900c3 631
8a6d8ec6
HS
632/* Used to point after bad bytes for an error message, but avoid skipping
633 * past a nul byte. */
634#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
635
b45f050a
JF
636/*
637 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
638 * arg. Show regex, up to a maximum length. If it's too long, chop and add
639 * "...".
640 */
58e23c8d 641#define _FAIL(code) STMT_START { \
bfed75c6 642 const char *ellipses = ""; \
711b303b 643 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
644 \
645 if (!SIZE_ONLY) \
a5e7bc51 646 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
647 if (len > RegexLengthToShowInErrorMessages) { \
648 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
649 len = RegexLengthToShowInErrorMessages - 10; \
650 ellipses = "..."; \
651 } \
58e23c8d 652 code; \
ccb2c380 653} STMT_END
8615cb43 654
58e23c8d 655#define FAIL(msg) _FAIL( \
147e3846 656 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 657 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
658
659#define FAIL2(msg,arg) _FAIL( \
147e3846 660 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 661 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 662
b45f050a 663/*
b45f050a
JF
664 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
665 */
ccb2c380 666#define Simple_vFAIL(m) STMT_START { \
ccb2c380 667 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 668 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 669} STMT_END
b45f050a
JF
670
671/*
672 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
673 */
ccb2c380
MP
674#define vFAIL(m) STMT_START { \
675 if (!SIZE_ONLY) \
a5e7bc51 676 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
677 Simple_vFAIL(m); \
678} STMT_END
b45f050a
JF
679
680/*
681 * Like Simple_vFAIL(), but accepts two arguments.
682 */
ccb2c380 683#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
684 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
685 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 686} STMT_END
b45f050a
JF
687
688/*
689 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
690 */
ccb2c380
MP
691#define vFAIL2(m,a1) STMT_START { \
692 if (!SIZE_ONLY) \
a5e7bc51 693 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
694 Simple_vFAIL2(m, a1); \
695} STMT_END
b45f050a
JF
696
697
698/*
699 * Like Simple_vFAIL(), but accepts three arguments.
700 */
ccb2c380 701#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 702 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 703 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 704} STMT_END
b45f050a
JF
705
706/*
707 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
708 */
ccb2c380
MP
709#define vFAIL3(m,a1,a2) STMT_START { \
710 if (!SIZE_ONLY) \
a5e7bc51 711 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
712 Simple_vFAIL3(m, a1, a2); \
713} STMT_END
b45f050a
JF
714
715/*
716 * Like Simple_vFAIL(), but accepts four arguments.
717 */
ccb2c380 718#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
719 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
720 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 721} STMT_END
b45f050a 722
95db3ffa
KW
723#define vFAIL4(m,a1,a2,a3) STMT_START { \
724 if (!SIZE_ONLY) \
725 SAVEFREESV(RExC_rx_sv); \
726 Simple_vFAIL4(m, a1, a2, a3); \
727} STMT_END
728
946095af 729/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
730#define vFAIL2utf8f(m, a1) STMT_START { \
731 if (!SIZE_ONLY) \
732 SAVEFREESV(RExC_rx_sv); \
733 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
734 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
735} STMT_END
736
3ba22297 737#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
738 if (!SIZE_ONLY) \
739 SAVEFREESV(RExC_rx_sv); \
740 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 741 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
742} STMT_END
743
499333dc
KW
744/* These have asserts in them because of [perl #122671] Many warnings in
745 * regcomp.c can occur twice. If they get output in pass1 and later in that
746 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
747 * would get output again. So they should be output in pass2, and these
748 * asserts make sure new warnings follow that paradigm. */
946095af 749
5e0a247b
KW
750/* m is not necessarily a "literal string", in this macro */
751#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
752 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
753 "%s" REPORT_LOCATION, \
754 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
755} STMT_END
756
668c081a 757#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
758 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
759 m REPORT_LOCATION, \
760 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
761} STMT_END
762
b927b7e9 763#define vWARN(loc, m) STMT_START { \
d528642a
KW
764 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
765 m REPORT_LOCATION, \
766 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
767} STMT_END
768
0d6106aa 769#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
770 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
771 m REPORT_LOCATION, \
772 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
773} STMT_END
774
147508a2 775#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
776 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
777 m REPORT_LOCATION, \
778 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
779} STMT_END
780
d528642a
KW
781#define ckWARNregdep(loc,m) STMT_START { \
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
783 WARN_REGEXP), \
784 m REPORT_LOCATION, \
785 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
786} STMT_END
787
d528642a
KW
788#define ckWARN2reg_d(loc,m, a1) STMT_START { \
789 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
790 m REPORT_LOCATION, \
791 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
792} STMT_END
793
d528642a
KW
794#define ckWARN2reg(loc, m, a1) STMT_START { \
795 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
796 m REPORT_LOCATION, \
797 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
798} STMT_END
799
d528642a
KW
800#define vWARN3(loc, m, a1, a2) STMT_START { \
801 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
802 m REPORT_LOCATION, \
803 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
804} STMT_END
805
d528642a
KW
806#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
807 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
808 m REPORT_LOCATION, \
809 a1, a2, \
810 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
811} STMT_END
812
ccb2c380 813#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
814 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
815 m REPORT_LOCATION, \
816 a1, a2, a3, \
817 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
818} STMT_END
819
668c081a 820#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
821 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
822 m REPORT_LOCATION, \
823 a1, a2, a3, \
824 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
825} STMT_END
826
ccb2c380 827#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
828 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
829 m REPORT_LOCATION, \
830 a1, a2, a3, a4, \
831 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 832} STMT_END
9d1d55b5 833
538e84ed 834/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
835 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
836 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
837 * Element 0 holds the number n.
07be1b83 838 * Position is 1 indexed.
fac92740 839 */
7122b237
YO
840#ifndef RE_TRACK_PATTERN_OFFSETS
841#define Set_Node_Offset_To_R(node,byte)
842#define Set_Node_Offset(node,byte)
843#define Set_Cur_Node_Offset
844#define Set_Node_Length_To_R(node,len)
845#define Set_Node_Length(node,len)
6a86c6ad 846#define Set_Node_Cur_Length(node,start)
538e84ed
KW
847#define Node_Offset(n)
848#define Node_Length(n)
7122b237
YO
849#define Set_Node_Offset_Length(node,offset,len)
850#define ProgLen(ri) ri->u.proglen
851#define SetProgLen(ri,x) ri->u.proglen = x
852#else
853#define ProgLen(ri) ri->u.offsets[0]
854#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
855#define Set_Node_Offset_To_R(node,byte) STMT_START { \
856 if (! SIZE_ONLY) { \
857 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 858 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 859 if((node) < 0) { \
538e84ed
KW
860 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
861 (int)(node)); \
ccb2c380
MP
862 } else { \
863 RExC_offsets[2*(node)-1] = (byte); \
864 } \
865 } \
866} STMT_END
867
868#define Set_Node_Offset(node,byte) \
869 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
870#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
871
872#define Set_Node_Length_To_R(node,len) STMT_START { \
873 if (! SIZE_ONLY) { \
874 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 875 __LINE__, (int)(node), (int)(len))); \
ccb2c380 876 if((node) < 0) { \
538e84ed
KW
877 Perl_croak(aTHX_ "value of node is %d in Length macro", \
878 (int)(node)); \
ccb2c380
MP
879 } else { \
880 RExC_offsets[2*(node)] = (len); \
881 } \
882 } \
883} STMT_END
884
885#define Set_Node_Length(node,len) \
886 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
887#define Set_Node_Cur_Length(node, start) \
888 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
889
890/* Get offsets and lengths */
891#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
892#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
893
07be1b83
YO
894#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
895 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
896 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
897} STMT_END
7122b237 898#endif
07be1b83
YO
899
900#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
901#define EXPERIMENTAL_INPLACESCAN
f427392e 902#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 903
cb41e5d6
YO
904#ifdef DEBUGGING
905int
6ad9a8ab 906Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
907{
908 va_list ap;
909 int result;
910 PerlIO *f= Perl_debug_log;
911 PERL_ARGS_ASSERT_RE_PRINTF;
912 va_start(ap, fmt);
913 result = PerlIO_vprintf(f, fmt, ap);
914 va_end(ap);
915 return result;
916}
917
918int
7b031478 919Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
920{
921 va_list ap;
922 int result;
923 PerlIO *f= Perl_debug_log;
924 PERL_ARGS_ASSERT_RE_INDENTF;
925 va_start(ap, depth);
daeb874b 926 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
927 result = PerlIO_vprintf(f, fmt, ap);
928 va_end(ap);
929 return result;
930}
931#endif /* DEBUGGING */
932
7b031478 933#define DEBUG_RExC_seen() \
538e84ed 934 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 935 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 936 \
e384d5c1 937 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 938 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 939 \
e384d5c1 940 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 941 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 942 \
e384d5c1 943 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 944 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 945 \
e384d5c1 946 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 947 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 948 \
7b031478 949 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 950 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 951 \
e384d5c1 952 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 953 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 954 \
e384d5c1 955 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 956 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 957 \
e384d5c1 958 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 959 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 960 \
e384d5c1 961 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 962 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 963 \
7b031478 964 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 965 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 966 \
6ad9a8ab 967 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
968 });
969
fdfb4f21 970#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 971 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21
YO
972
973#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
974 if ( ( flags ) ) { \
6ad9a8ab 975 Perl_re_printf( aTHX_ "%s", open_str); \
fdfb4f21
YO
976 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
977 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
978 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
979 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
980 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
981 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
982 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
983 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
984 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
985 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
986 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
987 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
988 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
989 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
990 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
6ad9a8ab 991 Perl_re_printf( aTHX_ "%s", close_str); \
fdfb4f21
YO
992 }
993
994
304ee84b
YO
995#define DEBUG_STUDYDATA(str,data,depth) \
996DEBUG_OPTIMISE_MORE_r(if(data){ \
147e3846
KW
997 Perl_re_indentf( aTHX_ "" str "Pos:%" IVdf "/%" IVdf \
998 " Flags: 0x%" UVXf, \
7b031478 999 depth, \
1de06328
YO
1000 (IV)((data)->pos_min), \
1001 (IV)((data)->pos_delta), \
fdfb4f21
YO
1002 (UV)((data)->flags) \
1003 ); \
1004 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
147e3846
KW
1005 Perl_re_printf( aTHX_ \
1006 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s", \
1de06328 1007 (IV)((data)->whilem_c), \
304ee84b
YO
1008 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
1009 is_inf ? "INF " : "" \
1de06328
YO
1010 ); \
1011 if ((data)->last_found) \
147e3846
KW
1012 Perl_re_printf( aTHX_ \
1013 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf \
1014 " %sFixed:'%s' @ %" IVdf \
1015 " %sFloat: '%s' @ %" IVdf "/%" IVdf, \
1de06328
YO
1016 SvPVX_const((data)->last_found), \
1017 (IV)((data)->last_end), \
1018 (IV)((data)->last_start_min), \
1019 (IV)((data)->last_start_max), \
1020 ((data)->longest && \
1021 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
1022 SvPVX_const((data)->longest_fixed), \
1023 (IV)((data)->offset_fixed), \
1024 ((data)->longest && \
1025 (data)->longest==&((data)->longest_float)) ? "*" : "", \
1026 SvPVX_const((data)->longest_float), \
1027 (IV)((data)->offset_float_min), \
1028 (IV)((data)->offset_float_max) \
1029 ); \
6ad9a8ab 1030 Perl_re_printf( aTHX_ "\n"); \
1de06328
YO
1031});
1032
cb41e5d6 1033
c6871b76
KW
1034/* =========================================================
1035 * BEGIN edit_distance stuff.
1036 *
1037 * This calculates how many single character changes of any type are needed to
1038 * transform a string into another one. It is taken from version 3.1 of
1039 *
1040 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1041 */
1042
1043/* Our unsorted dictionary linked list. */
1044/* Note we use UVs, not chars. */
1045
1046struct dictionary{
1047 UV key;
1048 UV value;
1049 struct dictionary* next;
1050};
1051typedef struct dictionary item;
1052
1053
1054PERL_STATIC_INLINE item*
1055push(UV key,item* curr)
1056{
1057 item* head;
1058 Newxz(head, 1, item);
1059 head->key = key;
1060 head->value = 0;
1061 head->next = curr;
1062 return head;
1063}
1064
1065
1066PERL_STATIC_INLINE item*
1067find(item* head, UV key)
1068{
1069 item* iterator = head;
1070 while (iterator){
1071 if (iterator->key == key){
1072 return iterator;
1073 }
1074 iterator = iterator->next;
1075 }
1076
1077 return NULL;
1078}
1079
1080PERL_STATIC_INLINE item*
1081uniquePush(item* head,UV key)
1082{
1083 item* iterator = head;
1084
1085 while (iterator){
1086 if (iterator->key == key) {
1087 return head;
1088 }
1089 iterator = iterator->next;
1090 }
1091
1092 return push(key,head);
1093}
1094
1095PERL_STATIC_INLINE void
1096dict_free(item* head)
1097{
1098 item* iterator = head;
1099
1100 while (iterator) {
1101 item* temp = iterator;
1102 iterator = iterator->next;
1103 Safefree(temp);
1104 }
1105
1106 head = NULL;
1107}
1108
1109/* End of Dictionary Stuff */
1110
1111/* All calculations/work are done here */
1112STATIC int
1113S_edit_distance(const UV* src,
1114 const UV* tgt,
1115 const STRLEN x, /* length of src[] */
1116 const STRLEN y, /* length of tgt[] */
1117 const SSize_t maxDistance
1118)
1119{
1120 item *head = NULL;
1121 UV swapCount,swapScore,targetCharCount,i,j;
1122 UV *scores;
1123 UV score_ceil = x + y;
1124
1125 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1126
1127 /* intialize matrix start values */
1128 Newxz(scores, ( (x + 2) * (y + 2)), UV);
1129 scores[0] = score_ceil;
1130 scores[1 * (y + 2) + 0] = score_ceil;
1131 scores[0 * (y + 2) + 1] = score_ceil;
1132 scores[1 * (y + 2) + 1] = 0;
1133 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1134
1135 /* work loops */
1136 /* i = src index */
1137 /* j = tgt index */
1138 for (i=1;i<=x;i++) {
1139 if (i < x)
1140 head = uniquePush(head,src[i]);
1141 scores[(i+1) * (y + 2) + 1] = i;
1142 scores[(i+1) * (y + 2) + 0] = score_ceil;
1143 swapCount = 0;
1144
1145 for (j=1;j<=y;j++) {
1146 if (i == 1) {
1147 if(j < y)
1148 head = uniquePush(head,tgt[j]);
1149 scores[1 * (y + 2) + (j + 1)] = j;
1150 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1151 }
1152
1153 targetCharCount = find(head,tgt[j-1])->value;
1154 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1155
1156 if (src[i-1] != tgt[j-1]){
1157 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));
1158 }
1159 else {
1160 swapCount = j;
1161 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1162 }
1163 }
1164
1165 find(head,src[i-1])->value = i;
1166 }
1167
1168 {
1169 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1170 dict_free(head);
1171 Safefree(scores);
1172 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1173 }
1174}
1175
1176/* END of edit_distance() stuff
1177 * ========================================================= */
1178
8e35b056
KW
1179/* is c a control character for which we have a mnemonic? */
1180#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1181
549b4e78
KW
1182STATIC const char *
1183S_cntrl_to_mnemonic(const U8 c)
1184{
1185 /* Returns the mnemonic string that represents character 'c', if one
1186 * exists; NULL otherwise. The only ones that exist for the purposes of
1187 * this routine are a few control characters */
1188
1189 switch (c) {
1190 case '\a': return "\\a";
1191 case '\b': return "\\b";
1192 case ESC_NATIVE: return "\\e";
1193 case '\f': return "\\f";
1194 case '\n': return "\\n";
1195 case '\r': return "\\r";
1196 case '\t': return "\\t";
1197 }
1198
1199 return NULL;
1200}
1201
653099ff 1202/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 1203 Update the longest found anchored substring and the longest found
653099ff
GS
1204 floating substrings if needed. */
1205
4327152a 1206STATIC void
ea3daa5d
FC
1207S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1208 SSize_t *minlenp, int is_inf)
c277df42 1209{
e1ec3a88
AL
1210 const STRLEN l = CHR_SVLEN(data->last_found);
1211 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 1212 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1213
7918f24d
NC
1214 PERL_ARGS_ASSERT_SCAN_COMMIT;
1215
c277df42 1216 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 1217 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
1218 if (*data->longest == data->longest_fixed) {
1219 data->offset_fixed = l ? data->last_start_min : data->pos_min;
1220 if (data->flags & SF_BEFORE_EOL)
b81d288d 1221 data->flags
c277df42
IZ
1222 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
1223 else
1224 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 1225 data->minlen_fixed=minlenp;
1de06328 1226 data->lookbehind_fixed=0;
a0ed51b3 1227 }
304ee84b 1228 else { /* *data->longest == data->longest_float */
c277df42 1229 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 1230 data->offset_float_max = (l
646e8787
DM
1231 ? data->last_start_max
1232 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1233 ? SSize_t_MAX
1234 : data->pos_min + data->pos_delta));
1235 if (is_inf
1236 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
1237 data->offset_float_max = SSize_t_MAX;
c277df42 1238 if (data->flags & SF_BEFORE_EOL)
b81d288d 1239 data->flags
c277df42
IZ
1240 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
1241 else
1242 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
1243 data->minlen_float=minlenp;
1244 data->lookbehind_float=0;
c277df42
IZ
1245 }
1246 }
1247 SvCUR_set(data->last_found, 0);
0eda9292 1248 {
a28509cc 1249 SV * const sv = data->last_found;
097eb12c
AL
1250 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1251 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1252 if (mg)
1253 mg->mg_len = 0;
1254 }
0eda9292 1255 }
c277df42
IZ
1256 data->last_end = -1;
1257 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1258 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1259}
1260
cdd87c1d
KW
1261/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1262 * list that describes which code points it matches */
1263
653099ff 1264STATIC void
3420edd7 1265S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1266{
cdd87c1d
KW
1267 /* Set the SSC 'ssc' to match an empty string or any code point */
1268
557bd3fb 1269 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1270
71068078 1271 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1272
0854ea0b
KW
1273 /* mortalize so won't leak */
1274 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1275 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1276}
1277
653099ff 1278STATIC int
dc3bf405 1279S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1280{
c144baaa
KW
1281 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1282 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1283 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1284 * in any way, so there's no point in using it */
cdd87c1d
KW
1285
1286 UV start, end;
1287 bool ret;
653099ff 1288
557bd3fb 1289 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1290
71068078 1291 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1292
93e92956 1293 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1294 return FALSE;
1295 }
1296
1297 /* See if the list consists solely of the range 0 - Infinity */
1298 invlist_iterinit(ssc->invlist);
1299 ret = invlist_iternext(ssc->invlist, &start, &end)
1300 && start == 0
1301 && end == UV_MAX;
1302
1303 invlist_iterfinish(ssc->invlist);
1304
1305 if (ret) {
1306 return TRUE;
1307 }
1308
1309 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1310 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1311 int i;
1312 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1313 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1314 return TRUE;
1315 }
1316 }
1317 }
1318
1319 return FALSE;
653099ff
GS
1320}
1321
653099ff 1322STATIC void
cdd87c1d 1323S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1324{
cdd87c1d
KW
1325 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1326 * string, any code point, or any posix class under locale */
1327
557bd3fb 1328 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1329
557bd3fb 1330 Zero(ssc, 1, regnode_ssc);
71068078 1331 set_ANYOF_SYNTHETIC(ssc);
93e92956 1332 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1333 ssc_anything(ssc);
cdd87c1d 1334
2f306ab9
KW
1335 /* If any portion of the regex is to operate under locale rules that aren't
1336 * fully known at compile time, initialization includes it. The reason
1337 * this isn't done for all regexes is that the optimizer was written under
1338 * the assumption that locale was all-or-nothing. Given the complexity and
1339 * lack of documentation in the optimizer, and that there are inadequate
1340 * test cases for locale, many parts of it may not work properly, it is
1341 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1342 if (RExC_contains_locale) {
1343 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1344 }
1345 else {
1346 ANYOF_POSIXL_ZERO(ssc);
1347 }
653099ff
GS
1348}
1349
b423522f 1350STATIC int
dc3bf405
BF
1351S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1352 const regnode_ssc *ssc)
b423522f
KW
1353{
1354 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1355 * to the list of code points matched, and locale posix classes; hence does
1356 * not check its flags) */
1357
1358 UV start, end;
1359 bool ret;
1360
1361 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1362
71068078 1363 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1364
1365 invlist_iterinit(ssc->invlist);
1366 ret = invlist_iternext(ssc->invlist, &start, &end)
1367 && start == 0
1368 && end == UV_MAX;
1369
1370 invlist_iterfinish(ssc->invlist);
1371
1372 if (! ret) {
1373 return FALSE;
1374 }
1375
e0e1be5f 1376 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1377 return FALSE;
b423522f
KW
1378 }
1379
1380 return TRUE;
1381}
1382
1383STATIC SV*
1384S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1385 const regnode_charclass* const node)
b423522f
KW
1386{
1387 /* Returns a mortal inversion list defining which code points are matched
1388 * by 'node', which is of type ANYOF. Handles complementing the result if
1389 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1390 * returned list must, and will, contain every code point that is a
1391 * possibility. */
b423522f 1392
e2506fa7 1393 SV* invlist = NULL;
1ee208c4 1394 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1395 unsigned int i;
1396 const U32 n = ARG(node);
31f05a37 1397 bool new_node_has_latin1 = FALSE;
b423522f
KW
1398
1399 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1400
1401 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1402 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1403 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1404 AV * const av = MUTABLE_AV(SvRV(rv));
1405 SV **const ary = AvARRAY(av);
1406 assert(RExC_rxi->data->what[n] == 's');
1407
1408 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1409 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1410 }
1411 else if (ary[0] && ary[0] != &PL_sv_undef) {
1412
1413 /* Here, no compile-time swash, and there are things that won't be
1414 * known until runtime -- we have to assume it could be anything */
e2506fa7 1415 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1416 return _add_range_to_invlist(invlist, 0, UV_MAX);
1417 }
1ee208c4 1418 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1419
1420 /* Here no compile-time swash, and no run-time only data. Use the
1421 * node's inversion list */
1ee208c4
KW
1422 invlist = sv_2mortal(invlist_clone(ary[3]));
1423 }
1424
1425 /* Get the code points valid only under UTF-8 locales */
037715a6 1426 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1427 && ary[2] && ary[2] != &PL_sv_undef)
1428 {
1429 only_utf8_locale_invlist = ary[2];
b423522f
KW
1430 }
1431 }
1432
e2506fa7
KW
1433 if (! invlist) {
1434 invlist = sv_2mortal(_new_invlist(0));
1435 }
1436
dcb20b36
KW
1437 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1438 * code points, and an inversion list for the others, but if there are code
1439 * points that should match only conditionally on the target string being
1440 * UTF-8, those are placed in the inversion list, and not the bitmap.
1441 * Since there are circumstances under which they could match, they are
1442 * included in the SSC. But if the ANYOF node is to be inverted, we have
1443 * to exclude them here, so that when we invert below, the end result
1444 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1445 * have to do this here before we add the unconditionally matched code
1446 * points */
b423522f
KW
1447 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1448 _invlist_intersection_complement_2nd(invlist,
1449 PL_UpperLatin1,
1450 &invlist);
1451 }
1452
1453 /* Add in the points from the bit map */
dcb20b36 1454 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1455 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1456 unsigned int start = i++;
1457
1458 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1459 /* empty */
1460 }
1461 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1462 new_node_has_latin1 = TRUE;
b423522f
KW
1463 }
1464 }
1465
1466 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1467 * as well. But don't add them if inverting, as when that gets done below,
1468 * it would exclude all these characters, including the ones it shouldn't
1469 * that were added just above */
1470 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1471 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1472 {
b423522f
KW
1473 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1474 }
1475
1476 /* Similarly for these */
93e92956 1477 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1478 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1479 }
1480
1481 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1482 _invlist_invert(invlist);
1483 }
037715a6 1484 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1485
1486 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1487 * locale. We can skip this if there are no 0-255 at all. */
1488 _invlist_union(invlist, PL_Latin1, &invlist);
1489 }
1490
1ee208c4
KW
1491 /* Similarly add the UTF-8 locale possible matches. These have to be
1492 * deferred until after the non-UTF-8 locale ones are taken care of just
1493 * above, or it leads to wrong results under ANYOF_INVERT */
1494 if (only_utf8_locale_invlist) {
31f05a37 1495 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1496 only_utf8_locale_invlist,
31f05a37
KW
1497 ANYOF_FLAGS(node) & ANYOF_INVERT,
1498 &invlist);
1499 }
b423522f
KW
1500
1501 return invlist;
1502}
1503
1051e1c4 1504/* These two functions currently do the exact same thing */
557bd3fb 1505#define ssc_init_zero ssc_init
653099ff 1506
cdd87c1d
KW
1507#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1508#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1509
557bd3fb 1510/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1511 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1512 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1513
653099ff 1514STATIC void
b423522f 1515S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1516 const regnode_charclass *and_with)
653099ff 1517{
cdd87c1d
KW
1518 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1519 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1520
a0dd4231
KW
1521 SV* anded_cp_list;
1522 U8 anded_flags;
1e6ade67 1523
cdd87c1d 1524 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1525
71068078 1526 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1527
1528 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1529 * the code point inversion list and just the relevant flags */
71068078 1530 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1531 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1532 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1533
1534 /* XXX This is a kludge around what appears to be deficiencies in the
1535 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1536 * there are paths through the optimizer where it doesn't get weeded
1537 * out when it should. And if we don't make some extra provision for
1538 * it like the code just below, it doesn't get added when it should.
1539 * This solution is to add it only when AND'ing, which is here, and
1540 * only when what is being AND'ed is the pristine, original node
1541 * matching anything. Thus it is like adding it to ssc_anything() but
1542 * only when the result is to be AND'ed. Probably the same solution
1543 * could be adopted for the same problem we have with /l matching,
1544 * which is solved differently in S_ssc_init(), and that would lead to
1545 * fewer false positives than that solution has. But if this solution
1546 * creates bugs, the consequences are only that a warning isn't raised
1547 * that should be; while the consequences for having /l bugs is
1548 * incorrect matches */
7dcac5f6 1549 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1550 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1551 }
a0dd4231
KW
1552 }
1553 else {
5c0f85ef 1554 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1555 if (OP(and_with) == ANYOFD) {
1556 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1557 }
1558 else {
1559 anded_flags = ANYOF_FLAGS(and_with)
1560 &( ANYOF_COMMON_FLAGS
108316fb
KW
1561 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1562 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1563 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1564 anded_flags &=
1565 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1566 }
f240c685 1567 }
a0dd4231
KW
1568 }
1569
1570 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1571
1572 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1573 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1574 * 'and_with' may be inverted. When not inverted, we have the situation of
1575 * computing:
1576 * (C1 | P1) & (C2 | P2)
1577 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1578 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1579 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1580 * <= ((C1 & C2) | P1 | P2)
1581 * Alternatively, the last few steps could be:
1582 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1583 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1584 * <= (C1 | C2 | (P1 & P2))
1585 * We favor the second approach if either P1 or P2 is non-empty. This is
1586 * because these components are a barrier to doing optimizations, as what
1587 * they match cannot be known until the moment of matching as they are
1588 * dependent on the current locale, 'AND"ing them likely will reduce or
1589 * eliminate them.
1590 * But we can do better if we know that C1,P1 are in their initial state (a
1591 * frequent occurrence), each matching everything:
1592 * (<everything>) & (C2 | P2) = C2 | P2
1593 * Similarly, if C2,P2 are in their initial state (again a frequent
1594 * occurrence), the result is a no-op
1595 * (C1 | P1) & (<everything>) = C1 | P1
1596 *
1597 * Inverted, we have
1598 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1599 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1600 * <= (C1 & ~C2) | (P1 & ~P2)
1601 * */
1aa99e6b 1602
a0dd4231 1603 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1604 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1605 {
cdd87c1d 1606 unsigned int i;
8951c461 1607
cdd87c1d
KW
1608 ssc_intersection(ssc,
1609 anded_cp_list,
1610 FALSE /* Has already been inverted */
1611 );
c6b76537 1612
cdd87c1d
KW
1613 /* If either P1 or P2 is empty, the intersection will be also; can skip
1614 * the loop */
93e92956 1615 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1616 ANYOF_POSIXL_ZERO(ssc);
1617 }
e0e1be5f 1618 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1619
1620 /* Note that the Posix class component P from 'and_with' actually
1621 * looks like:
1622 * P = Pa | Pb | ... | Pn
1623 * where each component is one posix class, such as in [\w\s].
1624 * Thus
1625 * ~P = ~(Pa | Pb | ... | Pn)
1626 * = ~Pa & ~Pb & ... & ~Pn
1627 * <= ~Pa | ~Pb | ... | ~Pn
1628 * The last is something we can easily calculate, but unfortunately
1629 * is likely to have many false positives. We could do better
1630 * in some (but certainly not all) instances if two classes in
1631 * P have known relationships. For example
1632 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1633 * So
1634 * :lower: & :print: = :lower:
1635 * And similarly for classes that must be disjoint. For example,
1636 * since \s and \w can have no elements in common based on rules in
1637 * the POSIX standard,
1638 * \w & ^\S = nothing
1639 * Unfortunately, some vendor locales do not meet the Posix
1640 * standard, in particular almost everything by Microsoft.
1641 * The loop below just changes e.g., \w into \W and vice versa */
1642
1ee208c4 1643 regnode_charclass_posixl temp;
cdd87c1d
KW
1644 int add = 1; /* To calculate the index of the complement */
1645
1646 ANYOF_POSIXL_ZERO(&temp);
1647 for (i = 0; i < ANYOF_MAX; i++) {
1648 assert(i % 2 != 0
7dcac5f6
KW
1649 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1650 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1651
7dcac5f6 1652 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1653 ANYOF_POSIXL_SET(&temp, i + add);
1654 }
1655 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1656 }
1657 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1658
cdd87c1d
KW
1659 } /* else ssc already has no posixes */
1660 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1661 in its initial state */
71068078 1662 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1663 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1664 {
1665 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1666 * copy it over 'ssc' */
1667 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1668 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1669 StructCopy(and_with, ssc, regnode_ssc);
1670 }
1671 else {
1672 ssc->invlist = anded_cp_list;
1673 ANYOF_POSIXL_ZERO(ssc);
93e92956 1674 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1675 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1676 }
1677 }
1678 }
e0e1be5f 1679 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1680 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1681 {
1682 /* One or the other of P1, P2 is non-empty. */
93e92956 1683 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1684 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1685 }
cdd87c1d
KW
1686 ssc_union(ssc, anded_cp_list, FALSE);
1687 }
1688 else { /* P1 = P2 = empty */
1689 ssc_intersection(ssc, anded_cp_list, FALSE);
1690 }
137165a6 1691 }
653099ff
GS
1692}
1693
653099ff 1694STATIC void
cdd87c1d 1695S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1696 const regnode_charclass *or_with)
653099ff 1697{
cdd87c1d
KW
1698 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1699 * another SSC or a regular ANYOF class. Can create false positives if
1700 * 'or_with' is to be inverted. */
7918f24d 1701
a0dd4231
KW
1702 SV* ored_cp_list;
1703 U8 ored_flags;
c6b76537 1704
cdd87c1d 1705 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1706
71068078 1707 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1708
1709 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1710 * the code point inversion list and just the relevant flags */
71068078 1711 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1712 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1713 ored_flags = ANYOF_FLAGS(or_with);
1714 }
1715 else {
5c0f85ef 1716 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1717 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1718 if (OP(or_with) != ANYOFD) {
1719 ored_flags
1720 |= ANYOF_FLAGS(or_with)
108316fb
KW
1721 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1722 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1723 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1724 ored_flags |=
1725 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1726 }
f240c685 1727 }
a0dd4231
KW
1728 }
1729
1730 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1731
1732 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1733 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1734 * 'or_with' may be inverted. When not inverted, we have the simple
1735 * situation of computing:
1736 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1737 * If P1|P2 yields a situation with both a class and its complement are
1738 * set, like having both \w and \W, this matches all code points, and we
1739 * can delete these from the P component of the ssc going forward. XXX We
1740 * might be able to delete all the P components, but I (khw) am not certain
1741 * about this, and it is better to be safe.
1742 *
1743 * Inverted, we have
1744 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1745 * <= (C1 | P1) | ~C2
1746 * <= (C1 | ~C2) | P1
1747 * (which results in actually simpler code than the non-inverted case)
1748 * */
9826f543 1749
a0dd4231 1750 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1751 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1752 {
cdd87c1d 1753 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1754 } /* else Not inverted */
93e92956 1755 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1756 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1757 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1758 unsigned int i;
1759 for (i = 0; i < ANYOF_MAX; i += 2) {
1760 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1761 {
1762 ssc_match_all_cp(ssc);
1763 ANYOF_POSIXL_CLEAR(ssc, i);
1764 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1765 }
1766 }
1767 }
1aa99e6b 1768 }
cdd87c1d
KW
1769
1770 ssc_union(ssc,
1771 ored_cp_list,
1772 FALSE /* Already has been inverted */
1773 );
653099ff
GS
1774}
1775
b423522f
KW
1776PERL_STATIC_INLINE void
1777S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1778{
1779 PERL_ARGS_ASSERT_SSC_UNION;
1780
71068078 1781 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1782
1783 _invlist_union_maybe_complement_2nd(ssc->invlist,
1784 invlist,
1785 invert2nd,
1786 &ssc->invlist);
1787}
1788
1789PERL_STATIC_INLINE void
1790S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1791 SV* const invlist,
1792 const bool invert2nd)
1793{
1794 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1795
71068078 1796 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1797
1798 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1799 invlist,
1800 invert2nd,
1801 &ssc->invlist);
1802}
1803
1804PERL_STATIC_INLINE void
1805S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1806{
1807 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1808
71068078 1809 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1810
1811 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1812}
1813
1814PERL_STATIC_INLINE void
1815S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1816{
1817 /* AND just the single code point 'cp' into the SSC 'ssc' */
1818
1819 SV* cp_list = _new_invlist(2);
1820
1821 PERL_ARGS_ASSERT_SSC_CP_AND;
1822
71068078 1823 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1824
1825 cp_list = add_cp_to_invlist(cp_list, cp);
1826 ssc_intersection(ssc, cp_list,
1827 FALSE /* Not inverted */
1828 );
1829 SvREFCNT_dec_NN(cp_list);
1830}
1831
1832PERL_STATIC_INLINE void
dc3bf405 1833S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1834{
1835 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1836 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1837
71068078 1838 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1839
1840 ANYOF_POSIXL_ZERO(ssc);
1841 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1842}
1843
b35552de
KW
1844#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1845
1846STATIC bool
1847S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1848{
1849 /* The synthetic start class is used to hopefully quickly winnow down
1850 * places where a pattern could start a match in the target string. If it
1851 * doesn't really narrow things down that much, there isn't much point to
1852 * having the overhead of using it. This function uses some very crude
1853 * heuristics to decide if to use the ssc or not.
1854 *
1855 * It returns TRUE if 'ssc' rules out more than half what it considers to
1856 * be the "likely" possible matches, but of course it doesn't know what the
1857 * actual things being matched are going to be; these are only guesses
1858 *
1859 * For /l matches, it assumes that the only likely matches are going to be
1860 * in the 0-255 range, uniformly distributed, so half of that is 127
1861 * For /a and /d matches, it assumes that the likely matches will be just
1862 * the ASCII range, so half of that is 63
1863 * For /u and there isn't anything matching above the Latin1 range, it
1864 * assumes that that is the only range likely to be matched, and uses
1865 * half that as the cut-off: 127. If anything matches above Latin1,
1866 * it assumes that all of Unicode could match (uniformly), except for
1867 * non-Unicode code points and things in the General Category "Other"
1868 * (unassigned, private use, surrogates, controls and formats). This
1869 * is a much large number. */
1870
b35552de
KW
1871 U32 count = 0; /* Running total of number of code points matched by
1872 'ssc' */
1873 UV start, end; /* Start and end points of current range in inversion
1874 list */
72400949
KW
1875 const U32 max_code_points = (LOC)
1876 ? 256
1877 : (( ! UNI_SEMANTICS
1878 || invlist_highest(ssc->invlist) < 256)
1879 ? 128
1880 : NON_OTHER_COUNT);
1881 const U32 max_match = max_code_points / 2;
b35552de
KW
1882
1883 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1884
1885 invlist_iterinit(ssc->invlist);
1886 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1887 if (start >= max_code_points) {
1888 break;
b35552de 1889 }
72400949 1890 end = MIN(end, max_code_points - 1);
b35552de 1891 count += end - start + 1;
72400949 1892 if (count >= max_match) {
b35552de
KW
1893 invlist_iterfinish(ssc->invlist);
1894 return FALSE;
1895 }
1896 }
1897
1898 return TRUE;
1899}
1900
1901
b423522f
KW
1902STATIC void
1903S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1904{
1905 /* The inversion list in the SSC is marked mortal; now we need a more
1906 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1907 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1908 * map */
b423522f
KW
1909
1910 SV* invlist = invlist_clone(ssc->invlist);
1911
1912 PERL_ARGS_ASSERT_SSC_FINALIZE;
1913
71068078 1914 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1915
a0dd4231 1916 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1917 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1918 * by the time we reach here */
f240c685
KW
1919 assert(! (ANYOF_FLAGS(ssc)
1920 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
1921 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1922 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 1923
b423522f
KW
1924 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1925
1ee208c4
KW
1926 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1927 NULL, NULL, NULL, FALSE);
b423522f 1928
85c8e306
KW
1929 /* Make sure is clone-safe */
1930 ssc->invlist = NULL;
1931
e0e1be5f 1932 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1933 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1934 }
1462525b 1935
b2e90ddf
KW
1936 if (RExC_contains_locale) {
1937 OP(ssc) = ANYOFL;
1938 }
1939
1462525b 1940 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1941}
1942
a3621e74
YO
1943#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1944#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1945#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1946#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1947 ? (TRIE_LIST_CUR( idx ) - 1) \
1948 : 0 )
a3621e74 1949
3dab1dad
YO
1950
1951#ifdef DEBUGGING
07be1b83 1952/*
2b8b4781
NC
1953 dump_trie(trie,widecharmap,revcharmap)
1954 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1955 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1956
1957 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1958 The _interim_ variants are used for debugging the interim
1959 tables that are used to generate the final compressed
1960 representation which is what dump_trie expects.
1961
486ec47a 1962 Part of the reason for their existence is to provide a form
3dab1dad 1963 of documentation as to how the different representations function.
07be1b83
YO
1964
1965*/
3dab1dad
YO
1966
1967/*
3dab1dad
YO
1968 Dumps the final compressed table form of the trie to Perl_debug_log.
1969 Used for debugging make_trie().
1970*/
b9a59e08 1971
3dab1dad 1972STATIC void
2b8b4781
NC
1973S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1974 AV *revcharmap, U32 depth)
3dab1dad
YO
1975{
1976 U32 state;
ab3bbdeb 1977 SV *sv=sv_newmortal();
55eed653 1978 int colwidth= widecharmap ? 6 : 4;
2e64971a 1979 U16 word;
3dab1dad
YO
1980 GET_RE_DEBUG_FLAGS_DECL;
1981
7918f24d 1982 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1983
6ad9a8ab 1984 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 1985 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
1986
1987 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1988 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1989 if ( tmp ) {
6ad9a8ab 1990 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 1991 colwidth,
538e84ed 1992 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1993 PL_colors[0], PL_colors[1],
1994 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1995 PERL_PV_ESCAPE_FIRSTCHAR
1996 )
ab3bbdeb 1997 );
3dab1dad
YO
1998 }
1999 }
1e37780e
YO
2000 Perl_re_printf( aTHX_ "\n");
2001 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2002
2003 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2004 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2005 Perl_re_printf( aTHX_ "\n");
3dab1dad 2006
1e2e3d02 2007 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2008 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2009
147e3846 2010 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2011
2012 if ( trie->states[ state ].wordnum ) {
1e37780e 2013 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2014 } else {
6ad9a8ab 2015 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2016 }
2017
147e3846 2018 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2019
2020 if ( base ) {
2021 U32 ofs = 0;
2022
2023 while( ( base + ofs < trie->uniquecharcount ) ||
2024 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2025 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2026 != state))
3dab1dad
YO
2027 ofs++;
2028
147e3846 2029 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2030
2031 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2032 if ( ( base + ofs >= trie->uniquecharcount )
2033 && ( base + ofs - trie->uniquecharcount
2034 < trie->lasttrans )
2035 && trie->trans[ base + ofs
2036 - trie->uniquecharcount ].check == state )
3dab1dad 2037 {
147e3846 2038 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2039 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2040 );
3dab1dad 2041 } else {
6ad9a8ab 2042 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2043 }
2044 }
2045
6ad9a8ab 2046 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2047
2048 }
6ad9a8ab 2049 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2050 }
6ad9a8ab 2051 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2052 depth);
2e64971a 2053 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2054 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2055 (int)word, (int)(trie->wordinfo[word].prev),
2056 (int)(trie->wordinfo[word].len));
2057 }
6ad9a8ab 2058 Perl_re_printf( aTHX_ "\n" );
538e84ed 2059}
3dab1dad 2060/*
3dab1dad 2061 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2062 List tries normally only are used for construction when the number of
3dab1dad
YO
2063 possible chars (trie->uniquecharcount) is very high.
2064 Used for debugging make_trie().
2065*/
2066STATIC void
55eed653 2067S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2068 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2069 U32 depth)
3dab1dad
YO
2070{
2071 U32 state;
ab3bbdeb 2072 SV *sv=sv_newmortal();
55eed653 2073 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2074 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2075
2076 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2077
3dab1dad 2078 /* print out the table precompression. */
6ad9a8ab 2079 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2080 depth+1 );
6ad9a8ab 2081 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2082 depth+1, "------:-----+-----------------\n" );
538e84ed 2083
3dab1dad
YO
2084 for( state=1 ; state < next_alloc ; state ++ ) {
2085 U16 charid;
538e84ed 2086
147e3846 2087 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2088 depth+1, (UV)state );
3dab1dad 2089 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2090 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2091 } else {
6ad9a8ab 2092 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2093 trie->states[ state ].wordnum
2094 );
2095 }
2096 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2097 SV ** const tmp = av_fetch( revcharmap,
2098 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2099 if ( tmp ) {
147e3846 2100 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2101 colwidth,
538e84ed
KW
2102 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2103 colwidth,
2104 PL_colors[0], PL_colors[1],
2105 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2106 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2107 ) ,
1e2e3d02
YO
2108 TRIE_LIST_ITEM(state,charid).forid,
2109 (UV)TRIE_LIST_ITEM(state,charid).newstate
2110 );
538e84ed 2111 if (!(charid % 10))
6ad9a8ab 2112 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2113 (int)((depth * 2) + 14), "");
1e2e3d02 2114 }
ab3bbdeb 2115 }
6ad9a8ab 2116 Perl_re_printf( aTHX_ "\n");
3dab1dad 2117 }
538e84ed 2118}
3dab1dad
YO
2119
2120/*
3dab1dad 2121 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2122 This is the normal DFA style state transition table, with a few
2123 twists to facilitate compression later.
3dab1dad
YO
2124 Used for debugging make_trie().
2125*/
2126STATIC void
55eed653 2127S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2128 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2129 U32 depth)
3dab1dad
YO
2130{
2131 U32 state;
2132 U16 charid;
ab3bbdeb 2133 SV *sv=sv_newmortal();
55eed653 2134 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2135 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2136
2137 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2138
3dab1dad
YO
2139 /*
2140 print out the table precompression so that we can do a visual check
2141 that they are identical.
2142 */
538e84ed 2143
6ad9a8ab 2144 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2145
2146 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2147 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2148 if ( tmp ) {
6ad9a8ab 2149 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2150 colwidth,
538e84ed 2151 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2152 PL_colors[0], PL_colors[1],
2153 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2154 PERL_PV_ESCAPE_FIRSTCHAR
2155 )
ab3bbdeb 2156 );
3dab1dad
YO
2157 }
2158 }
2159
4aaafc03
YO
2160 Perl_re_printf( aTHX_ "\n");
2161 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2162
2163 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2164 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2165 }
2166
6ad9a8ab 2167 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2168
2169 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2170
147e3846 2171 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2172 depth+1,
3dab1dad
YO
2173 (UV)TRIE_NODENUM( state ) );
2174
2175 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2176 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2177 if (v)
147e3846 2178 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2179 else
6ad9a8ab 2180 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2181 }
2182 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2183 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2184 (UV)trie->trans[ state ].check );
3dab1dad 2185 } else {
147e3846 2186 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2187 (UV)trie->trans[ state ].check,
3dab1dad
YO
2188 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2189 }
2190 }
07be1b83 2191}
3dab1dad
YO
2192
2193#endif
2194
2e64971a 2195
786e8c11
YO
2196/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2197 startbranch: the first branch in the whole branch sequence
2198 first : start branch of sequence of branch-exact nodes.
2199 May be the same as startbranch
2200 last : Thing following the last branch.
2201 May be the same as tail.
2202 tail : item following the branch sequence
2203 count : words in the sequence
a4525e78 2204 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2205 depth : indent depth
3dab1dad 2206
786e8c11 2207Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2208
786e8c11
YO
2209A trie is an N'ary tree where the branches are determined by digital
2210decomposition of the key. IE, at the root node you look up the 1st character and
2211follow that branch repeat until you find the end of the branches. Nodes can be
2212marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2213
786e8c11 2214 /he|she|his|hers/
72f13be8 2215
786e8c11
YO
2216would convert into the following structure. Numbers represent states, letters
2217following numbers represent valid transitions on the letter from that state, if
2218the number is in square brackets it represents an accepting state, otherwise it
2219will be in parenthesis.
07be1b83 2220
786e8c11
YO
2221 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2222 | |
2223 | (2)
2224 | |
2225 (1) +-i->(6)-+-s->[7]
2226 |
2227 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2228
786e8c11
YO
2229 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2230
2231This shows that when matching against the string 'hers' we will begin at state 1
2232read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2233then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2234is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2235single traverse. We store a mapping from accepting to state to which word was
2236matched, and then when we have multiple possibilities we try to complete the
b8fda935 2237rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2238
2239The only prior NFA like behaviour that would be changed by the TRIE support is
2240the silent ignoring of duplicate alternations which are of the form:
2241
2242 / (DUPE|DUPE) X? (?{ ... }) Y /x
2243
4b714af6 2244Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2245and without the optimisation. With the optimisations dupes will be silently
486ec47a 2246ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2247the following demonstrates:
2248
2249 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2250
2251which prints out 'word' three times, but
2252
2253 'words'=~/(word|word|word)(?{ print $1 })S/
2254
2255which doesnt print it out at all. This is due to other optimisations kicking in.
2256
2257Example of what happens on a structural level:
2258
486ec47a 2259The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2260
2261 1: CURLYM[1] {1,32767}(18)
2262 5: BRANCH(8)
2263 6: EXACT <ac>(16)
2264 8: BRANCH(11)
2265 9: EXACT <ad>(16)
2266 11: BRANCH(14)
2267 12: EXACT <ab>(16)
2268 16: SUCCEED(0)
2269 17: NOTHING(18)
2270 18: END(0)
2271
2272This would be optimizable with startbranch=5, first=5, last=16, tail=16
2273and should turn into:
2274
2275 1: CURLYM[1] {1,32767}(18)
2276 5: TRIE(16)
2277 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2278 <ac>
2279 <ad>
2280 <ab>
2281 16: SUCCEED(0)
2282 17: NOTHING(18)
2283 18: END(0)
2284
2285Cases where tail != last would be like /(?foo|bar)baz/:
2286
2287 1: BRANCH(4)
2288 2: EXACT <foo>(8)
2289 4: BRANCH(7)
2290 5: EXACT <bar>(8)
2291 7: TAIL(8)
2292 8: EXACT <baz>(10)
2293 10: END(0)
2294
2295which would be optimizable with startbranch=1, first=1, last=7, tail=8
2296and would end up looking like:
2297
2298 1: TRIE(8)
2299 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2300 <foo>
2301 <bar>
2302 7: TAIL(8)
2303 8: EXACT <baz>(10)
2304 10: END(0)
2305
c80e42f3 2306 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2307
2308is the recommended Unicode-aware way of saying
2309
2310 *(d++) = uv;
2311*/
2312
fab2782b 2313#define TRIE_STORE_REVCHAR(val) \
786e8c11 2314 STMT_START { \
73031816 2315 if (UTF) { \
668fcfea 2316 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2317 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2318 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2319 SvCUR_set(zlopp, kapow - flrbbbbb); \
2320 SvPOK_on(zlopp); \
2321 SvUTF8_on(zlopp); \
2322 av_push(revcharmap, zlopp); \
2323 } else { \
fab2782b 2324 char ooooff = (char)val; \
73031816
NC
2325 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2326 } \
2327 } STMT_END
786e8c11 2328
914a25d5
KW
2329/* This gets the next character from the input, folding it if not already
2330 * folded. */
2331#define TRIE_READ_CHAR STMT_START { \
2332 wordlen++; \
2333 if ( UTF ) { \
2334 /* if it is UTF then it is either already folded, or does not need \
2335 * folding */ \
1c1d615a 2336 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2337 } \
2338 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2339 /* This folder implies Unicode rules, which in the range expressible \
2340 * by not UTF is the lower case, with the two exceptions, one of \
2341 * which should have been taken care of before calling this */ \
2342 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2343 uvc = toLOWER_L1(*uc); \
2344 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2345 len = 1; \
914a25d5
KW
2346 } else { \
2347 /* raw data, will be folded later if needed */ \
2348 uvc = (U32)*uc; \
2349 len = 1; \
2350 } \
786e8c11
YO
2351} STMT_END
2352
2353
2354
2355#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2356 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2357 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2358 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2359 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2360 } \
2361 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2362 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2363 TRIE_LIST_CUR( state )++; \
2364} STMT_END
07be1b83 2365
786e8c11
YO
2366#define TRIE_LIST_NEW(state) STMT_START { \
2367 Newxz( trie->states[ state ].trans.list, \
2368 4, reg_trie_trans_le ); \
2369 TRIE_LIST_CUR( state ) = 1; \
2370 TRIE_LIST_LEN( state ) = 4; \
2371} STMT_END
07be1b83 2372
786e8c11
YO
2373#define TRIE_HANDLE_WORD(state) STMT_START { \
2374 U16 dupe= trie->states[ state ].wordnum; \
2375 regnode * const noper_next = regnext( noper ); \
2376 \
786e8c11
YO
2377 DEBUG_r({ \
2378 /* store the word for dumping */ \
2379 SV* tmp; \
2380 if (OP(noper) != NOTHING) \
740cce10 2381 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2382 else \
740cce10 2383 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2384 av_push( trie_words, tmp ); \
786e8c11
YO
2385 }); \
2386 \
2387 curword++; \
2e64971a
DM
2388 trie->wordinfo[curword].prev = 0; \
2389 trie->wordinfo[curword].len = wordlen; \
2390 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2391 \
2392 if ( noper_next < tail ) { \
2393 if (!trie->jump) \
538e84ed
KW
2394 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2395 sizeof(U16) ); \
7f69552c 2396 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2397 if (!jumper) \
2398 jumper = noper_next; \
2399 if (!nextbranch) \
2400 nextbranch= regnext(cur); \
2401 } \
2402 \
2403 if ( dupe ) { \
2e64971a
DM
2404 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2405 /* chain, so that when the bits of chain are later */\
2406 /* linked together, the dups appear in the chain */\
2407 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2408 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2409 } else { \
2410 /* we haven't inserted this word yet. */ \
2411 trie->states[ state ].wordnum = curword; \
2412 } \
2413} STMT_END
07be1b83 2414
3dab1dad 2415
786e8c11
YO
2416#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2417 ( ( base + charid >= ucharcount \
2418 && base + charid < ubound \
2419 && state == trie->trans[ base - ucharcount + charid ].check \
2420 && trie->trans[ base - ucharcount + charid ].next ) \
2421 ? trie->trans[ base - ucharcount + charid ].next \
2422 : ( state==1 ? special : 0 ) \
2423 )
3dab1dad 2424
8bcafbf4
YO
2425#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2426STMT_START { \
2427 TRIE_BITMAP_SET(trie, uvc); \
2428 /* store the folded codepoint */ \
2429 if ( folder ) \
2430 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2431 \
2432 if ( !UTF ) { \
2433 /* store first byte of utf8 representation of */ \
2434 /* variant codepoints */ \
2435 if (! UVCHR_IS_INVARIANT(uvc)) { \
2436 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2437 } \
2438 } \
2439} STMT_END
786e8c11
YO
2440#define MADE_TRIE 1
2441#define MADE_JUMP_TRIE 2
2442#define MADE_EXACT_TRIE 4
3dab1dad 2443
a3621e74 2444STATIC I32
538e84ed
KW
2445S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2446 regnode *first, regnode *last, regnode *tail,
2447 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2448{
2449 /* first pass, loop through and scan words */
2450 reg_trie_data *trie;
55eed653 2451 HV *widecharmap = NULL;
2b8b4781 2452 AV *revcharmap = newAV();
a3621e74 2453 regnode *cur;
a3621e74
YO
2454 STRLEN len = 0;
2455 UV uvc = 0;
2456 U16 curword = 0;
2457 U32 next_alloc = 0;
786e8c11
YO
2458 regnode *jumper = NULL;
2459 regnode *nextbranch = NULL;
7f69552c 2460 regnode *convert = NULL;
2e64971a 2461 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2462 /* we just use folder as a flag in utf8 */
1e696034 2463 const U8 * folder = NULL;
a3621e74 2464
2b8b4781 2465#ifdef DEBUGGING
cf78de0b 2466 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2467 AV *trie_words = NULL;
2468 /* along with revcharmap, this only used during construction but both are
2469 * useful during debugging so we store them in the struct when debugging.
8e11feef 2470 */
2b8b4781 2471#else
cf78de0b 2472 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2473 STRLEN trie_charcount=0;
3dab1dad 2474#endif
2b8b4781 2475 SV *re_trie_maxbuff;
a3621e74 2476 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2477
2478 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2479#ifndef DEBUGGING
2480 PERL_UNUSED_ARG(depth);
2481#endif
a3621e74 2482
1e696034 2483 switch (flags) {
a4525e78 2484 case EXACT: case EXACTL: break;
2f7f8cb1 2485 case EXACTFA:
fab2782b 2486 case EXACTFU_SS:
a4525e78
KW
2487 case EXACTFU:
2488 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2489 case EXACTF: folder = PL_fold; break;
fab2782b 2490 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2491 }
2492
c944940b 2493 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2494 trie->refcount = 1;
3dab1dad 2495 trie->startstate = 1;
786e8c11 2496 trie->wordcount = word_count;
f8fc2ecf 2497 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2498 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2499 if (flags == EXACT || flags == EXACTL)
c944940b 2500 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2501 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2502 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2503
a3621e74 2504 DEBUG_r({
2b8b4781 2505 trie_words = newAV();
a3621e74 2506 });
a3621e74 2507
0111c4fd 2508 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2509 assert(re_trie_maxbuff);
a3621e74 2510 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2511 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2512 }
df826430 2513 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2514 Perl_re_indentf( aTHX_
cb41e5d6
YO
2515 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2516 depth+1,
538e84ed
KW
2517 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2518 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2519 });
538e84ed 2520
7f69552c
YO
2521 /* Find the node we are going to overwrite */
2522 if ( first == startbranch && OP( last ) != BRANCH ) {
2523 /* whole branch chain */
2524 convert = first;
2525 } else {
2526 /* branch sub-chain */
2527 convert = NEXTOPER( first );
2528 }
538e84ed 2529
a3621e74
YO
2530 /* -- First loop and Setup --
2531
2532 We first traverse the branches and scan each word to determine if it
2533 contains widechars, and how many unique chars there are, this is
2534 important as we have to build a table with at least as many columns as we
2535 have unique chars.
2536
2537 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2538 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2539 the native representation of the character value as the key and IV's for
2540 the coded index.
a3621e74
YO
2541
2542 *TODO* If we keep track of how many times each character is used we can
2543 remap the columns so that the table compression later on is more
3b753521 2544 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2545 middle and the least common are on the outside. IMO this would be better
2546 than a most to least common mapping as theres a decent chance the most
2547 common letter will share a node with the least common, meaning the node
486ec47a 2548 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2549 case is when we have the least common nodes twice.
2550
2551 */
2552
a3621e74 2553 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2554 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2555 const U8 *uc;
2556 const U8 *e;
bc031a7d 2557 int foldlen = 0;
07be1b83 2558 U32 wordlen = 0; /* required init */
bc031a7d
KW
2559 STRLEN minchars = 0;
2560 STRLEN maxchars = 0;
538e84ed
KW
2561 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2562 bitmap?*/
a3621e74 2563
3dab1dad 2564 if (OP(noper) == NOTHING) {
20ed8c88
YO
2565 /* skip past a NOTHING at the start of an alternation
2566 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2567 */
df826430 2568 regnode *noper_next= regnext(noper);
944e05e3
YO
2569 if (noper_next < tail)
2570 noper= noper_next;
2571 }
2572
dca5fc4c
YO
2573 if ( noper < tail &&
2574 (
2575 OP(noper) == flags ||
2576 (
2577 flags == EXACTFU &&
2578 OP(noper) == EXACTFU_SS
2579 )
2580 )
2581 ) {
944e05e3
YO
2582 uc= (U8*)STRING(noper);
2583 e= uc + STR_LEN(noper);
2584 } else {
2585 trie->minlen= 0;
2586 continue;
3dab1dad 2587 }
df826430 2588
944e05e3 2589
fab2782b 2590 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2591 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2592 regardless of encoding */
fab2782b
YO
2593 if (OP( noper ) == EXACTFU_SS) {
2594 /* false positives are ok, so just set this */
0dc4a61d 2595 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2596 }
2597 }
dca5fc4c 2598
bc031a7d
KW
2599 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2600 branch */
3dab1dad 2601 TRIE_CHARCOUNT(trie)++;
a3621e74 2602 TRIE_READ_CHAR;
645de4ce 2603
bc031a7d
KW
2604 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2605 * is in effect. Under /i, this character can match itself, or
2606 * anything that folds to it. If not under /i, it can match just
2607 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2608 * all fold to k, and all are single characters. But some folds
2609 * expand to more than one character, so for example LATIN SMALL
2610 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2611 * the string beginning at 'uc' is 'ffi', it could be matched by
2612 * three characters, or just by the one ligature character. (It
2613 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2614 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2615 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2616 * match.) The trie needs to know the minimum and maximum number
2617 * of characters that could match so that it can use size alone to
2618 * quickly reject many match attempts. The max is simple: it is
2619 * the number of folded characters in this branch (since a fold is
2620 * never shorter than what folds to it. */
2621
2622 maxchars++;
2623
2624 /* And the min is equal to the max if not under /i (indicated by
2625 * 'folder' being NULL), or there are no multi-character folds. If
2626 * there is a multi-character fold, the min is incremented just
2627 * once, for the character that folds to the sequence. Each
2628 * character in the sequence needs to be added to the list below of
2629 * characters in the trie, but we count only the first towards the
2630 * min number of characters needed. This is done through the
2631 * variable 'foldlen', which is returned by the macros that look
2632 * for these sequences as the number of bytes the sequence
2633 * occupies. Each time through the loop, we decrement 'foldlen' by
2634 * how many bytes the current char occupies. Only when it reaches
2635 * 0 do we increment 'minchars' or look for another multi-character
2636 * sequence. */
2637 if (folder == NULL) {
2638 minchars++;
2639 }
2640 else if (foldlen > 0) {
2641 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2642 }
2643 else {
bc031a7d
KW
2644 minchars++;
2645
2646 /* See if *uc is the beginning of a multi-character fold. If
2647 * so, we decrement the length remaining to look at, to account
2648 * for the current character this iteration. (We can use 'uc'
2649 * instead of the fold returned by TRIE_READ_CHAR because for
2650 * non-UTF, the latin1_safe macro is smart enough to account
2651 * for all the unfolded characters, and because for UTF, the
2652 * string will already have been folded earlier in the
2653 * compilation process */
2654 if (UTF) {
2655 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2656 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2657 }
2658 }
bc031a7d
KW
2659 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2660 foldlen--;
2661 }
645de4ce 2662 }
bc031a7d
KW
2663
2664 /* The current character (and any potential folds) should be added
2665 * to the possible matching characters for this position in this
2666 * branch */
a3621e74 2667 if ( uvc < 256 ) {
fab2782b
YO
2668 if ( folder ) {
2669 U8 folded= folder[ (U8) uvc ];
2670 if ( !trie->charmap[ folded ] ) {
2671 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2672 TRIE_STORE_REVCHAR( folded );
2673 }
2674 }
a3621e74
YO
2675 if ( !trie->charmap[ uvc ] ) {
2676 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2677 TRIE_STORE_REVCHAR( uvc );
a3621e74 2678 }
02daf0ab 2679 if ( set_bit ) {
62012aee
KW
2680 /* store the codepoint in the bitmap, and its folded
2681 * equivalent. */
8bcafbf4 2682 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2683 set_bit = 0; /* We've done our bit :-) */
2684 }
a3621e74 2685 } else {
bc031a7d
KW
2686
2687 /* XXX We could come up with the list of code points that fold
2688 * to this using PL_utf8_foldclosures, except not for
2689 * multi-char folds, as there may be multiple combinations
2690 * there that could work, which needs to wait until runtime to
2691 * resolve (The comment about LIGATURE FFI above is such an
2692 * example */
2693
a3621e74 2694 SV** svpp;
55eed653
NC
2695 if ( !widecharmap )
2696 widecharmap = newHV();
a3621e74 2697
55eed653 2698 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2699
2700 if ( !svpp )
147e3846 2701 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2702
2703 if ( !SvTRUE( *svpp ) ) {
2704 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2705 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2706 }
2707 }
bc031a7d
KW
2708 } /* end loop through characters in this branch of the trie */
2709
2710 /* We take the min and max for this branch and combine to find the min
2711 * and max for all branches processed so far */
3dab1dad 2712 if( cur == first ) {
bc031a7d
KW
2713 trie->minlen = minchars;
2714 trie->maxlen = maxchars;
2715 } else if (minchars < trie->minlen) {
2716 trie->minlen = minchars;
2717 } else if (maxchars > trie->maxlen) {
2718 trie->maxlen = maxchars;
fab2782b 2719 }
a3621e74
YO
2720 } /* end first pass */
2721 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2722 Perl_re_indentf( aTHX_
cb41e5d6
YO
2723 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2724 depth+1,
55eed653 2725 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2726 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2727 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2728 );
a3621e74
YO
2729
2730 /*
2731 We now know what we are dealing with in terms of unique chars and
2732 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2733 representation using a flat table will take. If it's over a reasonable
2734 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2735 conservative but potentially much slower representation using an array
2736 of lists.
2737
2738 At the end we convert both representations into the same compressed
2739 form that will be used in regexec.c for matching with. The latter
2740 is a form that cannot be used to construct with but has memory
2741 properties similar to the list form and access properties similar
2742 to the table form making it both suitable for fast searches and
2743 small enough that its feasable to store for the duration of a program.
2744
2745 See the comment in the code where the compressed table is produced
2746 inplace from the flat tabe representation for an explanation of how
2747 the compression works.
2748
2749 */
2750
2751
2e64971a
DM
2752 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2753 prev_states[1] = 0;
2754
538e84ed
KW
2755 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2756 > SvIV(re_trie_maxbuff) )
2757 {
a3621e74
YO
2758 /*
2759 Second Pass -- Array Of Lists Representation
2760
2761 Each state will be represented by a list of charid:state records
2762 (reg_trie_trans_le) the first such element holds the CUR and LEN
2763 points of the allocated array. (See defines above).
2764
2765 We build the initial structure using the lists, and then convert
2766 it into the compressed table form which allows faster lookups
2767 (but cant be modified once converted).
a3621e74
YO
2768 */
2769
a3621e74
YO
2770 STRLEN transcount = 1;
2771
6ad9a8ab 2772 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2773 depth+1));
686b73d4 2774
c944940b
JH
2775 trie->states = (reg_trie_state *)
2776 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2777 sizeof(reg_trie_state) );
a3621e74
YO
2778 TRIE_LIST_NEW(1);
2779 next_alloc = 2;
2780
2781 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2782
df826430 2783 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2784 U32 state = 1; /* required init */
2785 U16 charid = 0; /* sanity init */
07be1b83 2786 U32 wordlen = 0; /* required init */
c445ea15 2787
df826430
YO
2788 if (OP(noper) == NOTHING) {
2789 regnode *noper_next= regnext(noper);
944e05e3
YO
2790 if (noper_next < tail)
2791 noper= noper_next;
df826430
YO
2792 }
2793
944e05e3
YO
2794 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2795 const U8 *uc= (U8*)STRING(noper);
2796 const U8 *e= uc + STR_LEN(noper);
2797
786e8c11 2798 for ( ; uc < e ; uc += len ) {
c445ea15 2799
786e8c11 2800 TRIE_READ_CHAR;
c445ea15 2801
786e8c11
YO
2802 if ( uvc < 256 ) {
2803 charid = trie->charmap[ uvc ];
c445ea15 2804 } else {
538e84ed
KW
2805 SV** const svpp = hv_fetch( widecharmap,
2806 (char*)&uvc,
2807 sizeof( UV ),
2808 0);
786e8c11
YO
2809 if ( !svpp ) {
2810 charid = 0;
2811 } else {
2812 charid=(U16)SvIV( *svpp );
2813 }
c445ea15 2814 }
538e84ed
KW
2815 /* charid is now 0 if we dont know the char read, or
2816 * nonzero if we do */
786e8c11 2817 if ( charid ) {
a3621e74 2818
786e8c11
YO
2819 U16 check;
2820 U32 newstate = 0;
a3621e74 2821
786e8c11
YO
2822 charid--;
2823 if ( !trie->states[ state ].trans.list ) {
2824 TRIE_LIST_NEW( state );
c445ea15 2825 }
538e84ed
KW
2826 for ( check = 1;
2827 check <= TRIE_LIST_USED( state );
2828 check++ )
2829 {
2830 if ( TRIE_LIST_ITEM( state, check ).forid
2831 == charid )
2832 {
786e8c11
YO
2833 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2834 break;
2835 }
2836 }
2837 if ( ! newstate ) {
2838 newstate = next_alloc++;
2e64971a 2839 prev_states[newstate] = state;
786e8c11
YO
2840 TRIE_LIST_PUSH( state, charid, newstate );
2841 transcount++;
2842 }
2843 state = newstate;
2844 } else {
147e3846 2845 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 2846 }
a28509cc 2847 }
c445ea15 2848 }
3dab1dad 2849 TRIE_HANDLE_WORD(state);
a3621e74
YO
2850
2851 } /* end second pass */
2852
1e2e3d02 2853 /* next alloc is the NEXT state to be allocated */
538e84ed 2854 trie->statecount = next_alloc;
c944940b
JH
2855 trie->states = (reg_trie_state *)
2856 PerlMemShared_realloc( trie->states,
2857 next_alloc
2858 * sizeof(reg_trie_state) );
a3621e74 2859
3dab1dad 2860 /* and now dump it out before we compress it */
2b8b4781
NC
2861 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2862 revcharmap, next_alloc,
2863 depth+1)
1e2e3d02 2864 );
a3621e74 2865
c944940b
JH
2866 trie->trans = (reg_trie_trans *)
2867 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2868 {
2869 U32 state;
a3621e74
YO
2870 U32 tp = 0;
2871 U32 zp = 0;
2872
2873
2874 for( state=1 ; state < next_alloc ; state ++ ) {
2875 U32 base=0;
2876
2877 /*
2878 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2879 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
2880 );
2881 */
2882
2883 if (trie->states[state].trans.list) {
2884 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2885 U16 maxid=minid;
a28509cc 2886 U16 idx;
a3621e74
YO
2887
2888 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2889 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2890 if ( forid < minid ) {
2891 minid=forid;
2892 } else if ( forid > maxid ) {
2893 maxid=forid;
2894 }
a3621e74
YO
2895 }
2896 if ( transcount < tp + maxid - minid + 1) {
2897 transcount *= 2;
c944940b
JH
2898 trie->trans = (reg_trie_trans *)
2899 PerlMemShared_realloc( trie->trans,
446bd890
NC
2900 transcount
2901 * sizeof(reg_trie_trans) );
538e84ed
KW
2902 Zero( trie->trans + (transcount / 2),
2903 transcount / 2,
2904 reg_trie_trans );
a3621e74
YO
2905 }
2906 base = trie->uniquecharcount + tp - minid;
2907 if ( maxid == minid ) {
2908 U32 set = 0;
2909 for ( ; zp < tp ; zp++ ) {
2910 if ( ! trie->trans[ zp ].next ) {
2911 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2912 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2913 1).newstate;
a3621e74
YO
2914 trie->trans[ zp ].check = state;
2915 set = 1;
2916 break;
2917 }
2918 }
2919 if ( !set ) {
538e84ed
KW
2920 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2921 1).newstate;
a3621e74
YO
2922 trie->trans[ tp ].check = state;
2923 tp++;
2924 zp = tp;
2925 }
2926 } else {
2927 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2928 const U32 tid = base
2929 - trie->uniquecharcount
2930 + TRIE_LIST_ITEM( state, idx ).forid;
2931 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2932 idx ).newstate;
a3621e74
YO
2933 trie->trans[ tid ].check = state;
2934 }
2935 tp += ( maxid - minid + 1 );
2936 }
2937 Safefree(trie->states[ state ].trans.list);
2938 }
2939 /*
2940 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2941 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
2942 );
2943 */
2944 trie->states[ state ].trans.base=base;
2945 }
cc601c31 2946 trie->lasttrans = tp + 1;
a3621e74
YO
2947 }
2948 } else {
2949 /*
2950 Second Pass -- Flat Table Representation.
2951
b423522f
KW
2952 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2953 each. We know that we will need Charcount+1 trans at most to store
2954 the data (one row per char at worst case) So we preallocate both
2955 structures assuming worst case.
a3621e74
YO
2956
2957 We then construct the trie using only the .next slots of the entry
2958 structs.
2959
b423522f
KW
2960 We use the .check field of the first entry of the node temporarily
2961 to make compression both faster and easier by keeping track of how
2962 many non zero fields are in the node.
a3621e74
YO
2963
2964 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2965 transition.
2966
b423522f
KW
2967 There are two terms at use here: state as a TRIE_NODEIDX() which is
2968 a number representing the first entry of the node, and state as a
2969 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2970 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2971 if there are 2 entrys per node. eg:
a3621e74
YO
2972
2973 A B A B
2974 1. 2 4 1. 3 7
2975 2. 0 3 3. 0 5
2976 3. 0 0 5. 0 0
2977 4. 0 0 7. 0 0
2978
b423522f
KW
2979 The table is internally in the right hand, idx form. However as we
2980 also have to deal with the states array which is indexed by nodenum
2981 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2982
2983 */
6ad9a8ab 2984 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 2985 depth+1));
3dab1dad 2986
c944940b
JH
2987 trie->trans = (reg_trie_trans *)
2988 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2989 * trie->uniquecharcount + 1,
2990 sizeof(reg_trie_trans) );
2991 trie->states = (reg_trie_state *)
2992 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2993 sizeof(reg_trie_state) );
a3621e74
YO
2994 next_alloc = trie->uniquecharcount + 1;
2995
3dab1dad 2996
a3621e74
YO
2997 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2998
df826430 2999 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3000
3001 U32 state = 1; /* required init */
3002
3003 U16 charid = 0; /* sanity init */
3004 U32 accept_state = 0; /* sanity init */
a3621e74 3005
07be1b83 3006 U32 wordlen = 0; /* required init */
a3621e74 3007
df826430
YO
3008 if (OP(noper) == NOTHING) {
3009 regnode *noper_next= regnext(noper);
944e05e3
YO
3010 if (noper_next < tail)
3011 noper= noper_next;
df826430 3012 }
fab2782b 3013
944e05e3
YO
3014 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3015 const U8 *uc= (U8*)STRING(noper);
3016 const U8 *e= uc + STR_LEN(noper);
3017
786e8c11 3018 for ( ; uc < e ; uc += len ) {
a3621e74 3019
786e8c11 3020 TRIE_READ_CHAR;
a3621e74 3021
786e8c11
YO
3022 if ( uvc < 256 ) {
3023 charid = trie->charmap[ uvc ];
3024 } else {
538e84ed
KW
3025 SV* const * const svpp = hv_fetch( widecharmap,
3026 (char*)&uvc,
3027 sizeof( UV ),
3028 0);
786e8c11 3029 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3030 }
786e8c11
YO
3031 if ( charid ) {
3032 charid--;
3033 if ( !trie->trans[ state + charid ].next ) {
3034 trie->trans[ state + charid ].next = next_alloc;
3035 trie->trans[ state ].check++;
2e64971a
DM
3036 prev_states[TRIE_NODENUM(next_alloc)]
3037 = TRIE_NODENUM(state);
786e8c11
YO
3038 next_alloc += trie->uniquecharcount;
3039 }
3040 state = trie->trans[ state + charid ].next;
3041 } else {
147e3846 3042 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3043 }
538e84ed
KW
3044 /* charid is now 0 if we dont know the char read, or
3045 * nonzero if we do */
a3621e74 3046 }
a3621e74 3047 }
3dab1dad
YO
3048 accept_state = TRIE_NODENUM( state );
3049 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3050
3051 } /* end second pass */
3052
3dab1dad 3053 /* and now dump it out before we compress it */
2b8b4781
NC
3054 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3055 revcharmap,
3056 next_alloc, depth+1));
a3621e74 3057
a3621e74
YO
3058 {
3059 /*
3060 * Inplace compress the table.*
3061
3062 For sparse data sets the table constructed by the trie algorithm will
3063 be mostly 0/FAIL transitions or to put it another way mostly empty.
3064 (Note that leaf nodes will not contain any transitions.)
3065
3066 This algorithm compresses the tables by eliminating most such
3067 transitions, at the cost of a modest bit of extra work during lookup:
3068
3069 - Each states[] entry contains a .base field which indicates the
3070 index in the state[] array wheres its transition data is stored.
3071
3b753521 3072 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3073
3074 - If .base is nonzero then charid is added to it to find an entry in
3075 the trans array.
3076
3077 -If trans[states[state].base+charid].check!=state then the
3078 transition is taken to be a 0/Fail transition. Thus if there are fail
3079 transitions at the front of the node then the .base offset will point
3080 somewhere inside the previous nodes data (or maybe even into a node
3081 even earlier), but the .check field determines if the transition is
3082 valid.
3083
786e8c11 3084 XXX - wrong maybe?
a3621e74 3085 The following process inplace converts the table to the compressed
3b753521 3086 table: We first do not compress the root node 1,and mark all its
a3621e74 3087 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3088 allows us to do a DFA construction from the compressed table later,
3089 and ensures that any .base pointers we calculate later are greater
3090 than 0.
a3621e74
YO
3091
3092 - We set 'pos' to indicate the first entry of the second node.
3093
3094 - We then iterate over the columns of the node, finding the first and
3095 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3096 and set the .check pointers accordingly, and advance pos
3097 appropriately and repreat for the next node. Note that when we copy
3098 the next pointers we have to convert them from the original
3099 NODEIDX form to NODENUM form as the former is not valid post
3100 compression.
3101
3102 - If a node has no transitions used we mark its base as 0 and do not
3103 advance the pos pointer.
3104
3105 - If a node only has one transition we use a second pointer into the
3106 structure to fill in allocated fail transitions from other states.
3107 This pointer is independent of the main pointer and scans forward
3108 looking for null transitions that are allocated to a state. When it
3109 finds one it writes the single transition into the "hole". If the
786e8c11 3110 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3111
3112 - Once compressed we can Renew/realloc the structures to release the
3113 excess space.
3114
3115 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3116 specifically Fig 3.47 and the associated pseudocode.
3117
3118 demq
3119 */
a3b680e6 3120 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3121 U32 state, charid;
a3621e74 3122 U32 pos = 0, zp=0;
1e2e3d02 3123 trie->statecount = laststate;
a3621e74
YO
3124
3125 for ( state = 1 ; state < laststate ; state++ ) {
3126 U8 flag = 0;
a28509cc
AL
3127 const U32 stateidx = TRIE_NODEIDX( state );
3128 const U32 o_used = trie->trans[ stateidx ].check;
3129 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3130 trie->trans[ stateidx ].check = 0;
3131
538e84ed
KW
3132 for ( charid = 0;
3133 used && charid < trie->uniquecharcount;
3134 charid++ )
3135 {
a3621e74
YO
3136 if ( flag || trie->trans[ stateidx + charid ].next ) {
3137 if ( trie->trans[ stateidx + charid ].next ) {
3138 if (o_used == 1) {
3139 for ( ; zp < pos ; zp++ ) {
3140 if ( ! trie->trans[ zp ].next ) {
3141 break;
3142 }
3143 }
538e84ed
KW
3144 trie->states[ state ].trans.base
3145 = zp
3146 + trie->uniquecharcount
3147 - charid ;
3148 trie->trans[ zp ].next
3149 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3150 + charid ].next );
a3621e74
YO
3151 trie->trans[ zp ].check = state;
3152 if ( ++zp > pos ) pos = zp;
3153 break;
3154 }
3155 used--;
3156 }
3157 if ( !flag ) {
3158 flag = 1;
538e84ed
KW
3159 trie->states[ state ].trans.base
3160 = pos + trie->uniquecharcount - charid ;
a3621e74 3161 }
538e84ed
KW
3162 trie->trans[ pos ].next
3163 = SAFE_TRIE_NODENUM(
3164 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3165 trie->trans[ pos ].check = state;
3166 pos++;
3167 }
3168 }
3169 }
cc601c31 3170 trie->lasttrans = pos + 1;
c944940b
JH
3171 trie->states = (reg_trie_state *)
3172 PerlMemShared_realloc( trie->states, laststate
3173 * sizeof(reg_trie_state) );
a3621e74 3174 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3175 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
cb41e5d6 3176 depth+1,
538e84ed
KW
3177 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3178 + 1 ),
3179 (IV)next_alloc,
3180 (IV)pos,
3181 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3182 );
3183
3184 } /* end table compress */
3185 }
1e2e3d02 3186 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3187 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
cb41e5d6 3188 depth+1,
1e2e3d02
YO
3189 (UV)trie->statecount,
3190 (UV)trie->lasttrans)
3191 );
cc601c31 3192 /* resize the trans array to remove unused space */
c944940b
JH
3193 trie->trans = (reg_trie_trans *)
3194 PerlMemShared_realloc( trie->trans, trie->lasttrans
3195 * sizeof(reg_trie_trans) );
a3621e74 3196
538e84ed 3197 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3198 U8 nodetype =(U8)(flags & 0xFF);
3199 char *str=NULL;
538e84ed 3200
07be1b83 3201#ifdef DEBUGGING
e62cc96a 3202 regnode *optimize = NULL;
7122b237
YO
3203#ifdef RE_TRACK_PATTERN_OFFSETS
3204
b57a0404
JH
3205 U32 mjd_offset = 0;
3206 U32 mjd_nodelen = 0;
7122b237
YO
3207#endif /* RE_TRACK_PATTERN_OFFSETS */
3208#endif /* DEBUGGING */
a3621e74 3209 /*
3dab1dad
YO
3210 This means we convert either the first branch or the first Exact,
3211 depending on whether the thing following (in 'last') is a branch
3212 or not and whther first is the startbranch (ie is it a sub part of
3213 the alternation or is it the whole thing.)
3b753521 3214 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3215 the whole branch sequence, including the first.
a3621e74 3216 */
3dab1dad 3217 /* Find the node we are going to overwrite */
7f69552c 3218 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3219 /* branch sub-chain */
3dab1dad 3220 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3221#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3222 DEBUG_r({
3223 mjd_offset= Node_Offset((convert));
3224 mjd_nodelen= Node_Length((convert));
3225 });
7122b237 3226#endif
7f69552c 3227 /* whole branch chain */
7122b237
YO
3228 }
3229#ifdef RE_TRACK_PATTERN_OFFSETS
3230 else {
7f69552c
YO
3231 DEBUG_r({
3232 const regnode *nop = NEXTOPER( convert );
3233 mjd_offset= Node_Offset((nop));
3234 mjd_nodelen= Node_Length((nop));
3235 });
07be1b83
YO
3236 }
3237 DEBUG_OPTIMISE_r(
147e3846 3238 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
cb41e5d6 3239 depth+1,
786e8c11 3240 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3241 );
7122b237 3242#endif
538e84ed 3243 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3244 split out as an EXACT and put in front of the TRIE node. */
3245 trie->startstate= 1;
55eed653 3246 if ( trie->bitmap && !widecharmap && !trie->jump ) {
5ee57374
YO
3247 /* we want to find the first state that has more than
3248 * one transition, if that state is not the first state
3249 * then we have a common prefix which we can remove.
3250 */
3dab1dad 3251 U32 state;
1e2e3d02 3252 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3253 U32 ofs = 0;
ee0dfd0b
YO
3254 I32 first_ofs = -1; /* keeps track of the ofs of the first
3255 transition, -1 means none */
8e11feef
RGS
3256 U32 count = 0;
3257 const U32 base = trie->states[ state ].trans.base;
a3621e74 3258
5ee57374 3259 /* does this state terminate an alternation? */
3dab1dad 3260 if ( trie->states[state].wordnum )
8e11feef 3261 count = 1;
a3621e74 3262
8e11feef 3263 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3264 if ( ( base + ofs >= trie->uniquecharcount ) &&
3265 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3266 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3267 {
3dab1dad 3268 if ( ++count > 1 ) {
5ee57374 3269 /* we have more than one transition */
d3d91c7c
YO
3270 SV **tmp;
3271 U8 *ch;
5ee57374
YO
3272 /* if this is the first state there is no common prefix
3273 * to extract, so we can exit */
8e11feef 3274 if ( state == 1 ) break;
d3d91c7c
YO
3275 tmp = av_fetch( revcharmap, ofs, 0);
3276 ch = (U8*)SvPV_nolen_const( *tmp );
3277
5ee57374
YO
3278 /* if we are on count 2 then we need to initialize the
3279 * bitmap, and store the previous char if there was one
3280 * in it*/
3dab1dad 3281 if ( count == 2 ) {
5ee57374 3282 /* clear the bitmap */
3dab1dad
YO
3283 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3284 DEBUG_OPTIMISE_r(
147e3846 3285 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
cb41e5d6 3286 depth+1,
786e8c11 3287 (UV)state));
ee0dfd0b
YO
3288 if (first_ofs >= 0) {
3289 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
be8e71aa 3290 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3291
8bcafbf4 3292 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3dab1dad 3293 DEBUG_OPTIMISE_r(
6ad9a8ab 3294 Perl_re_printf( aTHX_ "%s", (char*)ch)
3dab1dad 3295 );
8e11feef
RGS
3296 }
3297 }
8bcafbf4
YO
3298 /* store the current firstchar in the bitmap */
3299 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
6ad9a8ab 3300 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
8e11feef 3301 }
ee0dfd0b 3302 first_ofs = ofs;
8e11feef 3303 }
3dab1dad
YO
3304 }
3305 if ( count == 1 ) {
5ee57374
YO
3306 /* This state has only one transition, its transition is part
3307 * of a common prefix - we need to concatenate the char it
3308 * represents to what we have so far. */
ee0dfd0b 3309 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
c490c714
YO
3310 STRLEN len;
3311 char *ch = SvPV( *tmp, len );
de734bd5
A
3312 DEBUG_OPTIMISE_r({
3313 SV *sv=sv_newmortal();
147e3846 3314 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
cb41e5d6 3315 depth+1,
ee0dfd0b 3316 (UV)state, (UV)first_ofs,
538e84ed 3317 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3318 PL_colors[0], PL_colors[1],
3319 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3320 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3321 )
3322 );
3323 });
3dab1dad
YO
3324 if ( state==1 ) {
3325 OP( convert ) = nodetype;
3326 str=STRING(convert);
3327 STR_LEN(convert)=0;
3328 }
c490c714
YO
3329 STR_LEN(convert) += len;
3330 while (len--)
de734bd5 3331 *str++ = *ch++;
8e11feef 3332 } else {
538e84ed 3333#ifdef DEBUGGING
8e11feef 3334 if (state>1)
6ad9a8ab 3335 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
f9049ba1 3336#endif
8e11feef
RGS
3337 break;
3338 }
3339 }
2e64971a 3340 trie->prefixlen = (state-1);
3dab1dad 3341 if (str) {
8e11feef 3342 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3343 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3344 trie->startstate = state;
07be1b83
YO
3345 trie->minlen -= (state - 1);
3346 trie->maxlen -= (state - 1);
33809eae
JH
3347#ifdef DEBUGGING
3348 /* At least the UNICOS C compiler choked on this
3349 * being argument to DEBUG_r(), so let's just have
3350 * it right here. */
3351 if (
3352#ifdef PERL_EXT_RE_BUILD
3353 1
3354#else
3355 DEBUG_r_TEST
3356#endif
3357 ) {
3358 regnode *fix = convert;
3359 U32 word = trie->wordcount;
3360 mjd_nodelen++;
3361 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3362 while( ++fix < n ) {
3363 Set_Node_Offset_Length(fix, 0, 0);
3364 }
3365 while (word--) {
3366 SV ** const tmp = av_fetch( trie_words, word, 0 );
3367 if (tmp) {
3368 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3369 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3370 else
3371 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3372 }
3373 }
3374 }
3375#endif
8e11feef
RGS
3376 if (trie->maxlen) {
3377 convert = n;
3378 } else {
3dab1dad 3379 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3380 DEBUG_r(optimize= n);
3dab1dad
YO
3381 }
3382 }
3383 }
538e84ed
KW
3384 if (!jumper)
3385 jumper = last;
3dab1dad 3386 if ( trie->maxlen ) {
8e11feef
RGS
3387 NEXT_OFF( convert ) = (U16)(tail - convert);
3388 ARG_SET( convert, data_slot );
538e84ed
KW
3389 /* Store the offset to the first unabsorbed branch in
3390 jump[0], which is otherwise unused by the jump logic.
786e8c11 3391 We use this when dumping a trie and during optimisation. */
538e84ed 3392 if (trie->jump)
7f69552c 3393 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3394
6c48061a
YO
3395 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3396 * and there is a bitmap
3397 * and the first "jump target" node we found leaves enough room
3398 * then convert the TRIE node into a TRIEC node, with the bitmap
3399 * embedded inline in the opcode - this is hypothetically faster.
3400 */
3401 if ( !trie->states[trie->startstate].wordnum
3402 && trie->bitmap
3403 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3404 {
3405 OP( convert ) = TRIEC;
3406 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3407 PerlMemShared_free(trie->bitmap);
786e8c11 3408 trie->bitmap= NULL;
538e84ed 3409 } else
786e8c11 3410 OP( convert ) = TRIE;
a3621e74 3411
3dab1dad
YO
3412 /* store the type in the flags */
3413 convert->flags = nodetype;
a5ca303d 3414 DEBUG_r({
538e84ed
KW
3415 optimize = convert
3416 + NODE_STEP_REGNODE
a5ca303d
YO
3417 + regarglen[ OP( convert ) ];
3418 });
538e84ed 3419 /* XXX We really should free up the resource in trie now,
a5ca303d 3420 as we won't use them - (which resources?) dmq */
3dab1dad 3421 }
a3621e74 3422 /* needed for dumping*/
e62cc96a 3423 DEBUG_r(if (optimize) {
07be1b83 3424 regnode *opt = convert;
bcdf7404 3425
e62cc96a 3426 while ( ++opt < optimize) {
07be1b83
YO
3427 Set_Node_Offset_Length(opt,0,0);
3428 }
538e84ed
KW
3429 /*
3430 Try to clean up some of the debris left after the
786e8c11 3431 optimisation.
a3621e74 3432 */
786e8c11 3433 while( optimize < jumper ) {
07be1b83 3434 mjd_nodelen += Node_Length((optimize));
a3621e74 3435 OP( optimize ) = OPTIMIZED;
07be1b83 3436 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3437 optimize++;
3438 }
07be1b83 3439 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3440 });
3441 } /* end node insert */
2e64971a
DM
3442
3443 /* Finish populating the prev field of the wordinfo array. Walk back
3444 * from each accept state until we find another accept state, and if
3445 * so, point the first word's .prev field at the second word. If the
3446 * second already has a .prev field set, stop now. This will be the
3447 * case either if we've already processed that word's accept state,
3b753521
FN
3448 * or that state had multiple words, and the overspill words were
3449 * already linked up earlier.
2e64971a
DM
3450 */
3451 {
3452 U16 word;
3453 U32 state;
3454 U16 prev;
3455
3456 for (word=1; word <= trie->wordcount; word++) {
3457 prev = 0;
3458 if (trie->wordinfo[word].prev)
3459 continue;
3460 state = trie->wordinfo[word].accept;
3461 while (state) {
3462 state = prev_states[state];
3463 if (!state)
3464 break;
3465 prev = trie->states[state].wordnum;
3466 if (prev)
3467 break;
3468 }
3469 trie->wordinfo[word].prev = prev;
3470 }
3471 Safefree(prev_states);
3472 }
3473
3474