This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Split function into two functions
[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
dea37815
KW
104#ifndef MIN
105#define MIN(a,b) ((a) < (b) ? (a) : (b))
106#endif
b1603ef8 107
3f910e62
YO
108/* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
111struct scan_frame;
112typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
118
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
122} scan_frame;
123
f4ae5a27
KW
124/* Certain characters are output as a sequence with the first being a
125 * backslash. */
126#define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
128
129
09b2b2e6 130struct RExC_state_t {
514a91f1
DM
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 133 char *precomp; /* uncompiled string. */
288b8c02 134 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 135 regexp *rx; /* perl core regexp structure */
538e84ed
KW
136 regexp_internal *rxi; /* internal data for regexp object
137 pprivate field */
fac92740 138 char *start; /* Start of input for compile */
830247a4
IZ
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
ea3daa5d 141 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 142 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
143 regnode *emit_bound; /* First regnode outside of the
144 allocated space */
f7c7e32a
DM
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
9a81a976
KW
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
150 scratch in pass1 */
830247a4
IZ
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
153 U32 seen;
ea3daa5d 154 SSize_t size; /* Code size. */
538e84ed
KW
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
157 pattern)*/
158 I32 nestroot; /* root parens we are in - used by
159 accept */
830247a4
IZ
160 I32 extralen;
161 I32 seen_zerolen;
40d049e4
YO
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
02daf0ab
YO
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
e40e74fe
KW
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
171 utf8 */
81714fb9 172 HV *paren_names; /* Paren names */
538e84ed 173
40d049e4
YO
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
4286711a 176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 177 through */
09a65838 178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 179 I32 in_lookbehind;
4624b182 180 I32 contains_locale;
cfafade5 181 I32 contains_i;
bb3f3ed2 182 I32 override_recoding;
b6d67071
KW
183#ifdef EBCDIC
184 I32 recode_x_to_native;
185#endif
9d53c457 186 I32 in_multi_char_class;
3d2bd50a 187 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 188 within pattern */
b1603ef8
DM
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
ee273784 191 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
192 scan_frame *frame_head;
193 scan_frame *frame_last;
194 U32 frame_count;
67cdf558 195 U32 strict;
dc6d7f5c 196#ifdef ADD_TO_REGEXEC
830247a4
IZ
197 char *starttry; /* -Dr: where regtry was called. */
198#define RExC_starttry (pRExC_state->starttry)
199#endif
d24ca0c5 200 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 201#ifdef DEBUGGING
be8e71aa 202 const char *lastparse;
3dab1dad 203 I32 lastnum;
1f1031fe 204 AV *paren_name_list; /* idx -> name */
d9a72fcc 205 U32 study_chunk_recursed_count;
c9f0d54c
YO
206 SV *mysv1;
207 SV *mysv2;
3dab1dad
YO
208#define RExC_lastparse (pRExC_state->lastparse)
209#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 210#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 211#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
212#define RExC_mysv (pRExC_state->mysv1)
213#define RExC_mysv1 (pRExC_state->mysv1)
214#define RExC_mysv2 (pRExC_state->mysv2)
215
3dab1dad 216#endif
512c0f5a 217 bool seen_unfolded_sharp_s;
09b2b2e6 218};
830247a4 219
e2509266 220#define RExC_flags (pRExC_state->flags)
514a91f1 221#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 222#define RExC_precomp (pRExC_state->precomp)
288b8c02 223#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 224#define RExC_rx (pRExC_state->rx)
f8fc2ecf 225#define RExC_rxi (pRExC_state->rxi)
fac92740 226#define RExC_start (pRExC_state->start)
830247a4
IZ
227#define RExC_end (pRExC_state->end)
228#define RExC_parse (pRExC_state->parse)
229#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
230
231/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
232 * EXACTF node, hence was parsed under /di rules. If later in the parse,
233 * something forces the pattern into using /ui rules, the sharp s should be
234 * folded into the sequence 'ss', which takes up more space than previously
235 * calculated. This means that the sizing pass needs to be restarted. (The
236 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
237 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
238 * so there is no need to resize [perl #125990]. */
239#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
240
7122b237 241#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
242#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
243 others */
7122b237 244#endif
830247a4 245#define RExC_emit (pRExC_state->emit)
f7c7e32a 246#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 247#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 248#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
249#define RExC_sawback (pRExC_state->sawback)
250#define RExC_seen (pRExC_state->seen)
251#define RExC_size (pRExC_state->size)
ee273784 252#define RExC_maxlen (pRExC_state->maxlen)
830247a4 253#define RExC_npar (pRExC_state->npar)
e2e6a0f1 254#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
255#define RExC_extralen (pRExC_state->extralen)
256#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 257#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 258#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 259#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
260#define RExC_open_parens (pRExC_state->open_parens)
261#define RExC_close_parens (pRExC_state->close_parens)
262#define RExC_opend (pRExC_state->opend)
81714fb9 263#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
264#define RExC_recurse (pRExC_state->recurse)
265#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 266#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
267#define RExC_study_chunk_recursed_bytes \
268 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 269#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 270#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 271#define RExC_contains_i (pRExC_state->contains_i)
9d53c457 272#define RExC_override_recoding (pRExC_state->override_recoding)
b6d67071
KW
273#ifdef EBCDIC
274# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275#endif
9d53c457 276#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
277#define RExC_frame_head (pRExC_state->frame_head)
278#define RExC_frame_last (pRExC_state->frame_last)
279#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 280#define RExC_strict (pRExC_state->strict)
830247a4 281
99807a43
HS
282/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
283 * a flag to disable back-off on the fixed/floating substrings - if it's
284 * a high complexity pattern we assume the benefit of avoiding a full match
285 * is worth the cost of checking for the substrings even if they rarely help.
286 */
287#define RExC_naughty (pRExC_state->naughty)
288#define TOO_NAUGHTY (10)
289#define MARK_NAUGHTY(add) \
290 if (RExC_naughty < TOO_NAUGHTY) \
291 RExC_naughty += (add)
292#define MARK_NAUGHTY_EXP(exp, add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 295
a687059c
LW
296#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
297#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 298 ((*s) == '{' && regcurly(s)))
a687059c
LW
299
300/*
301 * Flags to be passed up and down.
302 */
a687059c 303#define WORST 0 /* Worst case. */
a3b492c3 304#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 305
e64f369d 306/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
307 * character. (There needs to be a case: in the switch statement in regexec.c
308 * for any node marked SIMPLE.) Note that this is not the same thing as
309 * REGNODE_SIMPLE */
fda99bee 310#define SIMPLE 0x02
e64f369d 311#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
312#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
313#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
314#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
315#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
316 calcuate sizes as UTF-8 */
a687059c 317
3dab1dad
YO
318#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
319
07be1b83
YO
320/* whether trie related optimizations are enabled */
321#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
322#define TRIE_STUDY_OPT
786e8c11 323#define FULL_TRIE_STUDY
07be1b83
YO
324#define TRIE_STCLASS
325#endif
1de06328
YO
326
327
40d049e4
YO
328
329#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
330#define PBITVAL(paren) (1 << ((paren) & 7))
331#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
332#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
333#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
334
82a6ada4 335#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 336 if (!UTF) { \
82a6ada4
KW
337 assert(PASS1); \
338 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
339 return NULL; \
340 } \
82a6ada4 341 } STMT_END
40d049e4 342
512c0f5a
KW
343/* Change from /d into /u rules, and restart the parse if we've already seen
344 * something whose size would increase as a result, by setting *flagp and
345 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
346 * we've change to /u during the parse. */
347#define REQUIRE_UNI_RULES(flagp, restart_retval) \
348 STMT_START { \
349 if (DEPENDS_SEMANTICS) { \
350 assert(PASS1); \
351 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
352 RExC_uni_semantics = 1; \
353 if (RExC_seen_unfolded_sharp_s) { \
354 *flagp |= RESTART_PASS1; \
355 return restart_retval; \
356 } \
357 } \
358 } STMT_END
359
f19b1a63
KW
360/* This converts the named class defined in regcomp.h to its equivalent class
361 * number defined in handy.h. */
362#define namedclass_to_classnum(class) ((int) ((class) / 2))
363#define classnum_to_namedclass(classnum) ((classnum) * 2)
364
de92f5e6
KW
365#define _invlist_union_complement_2nd(a, b, output) \
366 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
367#define _invlist_intersection_complement_2nd(a, b, output) \
368 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
369
1de06328
YO
370/* About scan_data_t.
371
372 During optimisation we recurse through the regexp program performing
373 various inplace (keyhole style) optimisations. In addition study_chunk
374 and scan_commit populate this data structure with information about
538e84ed 375 what strings MUST appear in the pattern. We look for the longest
3b753521 376 string that must appear at a fixed location, and we look for the
1de06328
YO
377 longest string that may appear at a floating location. So for instance
378 in the pattern:
538e84ed 379
1de06328 380 /FOO[xX]A.*B[xX]BAR/
538e84ed 381
1de06328
YO
382 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
383 strings (because they follow a .* construct). study_chunk will identify
384 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 385
1de06328 386 The strings can be composites, for instance
538e84ed 387
1de06328 388 /(f)(o)(o)/
538e84ed 389
1de06328 390 will result in a composite fixed substring 'foo'.
538e84ed 391
1de06328 392 For each string some basic information is maintained:
538e84ed 393
1de06328
YO
394 - offset or min_offset
395 This is the position the string must appear at, or not before.
396 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
397 characters must match before the string we are searching for.
398 Likewise when combined with minlenp and the length of the string it
538e84ed 399 tells us how many characters must appear after the string we have
1de06328 400 found.
538e84ed 401
1de06328
YO
402 - max_offset
403 Only used for floating strings. This is the rightmost point that
ea3daa5d 404 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 405 string can occur infinitely far to the right.
538e84ed 406
1de06328 407 - minlenp
2d608413
KW
408 A pointer to the minimum number of characters of the pattern that the
409 string was found inside. This is important as in the case of positive
538e84ed 410 lookahead or positive lookbehind we can have multiple patterns
1de06328 411 involved. Consider
538e84ed 412
1de06328 413 /(?=FOO).*F/
538e84ed 414
1de06328
YO
415 The minimum length of the pattern overall is 3, the minimum length
416 of the lookahead part is 3, but the minimum length of the part that
538e84ed 417 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 418 minimum length for the F is 1. This is important as the minimum length
538e84ed 419 is used to determine offsets in front of and behind the string being
1de06328 420 looked for. Since strings can be composites this is the length of the
486ec47a 421 pattern at the time it was committed with a scan_commit. Note that
1de06328 422 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 423 are not known until the full pattern has been compiled, thus the
1de06328 424 pointer to the value.
538e84ed 425
1de06328 426 - lookbehind
538e84ed 427
1de06328 428 In the case of lookbehind the string being searched for can be
538e84ed 429 offset past the start point of the final matching string.
1de06328
YO
430 If this value was just blithely removed from the min_offset it would
431 invalidate some of the calculations for how many chars must match
432 before or after (as they are derived from min_offset and minlen and
538e84ed 433 the length of the string being searched for).
1de06328
YO
434 When the final pattern is compiled and the data is moved from the
435 scan_data_t structure into the regexp structure the information
538e84ed
KW
436 about lookbehind is factored in, with the information that would
437 have been lost precalculated in the end_shift field for the
1de06328
YO
438 associated string.
439
440 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 441 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
442
443*/
2c2d71f5
JH
444
445typedef struct scan_data_t {
1de06328
YO
446 /*I32 len_min; unused */
447 /*I32 len_delta; unused */
49f55535 448 SSize_t pos_min;
ea3daa5d 449 SSize_t pos_delta;
2c2d71f5 450 SV *last_found;
ea3daa5d 451 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 452 SSize_t last_start_min;
ea3daa5d 453 SSize_t last_start_max;
1de06328
YO
454 SV **longest; /* Either &l_fixed, or &l_float. */
455 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 456 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 457 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
458 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
459 SV *longest_float; /* longest floating string found in pattern */
49f55535 460 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
461 SSize_t offset_float_max; /* latest point in string it can appear */
462 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 463 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
464 I32 flags;
465 I32 whilem_c;
49f55535 466 SSize_t *last_closep;
b8f7bb16 467 regnode_ssc *start_class;
2c2d71f5
JH
468} scan_data_t;
469
a687059c 470/*
e50aee73 471 * Forward declarations for pregcomp()'s friends.
a687059c 472 */
a0d0e21e 473
27da23d5 474static const scan_data_t zero_scan_data =
1de06328 475 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
476
477#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
478#define SF_BEFORE_SEOL 0x0001
479#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
480#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
481#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
482
44e3dfd2
BF
483#define SF_FIX_SHIFT_EOL (+2)
484#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
485
486#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
487#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
488
489#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
490#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
491#define SF_IS_INF 0x0040
492#define SF_HAS_PAR 0x0080
493#define SF_IN_PAR 0x0100
494#define SF_HAS_EVAL 0x0200
495#define SCF_DO_SUBSTR 0x0400
653099ff
GS
496#define SCF_DO_STCLASS_AND 0x0800
497#define SCF_DO_STCLASS_OR 0x1000
498#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 499#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 500
786e8c11 501#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 502#define SCF_SEEN_ACCEPT 0x8000
688e0391 503#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
504#define SCF_IN_DEFINE 0x20000
505
506
507
07be1b83 508
43fead97 509#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
510
511/* The enums for all these are ordered so things work out correctly */
a62b1201 512#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
513#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
514 == REGEX_DEPENDS_CHARSET)
00b27cfc 515#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
516#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
517 >= REGEX_UNICODE_CHARSET)
518#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
519 == REGEX_ASCII_RESTRICTED_CHARSET)
520#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
521 >= REGEX_ASCII_RESTRICTED_CHARSET)
522#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
523 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 524
43fead97 525#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 526
f2c2a6ab
KW
527/* For programs that want to be strictly Unicode compatible by dying if any
528 * attempt is made to match a non-Unicode code point against a Unicode
529 * property. */
530#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
531
93733859 532#define OOB_NAMEDCLASS -1
b8c5462f 533
8e661ac5
KW
534/* There is no code point that is out-of-bounds, so this is problematic. But
535 * its only current use is to initialize a variable that is always set before
536 * looked at. */
537#define OOB_UNICODE 0xDEADBEEF
538
a0ed51b3
LW
539#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
540#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
541
8615cb43 542
b45f050a
JF
543/* length of regex to show in messages that don't mark a position within */
544#define RegexLengthToShowInErrorMessages 127
545
546/*
547 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
548 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
549 * op/pragma/warn/regcomp.
550 */
7253e4e3
RK
551#define MARKER1 "<-- HERE" /* marker as it appears in the description */
552#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 553
538e84ed
KW
554#define REPORT_LOCATION " in regex; marked by " MARKER1 \
555 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 556
c1d900c3
BF
557#define REPORT_LOCATION_ARGS(offset) \
558 UTF8fARG(UTF, offset, RExC_precomp), \
559 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
560
8a6d8ec6
HS
561/* Used to point after bad bytes for an error message, but avoid skipping
562 * past a nul byte. */
563#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
564
b45f050a
JF
565/*
566 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
567 * arg. Show regex, up to a maximum length. If it's too long, chop and add
568 * "...".
569 */
58e23c8d 570#define _FAIL(code) STMT_START { \
bfed75c6 571 const char *ellipses = ""; \
ccb2c380
MP
572 IV len = RExC_end - RExC_precomp; \
573 \
574 if (!SIZE_ONLY) \
a5e7bc51 575 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
576 if (len > RegexLengthToShowInErrorMessages) { \
577 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
578 len = RegexLengthToShowInErrorMessages - 10; \
579 ellipses = "..."; \
580 } \
58e23c8d 581 code; \
ccb2c380 582} STMT_END
8615cb43 583
58e23c8d 584#define FAIL(msg) _FAIL( \
c1d900c3
BF
585 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
586 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
587
588#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
589 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
590 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 591
b45f050a 592/*
b45f050a
JF
593 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
594 */
ccb2c380 595#define Simple_vFAIL(m) STMT_START { \
8fb0127d
YO
596 const IV offset = \
597 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
ccb2c380 598 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
c1d900c3 599 m, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 600} STMT_END
b45f050a
JF
601
602/*
603 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
604 */
ccb2c380
MP
605#define vFAIL(m) STMT_START { \
606 if (!SIZE_ONLY) \
a5e7bc51 607 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
608 Simple_vFAIL(m); \
609} STMT_END
b45f050a
JF
610
611/*
612 * Like Simple_vFAIL(), but accepts two arguments.
613 */
ccb2c380 614#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 615 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
616 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
617 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 618} STMT_END
b45f050a
JF
619
620/*
621 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
622 */
ccb2c380
MP
623#define vFAIL2(m,a1) STMT_START { \
624 if (!SIZE_ONLY) \
a5e7bc51 625 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
626 Simple_vFAIL2(m, a1); \
627} STMT_END
b45f050a
JF
628
629
630/*
631 * Like Simple_vFAIL(), but accepts three arguments.
632 */
ccb2c380 633#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 634 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
635 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
636 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 637} STMT_END
b45f050a
JF
638
639/*
640 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
641 */
ccb2c380
MP
642#define vFAIL3(m,a1,a2) STMT_START { \
643 if (!SIZE_ONLY) \
a5e7bc51 644 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
645 Simple_vFAIL3(m, a1, a2); \
646} STMT_END
b45f050a
JF
647
648/*
649 * Like Simple_vFAIL(), but accepts four arguments.
650 */
ccb2c380 651#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 652 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
653 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
654 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 655} STMT_END
b45f050a 656
95db3ffa
KW
657#define vFAIL4(m,a1,a2,a3) STMT_START { \
658 if (!SIZE_ONLY) \
659 SAVEFREESV(RExC_rx_sv); \
660 Simple_vFAIL4(m, a1, a2, a3); \
661} STMT_END
662
946095af
BF
663/* A specialized version of vFAIL2 that works with UTF8f */
664#define vFAIL2utf8f(m, a1) STMT_START { \
ef3f731d
BF
665 const IV offset = RExC_parse - RExC_precomp; \
666 if (!SIZE_ONLY) \
667 SAVEFREESV(RExC_rx_sv); \
668 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
669 REPORT_LOCATION_ARGS(offset)); \
946095af
BF
670} STMT_END
671
499333dc
KW
672/* These have asserts in them because of [perl #122671] Many warnings in
673 * regcomp.c can occur twice. If they get output in pass1 and later in that
674 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
675 * would get output again. So they should be output in pass2, and these
676 * asserts make sure new warnings follow that paradigm. */
946095af 677
5e0a247b
KW
678/* m is not necessarily a "literal string", in this macro */
679#define reg_warn_non_literal_string(loc, m) STMT_START { \
680 const IV offset = loc - RExC_precomp; \
499333dc 681 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
c1d900c3 682 m, REPORT_LOCATION_ARGS(offset)); \
5e0a247b
KW
683} STMT_END
684
668c081a 685#define ckWARNreg(loc,m) STMT_START { \
a28509cc 686 const IV offset = loc - RExC_precomp; \
499333dc 687 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 688 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
689} STMT_END
690
b927b7e9
KW
691#define vWARN(loc, m) STMT_START { \
692 const IV offset = loc - RExC_precomp; \
693 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
694 REPORT_LOCATION_ARGS(offset)); \
695} STMT_END
696
0d6106aa
KW
697#define vWARN_dep(loc, m) STMT_START { \
698 const IV offset = loc - RExC_precomp; \
499333dc 699 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
c1d900c3 700 REPORT_LOCATION_ARGS(offset)); \
0d6106aa
KW
701} STMT_END
702
147508a2
KW
703#define ckWARNdep(loc,m) STMT_START { \
704 const IV offset = loc - RExC_precomp; \
499333dc 705 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
147508a2 706 m REPORT_LOCATION, \
c1d900c3 707 REPORT_LOCATION_ARGS(offset)); \
147508a2
KW
708} STMT_END
709
668c081a 710#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 711 const IV offset = loc - RExC_precomp; \
499333dc 712 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18 713 m REPORT_LOCATION, \
c1d900c3 714 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
715} STMT_END
716
b23eb183 717#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 718 const IV offset = loc - RExC_precomp; \
499333dc 719 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3 720 m REPORT_LOCATION, \
c1d900c3 721 a1, REPORT_LOCATION_ARGS(offset)); \
2335b3d3
KW
722} STMT_END
723
668c081a 724#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 725 const IV offset = loc - RExC_precomp; \
499333dc 726 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 727 a1, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
728} STMT_END
729
730#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 731 const IV offset = loc - RExC_precomp; \
499333dc 732 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 733 a1, a2, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
734} STMT_END
735
668c081a
NC
736#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
737 const IV offset = loc - RExC_precomp; \
499333dc 738 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 739 a1, a2, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
740} STMT_END
741
ccb2c380 742#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 743 const IV offset = loc - RExC_precomp; \
499333dc 744 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 745 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
746} STMT_END
747
668c081a
NC
748#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
749 const IV offset = loc - RExC_precomp; \
499333dc 750 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 751 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
752} STMT_END
753
ccb2c380 754#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 755 const IV offset = loc - RExC_precomp; \
499333dc 756 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 757 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 758} STMT_END
9d1d55b5 759
538e84ed 760/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
761 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
762 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
763 * Element 0 holds the number n.
07be1b83 764 * Position is 1 indexed.
fac92740 765 */
7122b237
YO
766#ifndef RE_TRACK_PATTERN_OFFSETS
767#define Set_Node_Offset_To_R(node,byte)
768#define Set_Node_Offset(node,byte)
769#define Set_Cur_Node_Offset
770#define Set_Node_Length_To_R(node,len)
771#define Set_Node_Length(node,len)
6a86c6ad 772#define Set_Node_Cur_Length(node,start)
538e84ed
KW
773#define Node_Offset(n)
774#define Node_Length(n)
7122b237
YO
775#define Set_Node_Offset_Length(node,offset,len)
776#define ProgLen(ri) ri->u.proglen
777#define SetProgLen(ri,x) ri->u.proglen = x
778#else
779#define ProgLen(ri) ri->u.offsets[0]
780#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
781#define Set_Node_Offset_To_R(node,byte) STMT_START { \
782 if (! SIZE_ONLY) { \
783 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 784 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 785 if((node) < 0) { \
538e84ed
KW
786 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
787 (int)(node)); \
ccb2c380
MP
788 } else { \
789 RExC_offsets[2*(node)-1] = (byte); \
790 } \
791 } \
792} STMT_END
793
794#define Set_Node_Offset(node,byte) \
795 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
796#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
797
798#define Set_Node_Length_To_R(node,len) STMT_START { \
799 if (! SIZE_ONLY) { \
800 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 801 __LINE__, (int)(node), (int)(len))); \
ccb2c380 802 if((node) < 0) { \
538e84ed
KW
803 Perl_croak(aTHX_ "value of node is %d in Length macro", \
804 (int)(node)); \
ccb2c380
MP
805 } else { \
806 RExC_offsets[2*(node)] = (len); \
807 } \
808 } \
809} STMT_END
810
811#define Set_Node_Length(node,len) \
812 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
813#define Set_Node_Cur_Length(node, start) \
814 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
815
816/* Get offsets and lengths */
817#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
818#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
819
07be1b83
YO
820#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
821 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
822 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
823} STMT_END
7122b237 824#endif
07be1b83
YO
825
826#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
827#define EXPERIMENTAL_INPLACESCAN
f427392e 828#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 829
9e9ecfdf 830#define DEBUG_RExC_seen() \
538e84ed
KW
831 DEBUG_OPTIMISE_MORE_r({ \
832 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
833 \
e384d5c1
YO
834 if (RExC_seen & REG_ZERO_LEN_SEEN) \
835 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
538e84ed 836 \
e384d5c1
YO
837 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
838 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
538e84ed 839 \
e384d5c1
YO
840 if (RExC_seen & REG_GPOS_SEEN) \
841 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
538e84ed 842 \
e384d5c1
YO
843 if (RExC_seen & REG_RECURSE_SEEN) \
844 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
538e84ed 845 \
e384d5c1
YO
846 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
847 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 848 \
e384d5c1
YO
849 if (RExC_seen & REG_VERBARG_SEEN) \
850 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
538e84ed 851 \
e384d5c1
YO
852 if (RExC_seen & REG_CUTGROUP_SEEN) \
853 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
538e84ed 854 \
e384d5c1
YO
855 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
856 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 857 \
e384d5c1
YO
858 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
859 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 860 \
e384d5c1
YO
861 if (RExC_seen & REG_GOSTART_SEEN) \
862 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
538e84ed 863 \
ee273784
YO
864 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
865 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
866 \
538e84ed 867 PerlIO_printf(Perl_debug_log,"\n"); \
9e9ecfdf
YO
868 });
869
fdfb4f21
YO
870#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
871 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
872
873#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
874 if ( ( flags ) ) { \
875 PerlIO_printf(Perl_debug_log, "%s", open_str); \
876 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
877 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
878 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
879 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
880 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
881 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
882 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
883 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
884 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
885 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
886 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
887 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
888 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
889 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
890 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
891 PerlIO_printf(Perl_debug_log, "%s", close_str); \
892 }
893
894
304ee84b
YO
895#define DEBUG_STUDYDATA(str,data,depth) \
896DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 897 PerlIO_printf(Perl_debug_log, \
304ee84b 898 "%*s" str "Pos:%"IVdf"/%"IVdf \
fdfb4f21 899 " Flags: 0x%"UVXf, \
1de06328
YO
900 (int)(depth)*2, "", \
901 (IV)((data)->pos_min), \
902 (IV)((data)->pos_delta), \
fdfb4f21
YO
903 (UV)((data)->flags) \
904 ); \
905 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
906 PerlIO_printf(Perl_debug_log, \
907 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328 908 (IV)((data)->whilem_c), \
304ee84b
YO
909 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
910 is_inf ? "INF " : "" \
1de06328
YO
911 ); \
912 if ((data)->last_found) \
913 PerlIO_printf(Perl_debug_log, \
914 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
915 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
916 SvPVX_const((data)->last_found), \
917 (IV)((data)->last_end), \
918 (IV)((data)->last_start_min), \
919 (IV)((data)->last_start_max), \
920 ((data)->longest && \
921 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
922 SvPVX_const((data)->longest_fixed), \
923 (IV)((data)->offset_fixed), \
924 ((data)->longest && \
925 (data)->longest==&((data)->longest_float)) ? "*" : "", \
926 SvPVX_const((data)->longest_float), \
927 (IV)((data)->offset_float_min), \
928 (IV)((data)->offset_float_max) \
929 ); \
930 PerlIO_printf(Perl_debug_log,"\n"); \
931});
932
8e35b056
KW
933/* is c a control character for which we have a mnemonic? */
934#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
935
549b4e78
KW
936STATIC const char *
937S_cntrl_to_mnemonic(const U8 c)
938{
939 /* Returns the mnemonic string that represents character 'c', if one
940 * exists; NULL otherwise. The only ones that exist for the purposes of
941 * this routine are a few control characters */
942
943 switch (c) {
944 case '\a': return "\\a";
945 case '\b': return "\\b";
946 case ESC_NATIVE: return "\\e";
947 case '\f': return "\\f";
948 case '\n': return "\\n";
949 case '\r': return "\\r";
950 case '\t': return "\\t";
951 }
952
953 return NULL;
954}
955
653099ff 956/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 957 Update the longest found anchored substring and the longest found
653099ff
GS
958 floating substrings if needed. */
959
4327152a 960STATIC void
ea3daa5d
FC
961S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
962 SSize_t *minlenp, int is_inf)
c277df42 963{
e1ec3a88
AL
964 const STRLEN l = CHR_SVLEN(data->last_found);
965 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 966 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 967
7918f24d
NC
968 PERL_ARGS_ASSERT_SCAN_COMMIT;
969
c277df42 970 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 971 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
972 if (*data->longest == data->longest_fixed) {
973 data->offset_fixed = l ? data->last_start_min : data->pos_min;
974 if (data->flags & SF_BEFORE_EOL)
b81d288d 975 data->flags
c277df42
IZ
976 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
977 else
978 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 979 data->minlen_fixed=minlenp;
1de06328 980 data->lookbehind_fixed=0;
a0ed51b3 981 }
304ee84b 982 else { /* *data->longest == data->longest_float */
c277df42 983 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 984 data->offset_float_max = (l
646e8787
DM
985 ? data->last_start_max
986 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
987 ? SSize_t_MAX
988 : data->pos_min + data->pos_delta));
989 if (is_inf
990 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
991 data->offset_float_max = SSize_t_MAX;
c277df42 992 if (data->flags & SF_BEFORE_EOL)
b81d288d 993 data->flags
c277df42
IZ
994 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
995 else
996 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
997 data->minlen_float=minlenp;
998 data->lookbehind_float=0;
c277df42
IZ
999 }
1000 }
1001 SvCUR_set(data->last_found, 0);
0eda9292 1002 {
a28509cc 1003 SV * const sv = data->last_found;
097eb12c
AL
1004 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1005 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1006 if (mg)
1007 mg->mg_len = 0;
1008 }
0eda9292 1009 }
c277df42
IZ
1010 data->last_end = -1;
1011 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 1012 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
1013}
1014
cdd87c1d
KW
1015/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1016 * list that describes which code points it matches */
1017
653099ff 1018STATIC void
3420edd7 1019S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1020{
cdd87c1d
KW
1021 /* Set the SSC 'ssc' to match an empty string or any code point */
1022
557bd3fb 1023 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1024
71068078 1025 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1026
cdd87c1d
KW
1027 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1028 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
93e92956 1029 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1030}
1031
653099ff 1032STATIC int
dc3bf405 1033S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1034{
c144baaa
KW
1035 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1036 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1037 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1038 * in any way, so there's no point in using it */
cdd87c1d
KW
1039
1040 UV start, end;
1041 bool ret;
653099ff 1042
557bd3fb 1043 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1044
71068078 1045 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1046
93e92956 1047 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1048 return FALSE;
1049 }
1050
1051 /* See if the list consists solely of the range 0 - Infinity */
1052 invlist_iterinit(ssc->invlist);
1053 ret = invlist_iternext(ssc->invlist, &start, &end)
1054 && start == 0
1055 && end == UV_MAX;
1056
1057 invlist_iterfinish(ssc->invlist);
1058
1059 if (ret) {
1060 return TRUE;
1061 }
1062
1063 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1064 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1065 int i;
1066 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1067 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1068 return TRUE;
1069 }
1070 }
1071 }
1072
1073 return FALSE;
653099ff
GS
1074}
1075
653099ff 1076STATIC void
cdd87c1d 1077S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1078{
cdd87c1d
KW
1079 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1080 * string, any code point, or any posix class under locale */
1081
557bd3fb 1082 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1083
557bd3fb 1084 Zero(ssc, 1, regnode_ssc);
71068078 1085 set_ANYOF_SYNTHETIC(ssc);
93e92956 1086 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1087 ssc_anything(ssc);
cdd87c1d 1088
2f306ab9
KW
1089 /* If any portion of the regex is to operate under locale rules that aren't
1090 * fully known at compile time, initialization includes it. The reason
1091 * this isn't done for all regexes is that the optimizer was written under
1092 * the assumption that locale was all-or-nothing. Given the complexity and
1093 * lack of documentation in the optimizer, and that there are inadequate
1094 * test cases for locale, many parts of it may not work properly, it is
1095 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1096 if (RExC_contains_locale) {
1097 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1098 }
1099 else {
1100 ANYOF_POSIXL_ZERO(ssc);
1101 }
653099ff
GS
1102}
1103
b423522f 1104STATIC int
dc3bf405
BF
1105S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1106 const regnode_ssc *ssc)
b423522f
KW
1107{
1108 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1109 * to the list of code points matched, and locale posix classes; hence does
1110 * not check its flags) */
1111
1112 UV start, end;
1113 bool ret;
1114
1115 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1116
71068078 1117 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1118
1119 invlist_iterinit(ssc->invlist);
1120 ret = invlist_iternext(ssc->invlist, &start, &end)
1121 && start == 0
1122 && end == UV_MAX;
1123
1124 invlist_iterfinish(ssc->invlist);
1125
1126 if (! ret) {
1127 return FALSE;
1128 }
1129
e0e1be5f 1130 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1131 return FALSE;
b423522f
KW
1132 }
1133
1134 return TRUE;
1135}
1136
1137STATIC SV*
1138S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1139 const regnode_charclass* const node)
b423522f
KW
1140{
1141 /* Returns a mortal inversion list defining which code points are matched
1142 * by 'node', which is of type ANYOF. Handles complementing the result if
1143 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1144 * returned list must, and will, contain every code point that is a
1145 * possibility. */
b423522f
KW
1146
1147 SV* invlist = sv_2mortal(_new_invlist(0));
1ee208c4 1148 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1149 unsigned int i;
1150 const U32 n = ARG(node);
31f05a37 1151 bool new_node_has_latin1 = FALSE;
b423522f
KW
1152
1153 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1154
1155 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1156 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1157 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1158 AV * const av = MUTABLE_AV(SvRV(rv));
1159 SV **const ary = AvARRAY(av);
1160 assert(RExC_rxi->data->what[n] == 's');
1161
1162 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1163 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1164 }
1165 else if (ary[0] && ary[0] != &PL_sv_undef) {
1166
1167 /* Here, no compile-time swash, and there are things that won't be
1168 * known until runtime -- we have to assume it could be anything */
1169 return _add_range_to_invlist(invlist, 0, UV_MAX);
1170 }
1ee208c4 1171 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1172
1173 /* Here no compile-time swash, and no run-time only data. Use the
1174 * node's inversion list */
1ee208c4
KW
1175 invlist = sv_2mortal(invlist_clone(ary[3]));
1176 }
1177
1178 /* Get the code points valid only under UTF-8 locales */
1179 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1180 && ary[2] && ary[2] != &PL_sv_undef)
1181 {
1182 only_utf8_locale_invlist = ary[2];
b423522f
KW
1183 }
1184 }
1185
dcb20b36
KW
1186 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1187 * code points, and an inversion list for the others, but if there are code
1188 * points that should match only conditionally on the target string being
1189 * UTF-8, those are placed in the inversion list, and not the bitmap.
1190 * Since there are circumstances under which they could match, they are
1191 * included in the SSC. But if the ANYOF node is to be inverted, we have
1192 * to exclude them here, so that when we invert below, the end result
1193 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1194 * have to do this here before we add the unconditionally matched code
1195 * points */
b423522f
KW
1196 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1197 _invlist_intersection_complement_2nd(invlist,
1198 PL_UpperLatin1,
1199 &invlist);
1200 }
1201
1202 /* Add in the points from the bit map */
dcb20b36 1203 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f
KW
1204 if (ANYOF_BITMAP_TEST(node, i)) {
1205 invlist = add_cp_to_invlist(invlist, i);
31f05a37 1206 new_node_has_latin1 = TRUE;
b423522f
KW
1207 }
1208 }
1209
1210 /* If this can match all upper Latin1 code points, have to add them
1211 * as well */
f240c685
KW
1212 if (OP(node) == ANYOFD
1213 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1214 {
b423522f
KW
1215 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1216 }
1217
1218 /* Similarly for these */
93e92956 1219 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1220 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1221 }
1222
1223 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1224 _invlist_invert(invlist);
1225 }
31f05a37
KW
1226 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1227
1228 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1229 * locale. We can skip this if there are no 0-255 at all. */
1230 _invlist_union(invlist, PL_Latin1, &invlist);
1231 }
1232
1ee208c4
KW
1233 /* Similarly add the UTF-8 locale possible matches. These have to be
1234 * deferred until after the non-UTF-8 locale ones are taken care of just
1235 * above, or it leads to wrong results under ANYOF_INVERT */
1236 if (only_utf8_locale_invlist) {
31f05a37 1237 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1238 only_utf8_locale_invlist,
31f05a37
KW
1239 ANYOF_FLAGS(node) & ANYOF_INVERT,
1240 &invlist);
1241 }
b423522f
KW
1242
1243 return invlist;
1244}
1245
1051e1c4 1246/* These two functions currently do the exact same thing */
557bd3fb 1247#define ssc_init_zero ssc_init
653099ff 1248
cdd87c1d
KW
1249#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1250#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1251
557bd3fb 1252/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1253 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1254 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1255
653099ff 1256STATIC void
b423522f 1257S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1258 const regnode_charclass *and_with)
653099ff 1259{
cdd87c1d
KW
1260 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1261 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1262
a0dd4231
KW
1263 SV* anded_cp_list;
1264 U8 anded_flags;
1e6ade67 1265
cdd87c1d 1266 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1267
71068078 1268 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1269
1270 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1271 * the code point inversion list and just the relevant flags */
71068078 1272 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1273 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1274 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1275
1276 /* XXX This is a kludge around what appears to be deficiencies in the
1277 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1278 * there are paths through the optimizer where it doesn't get weeded
1279 * out when it should. And if we don't make some extra provision for
1280 * it like the code just below, it doesn't get added when it should.
1281 * This solution is to add it only when AND'ing, which is here, and
1282 * only when what is being AND'ed is the pristine, original node
1283 * matching anything. Thus it is like adding it to ssc_anything() but
1284 * only when the result is to be AND'ed. Probably the same solution
1285 * could be adopted for the same problem we have with /l matching,
1286 * which is solved differently in S_ssc_init(), and that would lead to
1287 * fewer false positives than that solution has. But if this solution
1288 * creates bugs, the consequences are only that a warning isn't raised
1289 * that should be; while the consequences for having /l bugs is
1290 * incorrect matches */
7dcac5f6 1291 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1292 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1293 }
a0dd4231
KW
1294 }
1295 else {
5c0f85ef 1296 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1297 if (OP(and_with) == ANYOFD) {
1298 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1299 }
1300 else {
1301 anded_flags = ANYOF_FLAGS(and_with)
1302 &( ANYOF_COMMON_FLAGS
1303 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER);
1304 }
a0dd4231
KW
1305 }
1306
1307 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1308
1309 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1310 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1311 * 'and_with' may be inverted. When not inverted, we have the situation of
1312 * computing:
1313 * (C1 | P1) & (C2 | P2)
1314 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1315 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1316 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1317 * <= ((C1 & C2) | P1 | P2)
1318 * Alternatively, the last few steps could be:
1319 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1320 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1321 * <= (C1 | C2 | (P1 & P2))
1322 * We favor the second approach if either P1 or P2 is non-empty. This is
1323 * because these components are a barrier to doing optimizations, as what
1324 * they match cannot be known until the moment of matching as they are
1325 * dependent on the current locale, 'AND"ing them likely will reduce or
1326 * eliminate them.
1327 * But we can do better if we know that C1,P1 are in their initial state (a
1328 * frequent occurrence), each matching everything:
1329 * (<everything>) & (C2 | P2) = C2 | P2
1330 * Similarly, if C2,P2 are in their initial state (again a frequent
1331 * occurrence), the result is a no-op
1332 * (C1 | P1) & (<everything>) = C1 | P1
1333 *
1334 * Inverted, we have
1335 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1336 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1337 * <= (C1 & ~C2) | (P1 & ~P2)
1338 * */
1aa99e6b 1339
a0dd4231 1340 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1341 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1342 {
cdd87c1d 1343 unsigned int i;
8951c461 1344
cdd87c1d
KW
1345 ssc_intersection(ssc,
1346 anded_cp_list,
1347 FALSE /* Has already been inverted */
1348 );
c6b76537 1349
cdd87c1d
KW
1350 /* If either P1 or P2 is empty, the intersection will be also; can skip
1351 * the loop */
93e92956 1352 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1353 ANYOF_POSIXL_ZERO(ssc);
1354 }
e0e1be5f 1355 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1356
1357 /* Note that the Posix class component P from 'and_with' actually
1358 * looks like:
1359 * P = Pa | Pb | ... | Pn
1360 * where each component is one posix class, such as in [\w\s].
1361 * Thus
1362 * ~P = ~(Pa | Pb | ... | Pn)
1363 * = ~Pa & ~Pb & ... & ~Pn
1364 * <= ~Pa | ~Pb | ... | ~Pn
1365 * The last is something we can easily calculate, but unfortunately
1366 * is likely to have many false positives. We could do better
1367 * in some (but certainly not all) instances if two classes in
1368 * P have known relationships. For example
1369 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1370 * So
1371 * :lower: & :print: = :lower:
1372 * And similarly for classes that must be disjoint. For example,
1373 * since \s and \w can have no elements in common based on rules in
1374 * the POSIX standard,
1375 * \w & ^\S = nothing
1376 * Unfortunately, some vendor locales do not meet the Posix
1377 * standard, in particular almost everything by Microsoft.
1378 * The loop below just changes e.g., \w into \W and vice versa */
1379
1ee208c4 1380 regnode_charclass_posixl temp;
cdd87c1d
KW
1381 int add = 1; /* To calculate the index of the complement */
1382
1383 ANYOF_POSIXL_ZERO(&temp);
1384 for (i = 0; i < ANYOF_MAX; i++) {
1385 assert(i % 2 != 0
7dcac5f6
KW
1386 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1387 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1388
7dcac5f6 1389 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1390 ANYOF_POSIXL_SET(&temp, i + add);
1391 }
1392 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1393 }
1394 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1395
cdd87c1d
KW
1396 } /* else ssc already has no posixes */
1397 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1398 in its initial state */
71068078 1399 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1400 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1401 {
1402 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1403 * copy it over 'ssc' */
1404 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1405 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1406 StructCopy(and_with, ssc, regnode_ssc);
1407 }
1408 else {
1409 ssc->invlist = anded_cp_list;
1410 ANYOF_POSIXL_ZERO(ssc);
93e92956 1411 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1412 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1413 }
1414 }
1415 }
e0e1be5f 1416 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1417 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1418 {
1419 /* One or the other of P1, P2 is non-empty. */
93e92956 1420 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1421 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1422 }
cdd87c1d
KW
1423 ssc_union(ssc, anded_cp_list, FALSE);
1424 }
1425 else { /* P1 = P2 = empty */
1426 ssc_intersection(ssc, anded_cp_list, FALSE);
1427 }
137165a6 1428 }
653099ff
GS
1429}
1430
653099ff 1431STATIC void
cdd87c1d 1432S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1433 const regnode_charclass *or_with)
653099ff 1434{
cdd87c1d
KW
1435 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1436 * another SSC or a regular ANYOF class. Can create false positives if
1437 * 'or_with' is to be inverted. */
7918f24d 1438
a0dd4231
KW
1439 SV* ored_cp_list;
1440 U8 ored_flags;
c6b76537 1441
cdd87c1d 1442 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1443
71068078 1444 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1445
1446 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1447 * the code point inversion list and just the relevant flags */
71068078 1448 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1449 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1450 ored_flags = ANYOF_FLAGS(or_with);
1451 }
1452 else {
5c0f85ef 1453 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1454 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1455 if (OP(or_with) != ANYOFD) {
1456 ored_flags
1457 |= ANYOF_FLAGS(or_with)
1458 & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1459 }
a0dd4231
KW
1460 }
1461
1462 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1463
1464 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1465 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1466 * 'or_with' may be inverted. When not inverted, we have the simple
1467 * situation of computing:
1468 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1469 * If P1|P2 yields a situation with both a class and its complement are
1470 * set, like having both \w and \W, this matches all code points, and we
1471 * can delete these from the P component of the ssc going forward. XXX We
1472 * might be able to delete all the P components, but I (khw) am not certain
1473 * about this, and it is better to be safe.
1474 *
1475 * Inverted, we have
1476 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1477 * <= (C1 | P1) | ~C2
1478 * <= (C1 | ~C2) | P1
1479 * (which results in actually simpler code than the non-inverted case)
1480 * */
9826f543 1481
a0dd4231 1482 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1483 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1484 {
cdd87c1d 1485 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1486 } /* else Not inverted */
93e92956 1487 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1488 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1489 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1490 unsigned int i;
1491 for (i = 0; i < ANYOF_MAX; i += 2) {
1492 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1493 {
1494 ssc_match_all_cp(ssc);
1495 ANYOF_POSIXL_CLEAR(ssc, i);
1496 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1497 }
1498 }
1499 }
1aa99e6b 1500 }
cdd87c1d
KW
1501
1502 ssc_union(ssc,
1503 ored_cp_list,
1504 FALSE /* Already has been inverted */
1505 );
653099ff
GS
1506}
1507
b423522f
KW
1508PERL_STATIC_INLINE void
1509S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1510{
1511 PERL_ARGS_ASSERT_SSC_UNION;
1512
71068078 1513 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1514
1515 _invlist_union_maybe_complement_2nd(ssc->invlist,
1516 invlist,
1517 invert2nd,
1518 &ssc->invlist);
1519}
1520
1521PERL_STATIC_INLINE void
1522S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1523 SV* const invlist,
1524 const bool invert2nd)
1525{
1526 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1527
71068078 1528 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1529
1530 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1531 invlist,
1532 invert2nd,
1533 &ssc->invlist);
1534}
1535
1536PERL_STATIC_INLINE void
1537S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1538{
1539 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1540
71068078 1541 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1542
1543 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1544}
1545
1546PERL_STATIC_INLINE void
1547S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1548{
1549 /* AND just the single code point 'cp' into the SSC 'ssc' */
1550
1551 SV* cp_list = _new_invlist(2);
1552
1553 PERL_ARGS_ASSERT_SSC_CP_AND;
1554
71068078 1555 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1556
1557 cp_list = add_cp_to_invlist(cp_list, cp);
1558 ssc_intersection(ssc, cp_list,
1559 FALSE /* Not inverted */
1560 );
1561 SvREFCNT_dec_NN(cp_list);
1562}
1563
1564PERL_STATIC_INLINE void
dc3bf405 1565S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1566{
1567 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1568 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1569
71068078 1570 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1571
1572 ANYOF_POSIXL_ZERO(ssc);
1573 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1574}
1575
b35552de
KW
1576#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1577
1578STATIC bool
1579S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1580{
1581 /* The synthetic start class is used to hopefully quickly winnow down
1582 * places where a pattern could start a match in the target string. If it
1583 * doesn't really narrow things down that much, there isn't much point to
1584 * having the overhead of using it. This function uses some very crude
1585 * heuristics to decide if to use the ssc or not.
1586 *
1587 * It returns TRUE if 'ssc' rules out more than half what it considers to
1588 * be the "likely" possible matches, but of course it doesn't know what the
1589 * actual things being matched are going to be; these are only guesses
1590 *
1591 * For /l matches, it assumes that the only likely matches are going to be
1592 * in the 0-255 range, uniformly distributed, so half of that is 127
1593 * For /a and /d matches, it assumes that the likely matches will be just
1594 * the ASCII range, so half of that is 63
1595 * For /u and there isn't anything matching above the Latin1 range, it
1596 * assumes that that is the only range likely to be matched, and uses
1597 * half that as the cut-off: 127. If anything matches above Latin1,
1598 * it assumes that all of Unicode could match (uniformly), except for
1599 * non-Unicode code points and things in the General Category "Other"
1600 * (unassigned, private use, surrogates, controls and formats). This
1601 * is a much large number. */
1602
1603 const U32 max_match = (LOC)
1604 ? 127
1605 : (! UNI_SEMANTICS)
1606 ? 63
1607 : (invlist_highest(ssc->invlist) < 256)
1608 ? 127
1609 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1610 U32 count = 0; /* Running total of number of code points matched by
1611 'ssc' */
1612 UV start, end; /* Start and end points of current range in inversion
1613 list */
1614
1615 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1616
1617 invlist_iterinit(ssc->invlist);
1618 while (invlist_iternext(ssc->invlist, &start, &end)) {
1619
1620 /* /u is the only thing that we expect to match above 255; so if not /u
1621 * and even if there are matches above 255, ignore them. This catches
1622 * things like \d under /d which does match the digits above 255, but
1623 * since the pattern is /d, it is not likely to be expecting them */
1624 if (! UNI_SEMANTICS) {
1625 if (start > 255) {
1626 break;
1627 }
1628 end = MIN(end, 255);
1629 }
1630 count += end - start + 1;
1631 if (count > max_match) {
1632 invlist_iterfinish(ssc->invlist);
1633 return FALSE;
1634 }
1635 }
1636
1637 return TRUE;
1638}
1639
1640
b423522f
KW
1641STATIC void
1642S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1643{
1644 /* The inversion list in the SSC is marked mortal; now we need a more
1645 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1646 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1647 * map */
b423522f
KW
1648
1649 SV* invlist = invlist_clone(ssc->invlist);
1650
1651 PERL_ARGS_ASSERT_SSC_FINALIZE;
1652
71068078 1653 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1654
a0dd4231 1655 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1656 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1657 * by the time we reach here */
f240c685
KW
1658 assert(! (ANYOF_FLAGS(ssc)
1659 & ~( ANYOF_COMMON_FLAGS
1660 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)));
a0dd4231 1661
b423522f
KW
1662 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1663
1ee208c4
KW
1664 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1665 NULL, NULL, NULL, FALSE);
b423522f 1666
85c8e306
KW
1667 /* Make sure is clone-safe */
1668 ssc->invlist = NULL;
1669
e0e1be5f 1670 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1671 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1672 }
1462525b 1673
b2e90ddf
KW
1674 if (RExC_contains_locale) {
1675 OP(ssc) = ANYOFL;
1676 }
1677
1462525b 1678 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1679}
1680
a3621e74
YO
1681#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1682#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1683#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1684#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1685 ? (TRIE_LIST_CUR( idx ) - 1) \
1686 : 0 )
a3621e74 1687
3dab1dad
YO
1688
1689#ifdef DEBUGGING
07be1b83 1690/*
2b8b4781
NC
1691 dump_trie(trie,widecharmap,revcharmap)
1692 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1693 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1694
1695 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1696 The _interim_ variants are used for debugging the interim
1697 tables that are used to generate the final compressed
1698 representation which is what dump_trie expects.
1699
486ec47a 1700 Part of the reason for their existence is to provide a form
3dab1dad 1701 of documentation as to how the different representations function.
07be1b83
YO
1702
1703*/
3dab1dad
YO
1704
1705/*
3dab1dad
YO
1706 Dumps the final compressed table form of the trie to Perl_debug_log.
1707 Used for debugging make_trie().
1708*/
b9a59e08 1709
3dab1dad 1710STATIC void
2b8b4781
NC
1711S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1712 AV *revcharmap, U32 depth)
3dab1dad
YO
1713{
1714 U32 state;
ab3bbdeb 1715 SV *sv=sv_newmortal();
55eed653 1716 int colwidth= widecharmap ? 6 : 4;
2e64971a 1717 U16 word;
3dab1dad
YO
1718 GET_RE_DEBUG_FLAGS_DECL;
1719
7918f24d 1720 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1721
3dab1dad
YO
1722 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1723 (int)depth * 2 + 2,"",
1724 "Match","Base","Ofs" );
1725
1726 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1727 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1728 if ( tmp ) {
538e84ed 1729 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1730 colwidth,
538e84ed 1731 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1732 PL_colors[0], PL_colors[1],
1733 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1734 PERL_PV_ESCAPE_FIRSTCHAR
1735 )
ab3bbdeb 1736 );
3dab1dad
YO
1737 }
1738 }
1739 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1740 (int)depth * 2 + 2,"");
1741
1742 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1743 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1744 PerlIO_printf( Perl_debug_log, "\n");
1745
1e2e3d02 1746 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1747 const U32 base = trie->states[ state ].trans.base;
3dab1dad 1748
538e84ed
KW
1749 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1750 (int)depth * 2 + 2,"", (UV)state);
3dab1dad
YO
1751
1752 if ( trie->states[ state ].wordnum ) {
538e84ed
KW
1753 PerlIO_printf( Perl_debug_log, " W%4X",
1754 trie->states[ state ].wordnum );
3dab1dad
YO
1755 } else {
1756 PerlIO_printf( Perl_debug_log, "%6s", "" );
1757 }
1758
1759 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1760
1761 if ( base ) {
1762 U32 ofs = 0;
1763
1764 while( ( base + ofs < trie->uniquecharcount ) ||
1765 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
1766 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1767 != state))
3dab1dad
YO
1768 ofs++;
1769
1770 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1771
1772 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
1773 if ( ( base + ofs >= trie->uniquecharcount )
1774 && ( base + ofs - trie->uniquecharcount
1775 < trie->lasttrans )
1776 && trie->trans[ base + ofs
1777 - trie->uniquecharcount ].check == state )
3dab1dad 1778 {
ab3bbdeb
YO
1779 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1780 colwidth,
538e84ed
KW
1781 (UV)trie->trans[ base + ofs
1782 - trie->uniquecharcount ].next );
3dab1dad 1783 } else {
ab3bbdeb 1784 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1785 }
1786 }
1787
1788 PerlIO_printf( Perl_debug_log, "]");
1789
1790 }
1791 PerlIO_printf( Perl_debug_log, "\n" );
1792 }
538e84ed
KW
1793 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1794 (int)depth*2, "");
2e64971a
DM
1795 for (word=1; word <= trie->wordcount; word++) {
1796 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1797 (int)word, (int)(trie->wordinfo[word].prev),
1798 (int)(trie->wordinfo[word].len));
1799 }
1800 PerlIO_printf(Perl_debug_log, "\n" );
538e84ed 1801}
3dab1dad 1802/*
3dab1dad 1803 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 1804 List tries normally only are used for construction when the number of
3dab1dad
YO
1805 possible chars (trie->uniquecharcount) is very high.
1806 Used for debugging make_trie().
1807*/
1808STATIC void
55eed653 1809S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1810 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1811 U32 depth)
3dab1dad
YO
1812{
1813 U32 state;
ab3bbdeb 1814 SV *sv=sv_newmortal();
55eed653 1815 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1816 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1817
1818 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1819
3dab1dad 1820 /* print out the table precompression. */
ab3bbdeb
YO
1821 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1822 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1823 "------:-----+-----------------\n" );
538e84ed 1824
3dab1dad
YO
1825 for( state=1 ; state < next_alloc ; state ++ ) {
1826 U16 charid;
538e84ed 1827
ab3bbdeb 1828 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1829 (int)depth * 2 + 2,"", (UV)state );
1830 if ( ! trie->states[ state ].wordnum ) {
1831 PerlIO_printf( Perl_debug_log, "%5s| ","");
1832 } else {
1833 PerlIO_printf( Perl_debug_log, "W%4x| ",
1834 trie->states[ state ].wordnum
1835 );
1836 }
1837 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
1838 SV ** const tmp = av_fetch( revcharmap,
1839 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1840 if ( tmp ) {
1841 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1842 colwidth,
538e84ed
KW
1843 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1844 colwidth,
1845 PL_colors[0], PL_colors[1],
1846 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1847 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 1848 ) ,
1e2e3d02
YO
1849 TRIE_LIST_ITEM(state,charid).forid,
1850 (UV)TRIE_LIST_ITEM(state,charid).newstate
1851 );
538e84ed 1852 if (!(charid % 10))
664e119d
RGS
1853 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1854 (int)((depth * 2) + 14), "");
1e2e3d02 1855 }
ab3bbdeb
YO
1856 }
1857 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad 1858 }
538e84ed 1859}
3dab1dad
YO
1860
1861/*
3dab1dad 1862 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
1863 This is the normal DFA style state transition table, with a few
1864 twists to facilitate compression later.
3dab1dad
YO
1865 Used for debugging make_trie().
1866*/
1867STATIC void
55eed653 1868S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1869 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1870 U32 depth)
3dab1dad
YO
1871{
1872 U32 state;
1873 U16 charid;
ab3bbdeb 1874 SV *sv=sv_newmortal();
55eed653 1875 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1876 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1877
1878 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 1879
3dab1dad
YO
1880 /*
1881 print out the table precompression so that we can do a visual check
1882 that they are identical.
1883 */
538e84ed 1884
3dab1dad
YO
1885 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1886
1887 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1888 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1889 if ( tmp ) {
538e84ed 1890 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1891 colwidth,
538e84ed 1892 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1893 PL_colors[0], PL_colors[1],
1894 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1895 PERL_PV_ESCAPE_FIRSTCHAR
1896 )
ab3bbdeb 1897 );
3dab1dad
YO
1898 }
1899 }
1900
1901 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1902
1903 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1904 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1905 }
1906
1907 PerlIO_printf( Perl_debug_log, "\n" );
1908
1909 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1910
538e84ed 1911 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
3dab1dad
YO
1912 (int)depth * 2 + 2,"",
1913 (UV)TRIE_NODENUM( state ) );
1914
1915 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1916 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1917 if (v)
1918 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1919 else
1920 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1921 }
1922 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
538e84ed
KW
1923 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1924 (UV)trie->trans[ state ].check );
3dab1dad 1925 } else {
538e84ed
KW
1926 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1927 (UV)trie->trans[ state ].check,
3dab1dad
YO
1928 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1929 }
1930 }
07be1b83 1931}
3dab1dad
YO
1932
1933#endif
1934
2e64971a 1935
786e8c11
YO
1936/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1937 startbranch: the first branch in the whole branch sequence
1938 first : start branch of sequence of branch-exact nodes.
1939 May be the same as startbranch
1940 last : Thing following the last branch.
1941 May be the same as tail.
1942 tail : item following the branch sequence
1943 count : words in the sequence
a4525e78 1944 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 1945 depth : indent depth
3dab1dad 1946
786e8c11 1947Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1948
786e8c11
YO
1949A trie is an N'ary tree where the branches are determined by digital
1950decomposition of the key. IE, at the root node you look up the 1st character and
1951follow that branch repeat until you find the end of the branches. Nodes can be
1952marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1953
786e8c11 1954 /he|she|his|hers/
72f13be8 1955
786e8c11
YO
1956would convert into the following structure. Numbers represent states, letters
1957following numbers represent valid transitions on the letter from that state, if
1958the number is in square brackets it represents an accepting state, otherwise it
1959will be in parenthesis.
07be1b83 1960
786e8c11
YO
1961 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1962 | |
1963 | (2)
1964 | |
1965 (1) +-i->(6)-+-s->[7]
1966 |
1967 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1968
786e8c11
YO
1969 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1970
1971This shows that when matching against the string 'hers' we will begin at state 1
1972read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1973then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1974is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1975single traverse. We store a mapping from accepting to state to which word was
1976matched, and then when we have multiple possibilities we try to complete the
b8fda935 1977rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
1978
1979The only prior NFA like behaviour that would be changed by the TRIE support is
1980the silent ignoring of duplicate alternations which are of the form:
1981
1982 / (DUPE|DUPE) X? (?{ ... }) Y /x
1983
4b714af6 1984Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1985and without the optimisation. With the optimisations dupes will be silently
486ec47a 1986ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1987the following demonstrates:
1988
1989 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1990
1991which prints out 'word' three times, but
1992
1993 'words'=~/(word|word|word)(?{ print $1 })S/
1994
1995which doesnt print it out at all. This is due to other optimisations kicking in.
1996
1997Example of what happens on a structural level:
1998
486ec47a 1999The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2000
2001 1: CURLYM[1] {1,32767}(18)
2002 5: BRANCH(8)
2003 6: EXACT <ac>(16)
2004 8: BRANCH(11)
2005 9: EXACT <ad>(16)
2006 11: BRANCH(14)
2007 12: EXACT <ab>(16)
2008 16: SUCCEED(0)
2009 17: NOTHING(18)
2010 18: END(0)
2011
2012This would be optimizable with startbranch=5, first=5, last=16, tail=16
2013and should turn into:
2014
2015 1: CURLYM[1] {1,32767}(18)
2016 5: TRIE(16)
2017 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2018 <ac>
2019 <ad>
2020 <ab>
2021 16: SUCCEED(0)
2022 17: NOTHING(18)
2023 18: END(0)
2024
2025Cases where tail != last would be like /(?foo|bar)baz/:
2026
2027 1: BRANCH(4)
2028 2: EXACT <foo>(8)
2029 4: BRANCH(7)
2030 5: EXACT <bar>(8)
2031 7: TAIL(8)
2032 8: EXACT <baz>(10)
2033 10: END(0)
2034
2035which would be optimizable with startbranch=1, first=1, last=7, tail=8
2036and would end up looking like:
2037
2038 1: TRIE(8)
2039 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2040 <foo>
2041 <bar>
2042 7: TAIL(8)
2043 8: EXACT <baz>(10)
2044 10: END(0)
2045
c80e42f3 2046 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2047
2048is the recommended Unicode-aware way of saying
2049
2050 *(d++) = uv;
2051*/
2052
fab2782b 2053#define TRIE_STORE_REVCHAR(val) \
786e8c11 2054 STMT_START { \
73031816 2055 if (UTF) { \
668fcfea 2056 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2057 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2058 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2059 SvCUR_set(zlopp, kapow - flrbbbbb); \
2060 SvPOK_on(zlopp); \
2061 SvUTF8_on(zlopp); \
2062 av_push(revcharmap, zlopp); \
2063 } else { \
fab2782b 2064 char ooooff = (char)val; \
73031816
NC
2065 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2066 } \
2067 } STMT_END
786e8c11 2068
914a25d5
KW
2069/* This gets the next character from the input, folding it if not already
2070 * folded. */
2071#define TRIE_READ_CHAR STMT_START { \
2072 wordlen++; \
2073 if ( UTF ) { \
2074 /* if it is UTF then it is either already folded, or does not need \
2075 * folding */ \
1c1d615a 2076 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2077 } \
2078 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2079 /* This folder implies Unicode rules, which in the range expressible \
2080 * by not UTF is the lower case, with the two exceptions, one of \
2081 * which should have been taken care of before calling this */ \
2082 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2083 uvc = toLOWER_L1(*uc); \
2084 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2085 len = 1; \
914a25d5
KW
2086 } else { \
2087 /* raw data, will be folded later if needed */ \
2088 uvc = (U32)*uc; \
2089 len = 1; \
2090 } \
786e8c11
YO
2091} STMT_END
2092
2093
2094
2095#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2096 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
2097 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2098 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
2099 } \
2100 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2101 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2102 TRIE_LIST_CUR( state )++; \
2103} STMT_END
07be1b83 2104
786e8c11
YO
2105#define TRIE_LIST_NEW(state) STMT_START { \
2106 Newxz( trie->states[ state ].trans.list, \
2107 4, reg_trie_trans_le ); \
2108 TRIE_LIST_CUR( state ) = 1; \
2109 TRIE_LIST_LEN( state ) = 4; \
2110} STMT_END
07be1b83 2111
786e8c11
YO
2112#define TRIE_HANDLE_WORD(state) STMT_START { \
2113 U16 dupe= trie->states[ state ].wordnum; \
2114 regnode * const noper_next = regnext( noper ); \
2115 \
786e8c11
YO
2116 DEBUG_r({ \
2117 /* store the word for dumping */ \
2118 SV* tmp; \
2119 if (OP(noper) != NOTHING) \
740cce10 2120 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2121 else \
740cce10 2122 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2123 av_push( trie_words, tmp ); \
786e8c11
YO
2124 }); \
2125 \
2126 curword++; \
2e64971a
DM
2127 trie->wordinfo[curword].prev = 0; \
2128 trie->wordinfo[curword].len = wordlen; \
2129 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2130 \
2131 if ( noper_next < tail ) { \
2132 if (!trie->jump) \
538e84ed
KW
2133 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2134 sizeof(U16) ); \
7f69552c 2135 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2136 if (!jumper) \
2137 jumper = noper_next; \
2138 if (!nextbranch) \
2139 nextbranch= regnext(cur); \
2140 } \
2141 \
2142 if ( dupe ) { \
2e64971a
DM
2143 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2144 /* chain, so that when the bits of chain are later */\
2145 /* linked together, the dups appear in the chain */\
2146 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2147 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2148 } else { \
2149 /* we haven't inserted this word yet. */ \
2150 trie->states[ state ].wordnum = curword; \
2151 } \
2152} STMT_END
07be1b83 2153
3dab1dad 2154
786e8c11
YO
2155#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2156 ( ( base + charid >= ucharcount \
2157 && base + charid < ubound \
2158 && state == trie->trans[ base - ucharcount + charid ].check \
2159 && trie->trans[ base - ucharcount + charid ].next ) \
2160 ? trie->trans[ base - ucharcount + charid ].next \
2161 : ( state==1 ? special : 0 ) \
2162 )
3dab1dad 2163
786e8c11
YO
2164#define MADE_TRIE 1
2165#define MADE_JUMP_TRIE 2
2166#define MADE_EXACT_TRIE 4
3dab1dad 2167
a3621e74 2168STATIC I32
538e84ed
KW
2169S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2170 regnode *first, regnode *last, regnode *tail,
2171 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2172{
2173 /* first pass, loop through and scan words */
2174 reg_trie_data *trie;
55eed653 2175 HV *widecharmap = NULL;
2b8b4781 2176 AV *revcharmap = newAV();
a3621e74 2177 regnode *cur;
a3621e74
YO
2178 STRLEN len = 0;
2179 UV uvc = 0;
2180 U16 curword = 0;
2181 U32 next_alloc = 0;
786e8c11
YO
2182 regnode *jumper = NULL;
2183 regnode *nextbranch = NULL;
7f69552c 2184 regnode *convert = NULL;
2e64971a 2185 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2186 /* we just use folder as a flag in utf8 */
1e696034 2187 const U8 * folder = NULL;
a3621e74 2188
2b8b4781 2189#ifdef DEBUGGING
cf78de0b 2190 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2191 AV *trie_words = NULL;
2192 /* along with revcharmap, this only used during construction but both are
2193 * useful during debugging so we store them in the struct when debugging.
8e11feef 2194 */
2b8b4781 2195#else
cf78de0b 2196 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2197 STRLEN trie_charcount=0;
3dab1dad 2198#endif
2b8b4781 2199 SV *re_trie_maxbuff;
a3621e74 2200 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2201
2202 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2203#ifndef DEBUGGING
2204 PERL_UNUSED_ARG(depth);
2205#endif
a3621e74 2206
1e696034 2207 switch (flags) {
a4525e78 2208 case EXACT: case EXACTL: break;
2f7f8cb1 2209 case EXACTFA:
fab2782b 2210 case EXACTFU_SS:
a4525e78
KW
2211 case EXACTFU:
2212 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2213 case EXACTF: folder = PL_fold; break;
fab2782b 2214 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2215 }
2216
c944940b 2217 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2218 trie->refcount = 1;
3dab1dad 2219 trie->startstate = 1;
786e8c11 2220 trie->wordcount = word_count;
f8fc2ecf 2221 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2222 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2223 if (flags == EXACT || flags == EXACTL)
c944940b 2224 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2225 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2226 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2227
a3621e74 2228 DEBUG_r({
2b8b4781 2229 trie_words = newAV();
a3621e74 2230 });
a3621e74 2231
0111c4fd 2232 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2233 assert(re_trie_maxbuff);
a3621e74 2234 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2235 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2236 }
df826430 2237 DEBUG_TRIE_COMPILE_r({
538e84ed
KW
2238 PerlIO_printf( Perl_debug_log,
2239 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2240 (int)depth * 2 + 2, "",
2241 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2242 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2243 });
538e84ed 2244
7f69552c
YO
2245 /* Find the node we are going to overwrite */
2246 if ( first == startbranch && OP( last ) != BRANCH ) {
2247 /* whole branch chain */
2248 convert = first;
2249 } else {
2250 /* branch sub-chain */
2251 convert = NEXTOPER( first );
2252 }
538e84ed 2253
a3621e74
YO
2254 /* -- First loop and Setup --
2255
2256 We first traverse the branches and scan each word to determine if it
2257 contains widechars, and how many unique chars there are, this is
2258 important as we have to build a table with at least as many columns as we
2259 have unique chars.
2260
2261 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2262 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2263 the native representation of the character value as the key and IV's for
2264 the coded index.
a3621e74
YO
2265
2266 *TODO* If we keep track of how many times each character is used we can
2267 remap the columns so that the table compression later on is more
3b753521 2268 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2269 middle and the least common are on the outside. IMO this would be better
2270 than a most to least common mapping as theres a decent chance the most
2271 common letter will share a node with the least common, meaning the node
486ec47a 2272 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2273 case is when we have the least common nodes twice.
2274
2275 */
2276
a3621e74 2277 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2278 regnode *noper = NEXTOPER( cur );
e1ec3a88 2279 const U8 *uc = (U8*)STRING( noper );
df826430 2280 const U8 *e = uc + STR_LEN( noper );
bc031a7d 2281 int foldlen = 0;
07be1b83 2282 U32 wordlen = 0; /* required init */
bc031a7d
KW
2283 STRLEN minchars = 0;
2284 STRLEN maxchars = 0;
538e84ed
KW
2285 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2286 bitmap?*/
a3621e74 2287
3dab1dad 2288 if (OP(noper) == NOTHING) {
df826430
YO
2289 regnode *noper_next= regnext(noper);
2290 if (noper_next != tail && OP(noper_next) == flags) {
2291 noper = noper_next;
2292 uc= (U8*)STRING(noper);
2293 e= uc + STR_LEN(noper);
2294 trie->minlen= STR_LEN(noper);
2295 } else {
2296 trie->minlen= 0;
2297 continue;
2298 }
3dab1dad 2299 }
df826430 2300
fab2782b 2301 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2302 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2303 regardless of encoding */
fab2782b
YO
2304 if (OP( noper ) == EXACTFU_SS) {
2305 /* false positives are ok, so just set this */
0dc4a61d 2306 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2307 }
2308 }
bc031a7d
KW
2309 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2310 branch */
3dab1dad 2311 TRIE_CHARCOUNT(trie)++;
a3621e74 2312 TRIE_READ_CHAR;
645de4ce 2313
bc031a7d
KW
2314 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2315 * is in effect. Under /i, this character can match itself, or
2316 * anything that folds to it. If not under /i, it can match just
2317 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2318 * all fold to k, and all are single characters. But some folds
2319 * expand to more than one character, so for example LATIN SMALL
2320 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2321 * the string beginning at 'uc' is 'ffi', it could be matched by
2322 * three characters, or just by the one ligature character. (It
2323 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2324 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2325 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2326 * match.) The trie needs to know the minimum and maximum number
2327 * of characters that could match so that it can use size alone to
2328 * quickly reject many match attempts. The max is simple: it is
2329 * the number of folded characters in this branch (since a fold is
2330 * never shorter than what folds to it. */
2331
2332 maxchars++;
2333
2334 /* And the min is equal to the max if not under /i (indicated by
2335 * 'folder' being NULL), or there are no multi-character folds. If
2336 * there is a multi-character fold, the min is incremented just
2337 * once, for the character that folds to the sequence. Each
2338 * character in the sequence needs to be added to the list below of
2339 * characters in the trie, but we count only the first towards the
2340 * min number of characters needed. This is done through the
2341 * variable 'foldlen', which is returned by the macros that look
2342 * for these sequences as the number of bytes the sequence
2343 * occupies. Each time through the loop, we decrement 'foldlen' by
2344 * how many bytes the current char occupies. Only when it reaches
2345 * 0 do we increment 'minchars' or look for another multi-character
2346 * sequence. */
2347 if (folder == NULL) {
2348 minchars++;
2349 }
2350 else if (foldlen > 0) {
2351 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2352 }
2353 else {
bc031a7d
KW
2354 minchars++;
2355
2356 /* See if *uc is the beginning of a multi-character fold. If
2357 * so, we decrement the length remaining to look at, to account
2358 * for the current character this iteration. (We can use 'uc'
2359 * instead of the fold returned by TRIE_READ_CHAR because for
2360 * non-UTF, the latin1_safe macro is smart enough to account
2361 * for all the unfolded characters, and because for UTF, the
2362 * string will already have been folded earlier in the
2363 * compilation process */
2364 if (UTF) {
2365 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2366 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2367 }
2368 }
bc031a7d
KW
2369 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2370 foldlen--;
2371 }
645de4ce 2372 }
bc031a7d
KW
2373
2374 /* The current character (and any potential folds) should be added
2375 * to the possible matching characters for this position in this
2376 * branch */
a3621e74 2377 if ( uvc < 256 ) {
fab2782b
YO
2378 if ( folder ) {
2379 U8 folded= folder[ (U8) uvc ];
2380 if ( !trie->charmap[ folded ] ) {
2381 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2382 TRIE_STORE_REVCHAR( folded );
2383 }
2384 }
a3621e74
YO
2385 if ( !trie->charmap[ uvc ] ) {
2386 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2387 TRIE_STORE_REVCHAR( uvc );
a3621e74 2388 }
02daf0ab 2389 if ( set_bit ) {
62012aee
KW
2390 /* store the codepoint in the bitmap, and its folded
2391 * equivalent. */
fab2782b 2392 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2393
2394 /* store the folded codepoint */
fab2782b 2395 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2396
2397 if ( !UTF ) {
2398 /* store first byte of utf8 representation of
acdf4139 2399 variant codepoints */
6f2d5cbc 2400 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2401 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2402 }
2403 }
02daf0ab
YO
2404 set_bit = 0; /* We've done our bit :-) */
2405 }
a3621e74 2406 } else {
bc031a7d
KW
2407
2408 /* XXX We could come up with the list of code points that fold
2409 * to this using PL_utf8_foldclosures, except not for
2410 * multi-char folds, as there may be multiple combinations
2411 * there that could work, which needs to wait until runtime to
2412 * resolve (The comment about LIGATURE FFI above is such an
2413 * example */
2414
a3621e74 2415 SV** svpp;
55eed653
NC
2416 if ( !widecharmap )
2417 widecharmap = newHV();
a3621e74 2418
55eed653 2419 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2420
2421 if ( !svpp )
e4584336 2422 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2423
2424 if ( !SvTRUE( *svpp ) ) {
2425 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2426 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2427 }
2428 }
bc031a7d
KW
2429 } /* end loop through characters in this branch of the trie */
2430
2431 /* We take the min and max for this branch and combine to find the min
2432 * and max for all branches processed so far */
3dab1dad 2433 if( cur == first ) {
bc031a7d
KW
2434 trie->minlen = minchars;
2435 trie->maxlen = maxchars;
2436 } else if (minchars < trie->minlen) {
2437 trie->minlen = minchars;
2438 } else if (maxchars > trie->maxlen) {
2439 trie->maxlen = maxchars;
fab2782b 2440 }
a3621e74
YO
2441 } /* end first pass */
2442 DEBUG_TRIE_COMPILE_r(
538e84ed
KW
2443 PerlIO_printf( Perl_debug_log,
2444 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3dab1dad 2445 (int)depth * 2 + 2,"",
55eed653 2446 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2447 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2448 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2449 );
a3621e74
YO
2450
2451 /*
2452 We now know what we are dealing with in terms of unique chars and
2453 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2454 representation using a flat table will take. If it's over a reasonable
2455 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2456 conservative but potentially much slower representation using an array
2457 of lists.
2458
2459 At the end we convert both representations into the same compressed
2460 form that will be used in regexec.c for matching with. The latter
2461 is a form that cannot be used to construct with but has memory
2462 properties similar to the list form and access properties similar
2463 to the table form making it both suitable for fast searches and
2464 small enough that its feasable to store for the duration of a program.
2465
2466 See the comment in the code where the compressed table is produced
2467 inplace from the flat tabe representation for an explanation of how
2468 the compression works.
2469
2470 */
2471
2472
2e64971a
DM
2473 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2474 prev_states[1] = 0;
2475
538e84ed
KW
2476 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2477 > SvIV(re_trie_maxbuff) )
2478 {
a3621e74
YO
2479 /*
2480 Second Pass -- Array Of Lists Representation
2481
2482 Each state will be represented by a list of charid:state records
2483 (reg_trie_trans_le) the first such element holds the CUR and LEN
2484 points of the allocated array. (See defines above).
2485
2486 We build the initial structure using the lists, and then convert
2487 it into the compressed table form which allows faster lookups
2488 (but cant be modified once converted).
a3621e74
YO
2489 */
2490
a3621e74
YO
2491 STRLEN transcount = 1;
2492
538e84ed 2493 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2494 "%*sCompiling trie using list compiler\n",
2495 (int)depth * 2 + 2, ""));
686b73d4 2496
c944940b
JH
2497 trie->states = (reg_trie_state *)
2498 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2499 sizeof(reg_trie_state) );
a3621e74
YO
2500 TRIE_LIST_NEW(1);
2501 next_alloc = 2;
2502
2503 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2504
df826430 2505 regnode *noper = NEXTOPER( cur );
c445ea15 2506 U8 *uc = (U8*)STRING( noper );
df826430 2507 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
2508 U32 state = 1; /* required init */
2509 U16 charid = 0; /* sanity init */
07be1b83 2510 U32 wordlen = 0; /* required init */
c445ea15 2511
df826430
YO
2512 if (OP(noper) == NOTHING) {
2513 regnode *noper_next= regnext(noper);
2514 if (noper_next != tail && OP(noper_next) == flags) {
2515 noper = noper_next;
2516 uc= (U8*)STRING(noper);
2517 e= uc + STR_LEN(noper);
2518 }
2519 }
2520
3dab1dad 2521 if (OP(noper) != NOTHING) {
786e8c11 2522 for ( ; uc < e ; uc += len ) {
c445ea15 2523
786e8c11 2524 TRIE_READ_CHAR;
c445ea15 2525
786e8c11
YO
2526 if ( uvc < 256 ) {
2527 charid = trie->charmap[ uvc ];
c445ea15 2528 } else {
538e84ed
KW
2529 SV** const svpp = hv_fetch( widecharmap,
2530 (char*)&uvc,
2531 sizeof( UV ),
2532 0);
786e8c11
YO
2533 if ( !svpp ) {
2534 charid = 0;
2535 } else {
2536 charid=(U16)SvIV( *svpp );
2537 }
c445ea15 2538 }
538e84ed
KW
2539 /* charid is now 0 if we dont know the char read, or
2540 * nonzero if we do */
786e8c11 2541 if ( charid ) {
a3621e74 2542
786e8c11
YO
2543 U16 check;
2544 U32 newstate = 0;
a3621e74 2545
786e8c11
YO
2546 charid--;
2547 if ( !trie->states[ state ].trans.list ) {
2548 TRIE_LIST_NEW( state );
c445ea15 2549 }
538e84ed
KW
2550 for ( check = 1;
2551 check <= TRIE_LIST_USED( state );
2552 check++ )
2553 {
2554 if ( TRIE_LIST_ITEM( state, check ).forid
2555 == charid )
2556 {
786e8c11
YO
2557 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2558 break;
2559 }
2560 }
2561 if ( ! newstate ) {
2562 newstate = next_alloc++;
2e64971a 2563 prev_states[newstate] = state;
786e8c11
YO
2564 TRIE_LIST_PUSH( state, charid, newstate );
2565 transcount++;
2566 }
2567 state = newstate;
2568 } else {
2569 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2570 }
a28509cc 2571 }
c445ea15 2572 }
3dab1dad 2573 TRIE_HANDLE_WORD(state);
a3621e74
YO
2574
2575 } /* end second pass */
2576
1e2e3d02 2577 /* next alloc is the NEXT state to be allocated */
538e84ed 2578 trie->statecount = next_alloc;
c944940b
JH
2579 trie->states = (reg_trie_state *)
2580 PerlMemShared_realloc( trie->states,
2581 next_alloc
2582 * sizeof(reg_trie_state) );
a3621e74 2583
3dab1dad 2584 /* and now dump it out before we compress it */
2b8b4781
NC
2585 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2586 revcharmap, next_alloc,
2587 depth+1)
1e2e3d02 2588 );
a3621e74 2589
c944940b
JH
2590 trie->trans = (reg_trie_trans *)
2591 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2592 {
2593 U32 state;
a3621e74
YO
2594 U32 tp = 0;
2595 U32 zp = 0;
2596
2597
2598 for( state=1 ; state < next_alloc ; state ++ ) {
2599 U32 base=0;
2600
2601 /*
2602 DEBUG_TRIE_COMPILE_MORE_r(
2603 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2604 );
2605 */
2606
2607 if (trie->states[state].trans.list) {
2608 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2609 U16 maxid=minid;
a28509cc 2610 U16 idx;
a3621e74
YO
2611
2612 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2613 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2614 if ( forid < minid ) {
2615 minid=forid;
2616 } else if ( forid > maxid ) {
2617 maxid=forid;
2618 }
a3621e74
YO
2619 }
2620 if ( transcount < tp + maxid - minid + 1) {
2621 transcount *= 2;
c944940b
JH
2622 trie->trans = (reg_trie_trans *)
2623 PerlMemShared_realloc( trie->trans,
446bd890
NC
2624 transcount
2625 * sizeof(reg_trie_trans) );
538e84ed
KW
2626 Zero( trie->trans + (transcount / 2),
2627 transcount / 2,
2628 reg_trie_trans );
a3621e74
YO
2629 }
2630 base = trie->uniquecharcount + tp - minid;
2631 if ( maxid == minid ) {
2632 U32 set = 0;
2633 for ( ; zp < tp ; zp++ ) {
2634 if ( ! trie->trans[ zp ].next ) {
2635 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2636 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2637 1).newstate;
a3621e74
YO
2638 trie->trans[ zp ].check = state;
2639 set = 1;
2640 break;
2641 }
2642 }
2643 if ( !set ) {
538e84ed
KW
2644 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2645 1).newstate;
a3621e74
YO
2646 trie->trans[ tp ].check = state;
2647 tp++;
2648 zp = tp;
2649 }
2650 } else {
2651 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2652 const U32 tid = base
2653 - trie->uniquecharcount
2654 + TRIE_LIST_ITEM( state, idx ).forid;
2655 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2656 idx ).newstate;
a3621e74
YO
2657 trie->trans[ tid ].check = state;
2658 }
2659 tp += ( maxid - minid + 1 );
2660 }
2661 Safefree(trie->states[ state ].trans.list);
2662 }
2663 /*
2664 DEBUG_TRIE_COMPILE_MORE_r(
2665 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2666 );
2667 */
2668 trie->states[ state ].trans.base=base;
2669 }
cc601c31 2670 trie->lasttrans = tp + 1;
a3621e74
YO
2671 }
2672 } else {
2673 /*
2674 Second Pass -- Flat Table Representation.
2675
b423522f
KW
2676 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2677 each. We know that we will need Charcount+1 trans at most to store
2678 the data (one row per char at worst case) So we preallocate both
2679 structures assuming worst case.
a3621e74
YO
2680
2681 We then construct the trie using only the .next slots of the entry
2682 structs.
2683
b423522f
KW
2684 We use the .check field of the first entry of the node temporarily
2685 to make compression both faster and easier by keeping track of how
2686 many non zero fields are in the node.
a3621e74
YO
2687
2688 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2689 transition.
2690
b423522f
KW
2691 There are two terms at use here: state as a TRIE_NODEIDX() which is
2692 a number representing the first entry of the node, and state as a
2693 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2694 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2695 if there are 2 entrys per node. eg:
a3621e74
YO
2696
2697 A B A B
2698 1. 2 4 1. 3 7
2699 2. 0 3 3. 0 5
2700 3. 0 0 5. 0 0
2701 4. 0 0 7. 0 0
2702
b423522f
KW
2703 The table is internally in the right hand, idx form. However as we
2704 also have to deal with the states array which is indexed by nodenum
2705 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2706
2707 */
538e84ed 2708 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2709 "%*sCompiling trie using table compiler\n",
2710 (int)depth * 2 + 2, ""));
3dab1dad 2711
c944940b
JH
2712 trie->trans = (reg_trie_trans *)
2713 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2714 * trie->uniquecharcount + 1,
2715 sizeof(reg_trie_trans) );
2716 trie->states = (reg_trie_state *)
2717 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2718 sizeof(reg_trie_state) );
a3621e74
YO
2719 next_alloc = trie->uniquecharcount + 1;
2720
3dab1dad 2721
a3621e74
YO
2722 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2723
df826430 2724 regnode *noper = NEXTOPER( cur );
a28509cc 2725 const U8 *uc = (U8*)STRING( noper );
df826430 2726 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2727
2728 U32 state = 1; /* required init */
2729
2730 U16 charid = 0; /* sanity init */
2731 U32 accept_state = 0; /* sanity init */
a3621e74 2732
07be1b83 2733 U32 wordlen = 0; /* required init */
a3621e74 2734
df826430
YO
2735 if (OP(noper) == NOTHING) {
2736 regnode *noper_next= regnext(noper);
2737 if (noper_next != tail && OP(noper_next) == flags) {
2738 noper = noper_next;
2739 uc= (U8*)STRING(noper);
2740 e= uc + STR_LEN(noper);
2741 }
2742 }
fab2782b 2743
3dab1dad 2744 if ( OP(noper) != NOTHING ) {
786e8c11 2745 for ( ; uc < e ; uc += len ) {
a3621e74 2746
786e8c11 2747 TRIE_READ_CHAR;
a3621e74 2748
786e8c11
YO
2749 if ( uvc < 256 ) {
2750 charid = trie->charmap[ uvc ];
2751 } else {
538e84ed
KW
2752 SV* const * const svpp = hv_fetch( widecharmap,
2753 (char*)&uvc,
2754 sizeof( UV ),
2755 0);
786e8c11 2756 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2757 }
786e8c11
YO
2758 if ( charid ) {
2759 charid--;
2760 if ( !trie->trans[ state + charid ].next ) {
2761 trie->trans[ state + charid ].next = next_alloc;
2762 trie->trans[ state ].check++;
2e64971a
DM
2763 prev_states[TRIE_NODENUM(next_alloc)]
2764 = TRIE_NODENUM(state);
786e8c11
YO
2765 next_alloc += trie->uniquecharcount;
2766 }
2767 state = trie->trans[ state + charid ].next;
2768 } else {
2769 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2770 }
538e84ed
KW
2771 /* charid is now 0 if we dont know the char read, or
2772 * nonzero if we do */
a3621e74 2773 }
a3621e74 2774 }
3dab1dad
YO
2775 accept_state = TRIE_NODENUM( state );
2776 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2777
2778 } /* end second pass */
2779
3dab1dad 2780 /* and now dump it out before we compress it */
2b8b4781
NC
2781 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2782 revcharmap,
2783 next_alloc, depth+1));
a3621e74 2784
a3621e74
YO
2785 {
2786 /*
2787 * Inplace compress the table.*
2788
2789 For sparse data sets the table constructed by the trie algorithm will
2790 be mostly 0/FAIL transitions or to put it another way mostly empty.
2791 (Note that leaf nodes will not contain any transitions.)
2792
2793 This algorithm compresses the tables by eliminating most such
2794 transitions, at the cost of a modest bit of extra work during lookup:
2795
2796 - Each states[] entry contains a .base field which indicates the
2797 index in the state[] array wheres its transition data is stored.
2798
3b753521 2799 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2800
2801 - If .base is nonzero then charid is added to it to find an entry in
2802 the trans array.
2803
2804 -If trans[states[state].base+charid].check!=state then the
2805 transition is taken to be a 0/Fail transition. Thus if there are fail
2806 transitions at the front of the node then the .base offset will point
2807 somewhere inside the previous nodes data (or maybe even into a node
2808 even earlier), but the .check field determines if the transition is
2809 valid.
2810
786e8c11 2811 XXX - wrong maybe?
a3621e74 2812 The following process inplace converts the table to the compressed
3b753521 2813 table: We first do not compress the root node 1,and mark all its
a3621e74 2814 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2815 allows us to do a DFA construction from the compressed table later,
2816 and ensures that any .base pointers we calculate later are greater
2817 than 0.
a3621e74
YO
2818
2819 - We set 'pos' to indicate the first entry of the second node.
2820
2821 - We then iterate over the columns of the node, finding the first and
2822 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2823 and set the .check pointers accordingly, and advance pos
2824 appropriately and repreat for the next node. Note that when we copy
2825 the next pointers we have to convert them from the original
2826 NODEIDX form to NODENUM form as the former is not valid post
2827 compression.
2828
2829 - If a node has no transitions used we mark its base as 0 and do not
2830 advance the pos pointer.
2831
2832 - If a node only has one transition we use a second pointer into the
2833 structure to fill in allocated fail transitions from other states.
2834 This pointer is independent of the main pointer and scans forward
2835 looking for null transitions that are allocated to a state. When it
2836 finds one it writes the single transition into the "hole". If the
786e8c11 2837 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2838
2839 - Once compressed we can Renew/realloc the structures to release the
2840 excess space.
2841
2842 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2843 specifically Fig 3.47 and the associated pseudocode.
2844
2845 demq
2846 */
a3b680e6 2847 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2848 U32 state, charid;
a3621e74 2849 U32 pos = 0, zp=0;
1e2e3d02 2850 trie->statecount = laststate;
a3621e74
YO
2851
2852 for ( state = 1 ; state < laststate ; state++ ) {
2853 U8 flag = 0;
a28509cc
AL
2854 const U32 stateidx = TRIE_NODEIDX( state );
2855 const U32 o_used = trie->trans[ stateidx ].check;
2856 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2857 trie->trans[ stateidx ].check = 0;
2858
538e84ed
KW
2859 for ( charid = 0;
2860 used && charid < trie->uniquecharcount;
2861 charid++ )
2862 {
a3621e74
YO
2863 if ( flag || trie->trans[ stateidx + charid ].next ) {
2864 if ( trie->trans[ stateidx + charid ].next ) {
2865 if (o_used == 1) {
2866 for ( ; zp < pos ; zp++ ) {
2867 if ( ! trie->trans[ zp ].next ) {
2868 break;
2869 }
2870 }
538e84ed
KW
2871 trie->states[ state ].trans.base
2872 = zp
2873 + trie->uniquecharcount
2874 - charid ;
2875 trie->trans[ zp ].next
2876 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2877 + charid ].next );
a3621e74
YO
2878 trie->trans[ zp ].check = state;
2879 if ( ++zp > pos ) pos = zp;
2880 break;
2881 }
2882 used--;
2883 }
2884 if ( !flag ) {
2885 flag = 1;
538e84ed
KW
2886 trie->states[ state ].trans.base
2887 = pos + trie->uniquecharcount - charid ;
a3621e74 2888 }
538e84ed
KW
2889 trie->trans[ pos ].next
2890 = SAFE_TRIE_NODENUM(
2891 trie->trans[ stateidx + charid ].next );
a3621e74
YO
2892 trie->trans[ pos ].check = state;
2893 pos++;
2894 }
2895 }
2896 }
cc601c31 2897 trie->lasttrans = pos + 1;
c944940b
JH
2898 trie->states = (reg_trie_state *)
2899 PerlMemShared_realloc( trie->states, laststate
2900 * sizeof(reg_trie_state) );
a3621e74 2901 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2902 PerlIO_printf( Perl_debug_log,
2903 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2904 (int)depth * 2 + 2,"",
2905 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2906 + 1 ),
2907 (IV)next_alloc,
2908 (IV)pos,
2909 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
2910 );
2911
2912 } /* end table compress */
2913 }
1e2e3d02 2914 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2915 PerlIO_printf(Perl_debug_log,
2916 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1e2e3d02
YO
2917 (int)depth * 2 + 2, "",
2918 (UV)trie->statecount,
2919 (UV)trie->lasttrans)
2920 );
cc601c31 2921 /* resize the trans array to remove unused space */
c944940b
JH
2922 trie->trans = (reg_trie_trans *)
2923 PerlMemShared_realloc( trie->trans, trie->lasttrans
2924 * sizeof(reg_trie_trans) );
a3621e74 2925
538e84ed 2926 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2927 U8 nodetype =(U8)(flags & 0xFF);
2928 char *str=NULL;
538e84ed 2929
07be1b83 2930#ifdef DEBUGGING
e62cc96a 2931 regnode *optimize = NULL;
7122b237
YO
2932#ifdef RE_TRACK_PATTERN_OFFSETS
2933
b57a0404
JH
2934 U32 mjd_offset = 0;
2935 U32 mjd_nodelen = 0;
7122b237
YO
2936#endif /* RE_TRACK_PATTERN_OFFSETS */
2937#endif /* DEBUGGING */
a3621e74 2938 /*
3dab1dad
YO
2939 This means we convert either the first branch or the first Exact,
2940 depending on whether the thing following (in 'last') is a branch
2941 or not and whther first is the startbranch (ie is it a sub part of
2942 the alternation or is it the whole thing.)
3b753521 2943 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2944 the whole branch sequence, including the first.
a3621e74 2945 */
3dab1dad 2946 /* Find the node we are going to overwrite */
7f69552c 2947 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2948 /* branch sub-chain */
3dab1dad 2949 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2950#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2951 DEBUG_r({
2952 mjd_offset= Node_Offset((convert));
2953 mjd_nodelen= Node_Length((convert));
2954 });
7122b237 2955#endif
7f69552c 2956 /* whole branch chain */
7122b237
YO
2957 }
2958#ifdef RE_TRACK_PATTERN_OFFSETS
2959 else {
7f69552c
YO
2960 DEBUG_r({
2961 const regnode *nop = NEXTOPER( convert );
2962 mjd_offset= Node_Offset((nop));
2963 mjd_nodelen= Node_Length((nop));
2964 });
07be1b83
YO
2965 }
2966 DEBUG_OPTIMISE_r(
538e84ed
KW
2967 PerlIO_printf(Perl_debug_log,
2968 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
07be1b83 2969 (int)depth * 2 + 2, "",
786e8c11 2970 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2971 );
7122b237 2972#endif
538e84ed 2973 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
2974 split out as an EXACT and put in front of the TRIE node. */
2975 trie->startstate= 1;
55eed653 2976 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2977 U32 state;
1e2e3d02 2978 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2979 U32 ofs = 0;
8e11feef
RGS
2980 I32 idx = -1;
2981 U32 count = 0;
2982 const U32 base = trie->states[ state ].trans.base;
a3621e74 2983
3dab1dad 2984 if ( trie->states[state].wordnum )
8e11feef 2985 count = 1;
a3621e74 2986
8e11feef 2987 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2988 if ( ( base + ofs >= trie->uniquecharcount ) &&
2989 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2990 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2991 {
3dab1dad 2992 if ( ++count > 1 ) {
2b8b4781 2993 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2994 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2995 if ( state == 1 ) break;
3dab1dad
YO
2996 if ( count == 2 ) {
2997 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2998 DEBUG_OPTIMISE_r(
8e11feef
RGS
2999 PerlIO_printf(Perl_debug_log,
3000 "%*sNew Start State=%"UVuf" Class: [",
3001 (int)depth * 2 + 2, "",
786e8c11 3002 (UV)state));
be8e71aa 3003 if (idx >= 0) {
2b8b4781 3004 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 3005 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3006
3dab1dad 3007 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
3008 if ( folder )
3009 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 3010 DEBUG_OPTIMISE_r(
f1f66076 3011 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 3012 );
8e11feef
RGS
3013 }
3014 }
3015 TRIE_BITMAP_SET(trie,*ch);
3016 if ( folder )
3017 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3018 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3019 }
3020 idx = ofs;
3021 }
3dab1dad
YO
3022 }
3023 if ( count == 1 ) {
2b8b4781 3024 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
3025 STRLEN len;
3026 char *ch = SvPV( *tmp, len );
de734bd5
A
3027 DEBUG_OPTIMISE_r({
3028 SV *sv=sv_newmortal();
8e11feef
RGS
3029 PerlIO_printf( Perl_debug_log,
3030 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3031 (int)depth * 2 + 2, "",
538e84ed
KW
3032 (UV)state, (UV)idx,
3033 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3034 PL_colors[0], PL_colors[1],
3035 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3036 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3037 )
3038 );
3039 });
3dab1dad
YO
3040 if ( state==1 ) {
3041 OP( convert ) = nodetype;
3042 str=STRING(convert);
3043 STR_LEN(convert)=0;
3044 }
c490c714
YO
3045 STR_LEN(convert) += len;
3046 while (len--)
de734bd5 3047 *str++ = *ch++;
8e11feef 3048 } else {
538e84ed 3049#ifdef DEBUGGING
8e11feef
RGS
3050 if (state>1)
3051 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 3052#endif
8e11feef
RGS
3053 break;
3054 }
3055 }
2e64971a 3056 trie->prefixlen = (state-1);
3dab1dad 3057 if (str) {
8e11feef 3058 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3059 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3060 trie->startstate = state;
07be1b83
YO
3061 trie->minlen -= (state - 1);
3062 trie->maxlen -= (state - 1);
33809eae
JH
3063#ifdef DEBUGGING
3064 /* At least the UNICOS C compiler choked on this
3065 * being argument to DEBUG_r(), so let's just have
3066 * it right here. */
3067 if (
3068#ifdef PERL_EXT_RE_BUILD
3069 1
3070#else
3071 DEBUG_r_TEST
3072#endif
3073 ) {
3074 regnode *fix = convert;
3075 U32 word = trie->wordcount;
3076 mjd_nodelen++;
3077 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3078 while( ++fix < n ) {
3079 Set_Node_Offset_Length(fix, 0, 0);
3080 }
3081 while (word--) {
3082 SV ** const tmp = av_fetch( trie_words, word, 0 );
3083 if (tmp) {
3084 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3085 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3086 else
3087 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3088 }
3089 }
3090 }
3091#endif
8e11feef
RGS
3092 if (trie->maxlen) {
3093 convert = n;
3094 } else {
3dab1dad 3095 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3096 DEBUG_r(optimize= n);
3dab1dad
YO
3097 }
3098 }
3099 }
538e84ed
KW
3100 if (!jumper)
3101 jumper = last;
3dab1dad 3102 if ( trie->maxlen ) {
8e11feef
RGS
3103 NEXT_OFF( convert ) = (U16)(tail - convert);
3104 ARG_SET( convert, data_slot );
538e84ed
KW
3105 /* Store the offset to the first unabsorbed branch in
3106 jump[0], which is otherwise unused by the jump logic.
786e8c11 3107 We use this when dumping a trie and during optimisation. */
538e84ed 3108 if (trie->jump)
7f69552c 3109 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3110
6c48061a
YO
3111 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3112 * and there is a bitmap
3113 * and the first "jump target" node we found leaves enough room
3114 * then convert the TRIE node into a TRIEC node, with the bitmap
3115 * embedded inline in the opcode - this is hypothetically faster.
3116 */
3117 if ( !trie->states[trie->startstate].wordnum
3118 && trie->bitmap
3119 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3120 {
3121 OP( convert ) = TRIEC;
3122 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3123 PerlMemShared_free(trie->bitmap);
786e8c11 3124 trie->bitmap= NULL;
538e84ed 3125 } else
786e8c11 3126 OP( convert ) = TRIE;
a3621e74 3127
3dab1dad
YO
3128 /* store the type in the flags */
3129 convert->flags = nodetype;
a5ca303d 3130 DEBUG_r({
538e84ed
KW
3131 optimize = convert
3132 + NODE_STEP_REGNODE
a5ca303d
YO
3133 + regarglen[ OP( convert ) ];
3134 });
538e84ed 3135 /* XXX We really should free up the resource in trie now,
a5ca303d 3136 as we won't use them - (which resources?) dmq */
3dab1dad 3137 }
a3621e74 3138 /* needed for dumping*/
e62cc96a 3139 DEBUG_r(if (optimize) {
07be1b83 3140 regnode *opt = convert;
bcdf7404 3141
e62cc96a 3142 while ( ++opt < optimize) {
07be1b83
YO
3143 Set_Node_Offset_Length(opt,0,0);
3144 }
538e84ed
KW
3145 /*
3146 Try to clean up some of the debris left after the
786e8c11 3147 optimisation.
a3621e74 3148 */
786e8c11 3149 while( optimize < jumper ) {
07be1b83 3150 mjd_nodelen += Node_Length((optimize));
a3621e74 3151 OP( optimize ) = OPTIMIZED;
07be1b83 3152 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3153 optimize++;
3154 }
07be1b83 3155 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3156 });
3157 } /* end node insert */
2e64971a
DM
3158
3159 /* Finish populating the prev field of the wordinfo array. Walk back
3160 * from each accept state until we find another accept state, and if
3161 * so, point the first word's .prev field at the second word. If the
3162 * second already has a .prev field set, stop now. This will be the
3163 * case either if we've already processed that word's accept state,
3b753521
FN
3164 * or that state had multiple words, and the overspill words were
3165 * already linked up earlier.
2e64971a
DM
3166 */
3167 {
3168 U16 word;
3169 U32 state;
3170 U16 prev;
3171
3172 for (word=1; word <= trie->wordcount; word++) {
3173 prev = 0;
3174 if (trie->wordinfo[word].prev)
3175 continue;
3176 state = trie->wordinfo[word].accept;
3177 while (state) {
3178 state = prev_states[state];
3179 if (!state)
3180 break;
3181 prev = trie->states[state].wordnum;
3182 if (prev)
3183 break;
3184 }
3185 trie->wordinfo[word].prev = prev;
3186 }
3187 Safefree(prev_states);
3188 }
3189
3190
3191 /* and now dump out the compressed format */
3192 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3193
55eed653 3194 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
3195#ifdef DEBUGGING
3196 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3197 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3198#else
03e70be4 3199 SvREFCNT_dec_NN(revcharmap);
07be1b83 3200#endif
538e84ed
KW
3201 return trie->jump
3202 ? MADE_JUMP_TRIE
3203 : trie->startstate>1
3204 ? MADE_EXACT_TRIE
786e8c11
YO
3205 : MADE_TRIE;
3206}
3207
615a2e7f
YO
3208STATIC regnode *
3209S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
786e8c11 3210{
b423522f
KW
3211/* The Trie is constructed and compressed now so we can build a fail array if
3212 * it's needed
786e8c11 3213
b423522f
KW
3214 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3215 3.32 in the
3216 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3217 Ullman 1985/88
786e8c11
YO
3218 ISBN 0-201-10088-6
3219
b423522f
KW
3220 We find the fail state for each state in the trie, this state is the longest
3221 proper suffix of the current state's 'word' that is also a proper prefix of
3222 another word in our trie. State 1 represents the word '' and is thus the
3223 default fail state. This allows the DFA not to have to restart after its
3224 tried and failed a word at a given point, it simply continues as though it
3225 had been matching the other word in the first place.
786e8c11
YO
3226 Consider
3227 'abcdgu'=~/abcdefg|cdgu/
b423522f
KW
3228 When we get to 'd' we are still matching the first word, we would encounter
3229 'g' which would fail, which would bring us to the state representing 'd' in
3230 the second word where we would try 'g' and succeed, proceeding to match
3231 'cdgu'.
786e8c11
YO
3232 */
3233 /* add a fail transition */
3251b653
NC
3234 const U32 trie_offset = ARG(source);
3235 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
3236 U32 *q;
3237 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 3238 const U32 numstates = trie->statecount;
786e8c11
YO
3239 const U32 ubound = trie->lasttrans + ucharcount;
3240 U32 q_read = 0;
3241 U32 q_write = 0;
3242 U32 charid;
3243 U32 base = trie->states[ 1 ].trans.base;
3244 U32 *fail;
3245 reg_ac_data *aho;
cf78de0b 3246 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
615a2e7f 3247 regnode *stclass;
786e8c11 3248 GET_RE_DEBUG_FLAGS_DECL;
7918f24d 3249
615a2e7f 3250 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
81611534 3251 PERL_UNUSED_CONTEXT;
786e8c11
YO
3252#ifndef DEBUGGING
3253 PERL_UNUSED_ARG(depth);
3254#endif
3255
615a2e7f
YO
3256 if ( OP(source) == TRIE ) {
3257 struct regnode_1 *op = (struct regnode_1 *)
3258 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3259 StructCopy(source,op,struct regnode_1);
3260 stclass = (regnode *)op;
3261 } else {
3262 struct regnode_charclass *op = (struct regnode_charclass *)
3263 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3264 StructCopy(source,op,struct regnode_charclass);
3265 stclass = (regnode *)op;
3266 }
2f306ab9 3267 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
786e8c11
YO
3268
3269 ARG_SET( stclass, data_slot );
c944940b 3270 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 3271 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 3272 aho->trie=trie_offset;
446bd890
NC
3273 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3274 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 3275 Newxz( q, numstates, U32);
c944940b 3276 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
3277 aho->refcount = 1;
3278 fail = aho->fail;
3279 /* initialize fail[0..1] to be 1 so that we always have
3280 a valid final fail state */
3281 fail[ 0 ] = fail[ 1 ] = 1;
3282
3283 for ( charid = 0; charid < ucharcount ; charid++ ) {
3284 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3285 if ( newstate ) {
3286 q[ q_write ] = newstate;
3287 /* set to point at the root */
3288 fail[ q[ q_write++ ] ]=1;
3289 }
3290 }
3291 while ( q_read < q_write) {
3292 const U32 cur = q[ q_read++ % numstates ];
3293 base = trie->states[ cur ].trans.base;
3294
3295 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3296 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3297 if (ch_state) {
3298 U32 fail_state = cur;
3299 U32 fail_base;
3300 do {
3301 fail_state = fail[ fail_state ];
3302 fail_base = aho->states[ fail_state ].trans.base;
3303 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3304
3305 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3306 fail[ ch_state ] = fail_state;
3307 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3308 {
3309 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3310 }
3311 q[ q_write++ % numstates] = ch_state;
3312 }
3313 }
3314 }
3315 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3316 when we fail in state 1, this allows us to use the
3317 charclass scan to find a valid start char. This is based on the principle
3318 that theres a good chance the string being searched contains lots of stuff
3319 that cant be a start char.
3320 */
3321 fail[ 0 ] = fail[ 1 ] = 0;
3322 DEBUG_TRIE_COMPILE_r({
6d99fb9b 3323 PerlIO_printf(Perl_debug_log,
538e84ed 3324 "%*sStclass Failtable (%"UVuf" states): 0",
6d99fb9b 3325 (int)(depth * 2), "", (UV)numstates
1e2e3d02 3326 );
786e8c11
YO
3327 for( q_read=1; q_read<numstates; q_read++ ) {
3328 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3329 }
3330 PerlIO_printf(Perl_debug_log, "\n");
3331 });
3332 Safefree(q);
e384d5c1 3333 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
615a2e7f 3334 return stclass;
a3621e74
YO
3335}
3336
786e8c11 3337
07be1b83 3338#define DEBUG_PEEP(str,scan,depth) \
b515a41d 3339 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83 3340 regnode *Next = regnext(scan); \
8b9781c9 3341 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
fdfb4f21 3342 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
c9f0d54c
YO
3343 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3344 Next ? (REG_NODE_NUM(Next)) : 0 ); \
fdfb4f21
YO
3345 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3346 PerlIO_printf(Perl_debug_log, "\n"); \
b515a41d 3347 }});
07be1b83 3348
bb914485 3349/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
3350 * one. The regop may be changed if the node(s) contain certain sequences that
3351 * require special handling. The joining is only done if:
bb914485
KW
3352 * 1) there is room in the current conglomerated node to entirely contain the
3353 * next one.
3354 * 2) they are the exact same node type
3355 *
87b8b349 3356 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
3357 * these get optimized out
3358 *
0a982f06
KW
3359 * If a node is to match under /i (folded), the number of characters it matches
3360 * can be different than its character length if it contains a multi-character
31f05a37
KW
3361 * fold. *min_subtract is set to the total delta number of characters of the
3362 * input nodes.
bb914485 3363 *
fb29cc72 3364 * And *unfolded_multi_char is set to indicate whether or not the node contains
31f05a37
KW
3365 * an unfolded multi-char fold. This happens when whether the fold is valid or
3366 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3367 * SMALL LETTER SHARP S, as only if the target string being matched against
3368 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3369 * folding rules depend on the locale in force at runtime. (Multi-char folds
3370 * whose components are all above the Latin1 range are not run-time locale
3371 * dependent, and have already been folded by the time this function is
3372 * called.)
f758bddf 3373 *
bb914485 3374 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
3375 * multi-character fold sequences. It's been wrong in Perl for a very long
3376 * time. There are three code points in Unicode whose multi-character folds
3377 * were long ago discovered to mess things up. The previous designs for
3378 * dealing with these involved assigning a special node for them. This
538e84ed 3379 * approach doesn't always work, as evidenced by this example:
a0c4c608 3380 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
538e84ed 3381 * Both sides fold to "sss", but if the pattern is parsed to create a node that
0a982f06 3382 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
3383 * successful match would have to cross the node's boundary. The new approach
3384 * that hopefully generally solves the problem generates an EXACTFU_SS node
538e84ed 3385 * that is "sss" in this case.
bb914485 3386 *
0a982f06 3387 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
3388 * just these three. Now the code is general, for all such cases. The
3389 * approach taken is:
0a982f06 3390 * 1) This routine examines each EXACTFish node that could contain multi-
538e84ed
KW
3391 * character folded sequences. Since a single character can fold into
3392 * such a sequence, the minimum match length for this node is less than
3393 * the number of characters in the node. This routine returns in
31f05a37
KW
3394 * *min_subtract how many characters to subtract from the the actual
3395 * length of the string to get a real minimum match length; it is 0 if
3396 * there are no multi-char foldeds. This delta is used by the caller to
3397 * adjust the min length of the match, and the delta between min and max,
3398 * so that the optimizer doesn't reject these possibilities based on size
3399 * constraints.
cb117658 3400 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
3401 * is used for an EXACTFU node that contains at least one "ss" sequence in
3402 * it. For non-UTF-8 patterns and strings, this is the only case where
3403 * there is a possible fold length change. That means that a regular
3404 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3405 * with length changes, and so can be processed faster. regexec.c takes
3406 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
31f05a37
KW
3407 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3408 * known until runtime). This saves effort in regex matching. However,
3409 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3410 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3411 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3412 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
0a982f06
KW
3413 * possibilities for the non-UTF8 patterns are quite simple, except for
3414 * the sharp s. All the ones that don't involve a UTF-8 target string are
3415 * members of a fold-pair, and arrays are set up for all of them so that
3416 * the other member of the pair can be found quickly. Code elsewhere in
3417 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3418 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3419 * described in the next item.
31f05a37
KW
3420 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3421 * validity of the fold won't be known until runtime, and so must remain
3422 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3423 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3424 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3425 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3426 * The reason this is a problem is that the optimizer part of regexec.c
3427 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3428 * that a character in the pattern corresponds to at most a single
3429 * character in the target string. (And I do mean character, and not byte
3430 * here, unlike other parts of the documentation that have never been
3431 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3432 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3433 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3434 * nodes, violate the assumption, and they are the only instances where it
3435 * is violated. I'm reluctant to try to change the assumption, as the
3436 * code involved is impenetrable to me (khw), so instead the code here
3437 * punts. This routine examines EXACTFL nodes, and (when the pattern
3438 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3439 * boolean indicating whether or not the node contains such a fold. When
3440 * it is true, the caller sets a flag that later causes the optimizer in
3441 * this file to not set values for the floating and fixed string lengths,
3442 * and thus avoids the optimizer code in regexec.c that makes the invalid
1ca267a5 3443 * assumption. Thus, there is no optimization based on string lengths for
31f05a37
KW
3444 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3445 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3446 * assumption is wrong only in these cases is that all other non-UTF-8
3447 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3448 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3449 * EXACTF nodes because we don't know at compile time if it actually
3450 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3451 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3452 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3453 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3454 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3455 * string would require the pattern to be forced into UTF-8, the overhead
3456 * of which we want to avoid. Similarly the unfolded multi-char folds in
3457 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3458 * locale.)
098b07d5
KW
3459 *
3460 * Similarly, the code that generates tries doesn't currently handle
3461 * not-already-folded multi-char folds, and it looks like a pain to change
3462 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3463 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3464 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3465 * using /iaa matching will be doing so almost entirely with ASCII
3466 * strings, so this should rarely be encountered in practice */
1de06328 3467
fb29cc72 3468#define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
07be1b83 3469 if (PL_regkind[OP(scan)] == EXACT) \
fb29cc72 3470 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
07be1b83 3471
be8e71aa 3472STATIC U32
538e84ed 3473S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
fb29cc72