This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
f
[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
09b2b2e6 217};
830247a4 218
e2509266 219#define RExC_flags (pRExC_state->flags)
514a91f1 220#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 221#define RExC_precomp (pRExC_state->precomp)
288b8c02 222#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 223#define RExC_rx (pRExC_state->rx)
f8fc2ecf 224#define RExC_rxi (pRExC_state->rxi)
fac92740 225#define RExC_start (pRExC_state->start)
830247a4
IZ
226#define RExC_end (pRExC_state->end)
227#define RExC_parse (pRExC_state->parse)
228#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237 229#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
230#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
231 others */
7122b237 232#endif
830247a4 233#define RExC_emit (pRExC_state->emit)
f7c7e32a 234#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 235#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 236#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
237#define RExC_sawback (pRExC_state->sawback)
238#define RExC_seen (pRExC_state->seen)
239#define RExC_size (pRExC_state->size)
ee273784 240#define RExC_maxlen (pRExC_state->maxlen)
830247a4 241#define RExC_npar (pRExC_state->npar)
e2e6a0f1 242#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
243#define RExC_extralen (pRExC_state->extralen)
244#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 245#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 246#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 247#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
248#define RExC_open_parens (pRExC_state->open_parens)
249#define RExC_close_parens (pRExC_state->close_parens)
250#define RExC_opend (pRExC_state->opend)
81714fb9 251#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
252#define RExC_recurse (pRExC_state->recurse)
253#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 254#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
255#define RExC_study_chunk_recursed_bytes \
256 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 257#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 258#define RExC_contains_locale (pRExC_state->contains_locale)
cfafade5 259#define RExC_contains_i (pRExC_state->contains_i)
9d53c457 260#define RExC_override_recoding (pRExC_state->override_recoding)
b6d67071
KW
261#ifdef EBCDIC
262# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
263#endif
9d53c457 264#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
265#define RExC_frame_head (pRExC_state->frame_head)
266#define RExC_frame_last (pRExC_state->frame_last)
267#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 268#define RExC_strict (pRExC_state->strict)
830247a4 269
99807a43
HS
270/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
271 * a flag to disable back-off on the fixed/floating substrings - if it's
272 * a high complexity pattern we assume the benefit of avoiding a full match
273 * is worth the cost of checking for the substrings even if they rarely help.
274 */
275#define RExC_naughty (pRExC_state->naughty)
276#define TOO_NAUGHTY (10)
277#define MARK_NAUGHTY(add) \
278 if (RExC_naughty < TOO_NAUGHTY) \
279 RExC_naughty += (add)
280#define MARK_NAUGHTY_EXP(exp, add) \
281 if (RExC_naughty < TOO_NAUGHTY) \
282 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 283
a687059c
LW
284#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
285#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 286 ((*s) == '{' && regcurly(s)))
a687059c
LW
287
288/*
289 * Flags to be passed up and down.
290 */
a687059c 291#define WORST 0 /* Worst case. */
a3b492c3 292#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 293
e64f369d 294/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
295 * character. (There needs to be a case: in the switch statement in regexec.c
296 * for any node marked SIMPLE.) Note that this is not the same thing as
297 * REGNODE_SIMPLE */
fda99bee 298#define SIMPLE 0x02
e64f369d 299#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
300#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
301#define TRYAGAIN 0x10 /* Weeded out a declaration. */
302#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 303
3dab1dad
YO
304#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
305
07be1b83
YO
306/* whether trie related optimizations are enabled */
307#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
308#define TRIE_STUDY_OPT
786e8c11 309#define FULL_TRIE_STUDY
07be1b83
YO
310#define TRIE_STCLASS
311#endif
1de06328
YO
312
313
40d049e4
YO
314
315#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
316#define PBITVAL(paren) (1 << ((paren) & 7))
317#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
318#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
319#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
320
bbd61b5f 321#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
322 if (!UTF) { \
323 *flagp = RESTART_UTF8; \
324 return NULL; \
325 } \
bbd61b5f 326 } STMT_END
40d049e4 327
f19b1a63
KW
328/* This converts the named class defined in regcomp.h to its equivalent class
329 * number defined in handy.h. */
330#define namedclass_to_classnum(class) ((int) ((class) / 2))
331#define classnum_to_namedclass(classnum) ((classnum) * 2)
332
de92f5e6
KW
333#define _invlist_union_complement_2nd(a, b, output) \
334 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
335#define _invlist_intersection_complement_2nd(a, b, output) \
336 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
337
1de06328
YO
338/* About scan_data_t.
339
340 During optimisation we recurse through the regexp program performing
341 various inplace (keyhole style) optimisations. In addition study_chunk
342 and scan_commit populate this data structure with information about
538e84ed 343 what strings MUST appear in the pattern. We look for the longest
3b753521 344 string that must appear at a fixed location, and we look for the
1de06328
YO
345 longest string that may appear at a floating location. So for instance
346 in the pattern:
538e84ed 347
1de06328 348 /FOO[xX]A.*B[xX]BAR/
538e84ed 349
1de06328
YO
350 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
351 strings (because they follow a .* construct). study_chunk will identify
352 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 353
1de06328 354 The strings can be composites, for instance
538e84ed 355
1de06328 356 /(f)(o)(o)/
538e84ed 357
1de06328 358 will result in a composite fixed substring 'foo'.
538e84ed 359
1de06328 360 For each string some basic information is maintained:
538e84ed 361
1de06328
YO
362 - offset or min_offset
363 This is the position the string must appear at, or not before.
364 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
365 characters must match before the string we are searching for.
366 Likewise when combined with minlenp and the length of the string it
538e84ed 367 tells us how many characters must appear after the string we have
1de06328 368 found.
538e84ed 369
1de06328
YO
370 - max_offset
371 Only used for floating strings. This is the rightmost point that
ea3daa5d 372 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 373 string can occur infinitely far to the right.
538e84ed 374
1de06328 375 - minlenp
2d608413
KW
376 A pointer to the minimum number of characters of the pattern that the
377 string was found inside. This is important as in the case of positive
538e84ed 378 lookahead or positive lookbehind we can have multiple patterns
1de06328 379 involved. Consider
538e84ed 380
1de06328 381 /(?=FOO).*F/
538e84ed 382
1de06328
YO
383 The minimum length of the pattern overall is 3, the minimum length
384 of the lookahead part is 3, but the minimum length of the part that
538e84ed 385 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 386 minimum length for the F is 1. This is important as the minimum length
538e84ed 387 is used to determine offsets in front of and behind the string being
1de06328 388 looked for. Since strings can be composites this is the length of the
486ec47a 389 pattern at the time it was committed with a scan_commit. Note that
1de06328 390 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 391 are not known until the full pattern has been compiled, thus the
1de06328 392 pointer to the value.
538e84ed 393
1de06328 394 - lookbehind
538e84ed 395
1de06328 396 In the case of lookbehind the string being searched for can be
538e84ed 397 offset past the start point of the final matching string.
1de06328
YO
398 If this value was just blithely removed from the min_offset it would
399 invalidate some of the calculations for how many chars must match
400 before or after (as they are derived from min_offset and minlen and
538e84ed 401 the length of the string being searched for).
1de06328
YO
402 When the final pattern is compiled and the data is moved from the
403 scan_data_t structure into the regexp structure the information
538e84ed
KW
404 about lookbehind is factored in, with the information that would
405 have been lost precalculated in the end_shift field for the
1de06328
YO
406 associated string.
407
408 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 409 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
410
411*/
2c2d71f5
JH
412
413typedef struct scan_data_t {
1de06328
YO
414 /*I32 len_min; unused */
415 /*I32 len_delta; unused */
49f55535 416 SSize_t pos_min;
ea3daa5d 417 SSize_t pos_delta;
2c2d71f5 418 SV *last_found;
ea3daa5d 419 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 420 SSize_t last_start_min;
ea3daa5d 421 SSize_t last_start_max;
1de06328
YO
422 SV **longest; /* Either &l_fixed, or &l_float. */
423 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 424 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 425 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
426 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
427 SV *longest_float; /* longest floating string found in pattern */
49f55535 428 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
429 SSize_t offset_float_max; /* latest point in string it can appear */
430 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 431 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
432 I32 flags;
433 I32 whilem_c;
49f55535 434 SSize_t *last_closep;
b8f7bb16 435 regnode_ssc *start_class;
2c2d71f5
JH
436} scan_data_t;
437
a687059c 438/*
e50aee73 439 * Forward declarations for pregcomp()'s friends.
a687059c 440 */
a0d0e21e 441
27da23d5 442static const scan_data_t zero_scan_data =
1de06328 443 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
444
445#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
446#define SF_BEFORE_SEOL 0x0001
447#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
448#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
449#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
450
44e3dfd2
BF
451#define SF_FIX_SHIFT_EOL (+2)
452#define SF_FL_SHIFT_EOL (+4)
c277df42
IZ
453
454#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
455#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
456
457#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
458#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
459#define SF_IS_INF 0x0040
460#define SF_HAS_PAR 0x0080
461#define SF_IN_PAR 0x0100
462#define SF_HAS_EVAL 0x0200
463#define SCF_DO_SUBSTR 0x0400
653099ff
GS
464#define SCF_DO_STCLASS_AND 0x0800
465#define SCF_DO_STCLASS_OR 0x1000
466#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 467#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 468
786e8c11 469#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 470#define SCF_SEEN_ACCEPT 0x8000
688e0391 471#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
472#define SCF_IN_DEFINE 0x20000
473
474
475
07be1b83 476
43fead97 477#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
478
479/* The enums for all these are ordered so things work out correctly */
a62b1201 480#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
481#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
482 == REGEX_DEPENDS_CHARSET)
00b27cfc 483#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
484#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
485 >= REGEX_UNICODE_CHARSET)
486#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
487 == REGEX_ASCII_RESTRICTED_CHARSET)
488#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
489 >= REGEX_ASCII_RESTRICTED_CHARSET)
490#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
491 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 492
43fead97 493#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 494
f2c2a6ab
KW
495/* For programs that want to be strictly Unicode compatible by dying if any
496 * attempt is made to match a non-Unicode code point against a Unicode
497 * property. */
498#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
499
93733859 500#define OOB_NAMEDCLASS -1
b8c5462f 501
8e661ac5
KW
502/* There is no code point that is out-of-bounds, so this is problematic. But
503 * its only current use is to initialize a variable that is always set before
504 * looked at. */
505#define OOB_UNICODE 0xDEADBEEF
506
a0ed51b3
LW
507#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
508#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
509
8615cb43 510
b45f050a
JF
511/* length of regex to show in messages that don't mark a position within */
512#define RegexLengthToShowInErrorMessages 127
513
514/*
515 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
516 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
517 * op/pragma/warn/regcomp.
518 */
7253e4e3
RK
519#define MARKER1 "<-- HERE" /* marker as it appears in the description */
520#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 521
538e84ed
KW
522#define REPORT_LOCATION " in regex; marked by " MARKER1 \
523 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
b45f050a 524
c1d900c3
BF
525#define REPORT_LOCATION_ARGS(offset) \
526 UTF8fARG(UTF, offset, RExC_precomp), \
527 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
528
8a6d8ec6
HS
529/* Used to point after bad bytes for an error message, but avoid skipping
530 * past a nul byte. */
531#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
532
b45f050a
JF
533/*
534 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
535 * arg. Show regex, up to a maximum length. If it's too long, chop and add
536 * "...".
537 */
58e23c8d 538#define _FAIL(code) STMT_START { \
bfed75c6 539 const char *ellipses = ""; \
ccb2c380
MP
540 IV len = RExC_end - RExC_precomp; \
541 \
542 if (!SIZE_ONLY) \
a5e7bc51 543 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
544 if (len > RegexLengthToShowInErrorMessages) { \
545 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
546 len = RegexLengthToShowInErrorMessages - 10; \
547 ellipses = "..."; \
548 } \
58e23c8d 549 code; \
ccb2c380 550} STMT_END
8615cb43 551
58e23c8d 552#define FAIL(msg) _FAIL( \
c1d900c3
BF
553 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
554 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
555
556#define FAIL2(msg,arg) _FAIL( \
c1d900c3
BF
557 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
558 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 559
b45f050a 560/*
b45f050a
JF
561 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
562 */
ccb2c380 563#define Simple_vFAIL(m) STMT_START { \
8fb0127d
YO
564 const IV offset = \
565 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
ccb2c380 566 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
c1d900c3 567 m, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 568} STMT_END
b45f050a
JF
569
570/*
571 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
572 */
ccb2c380
MP
573#define vFAIL(m) STMT_START { \
574 if (!SIZE_ONLY) \
a5e7bc51 575 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
576 Simple_vFAIL(m); \
577} STMT_END
b45f050a
JF
578
579/*
580 * Like Simple_vFAIL(), but accepts two arguments.
581 */
ccb2c380 582#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 583 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
584 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
585 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 586} STMT_END
b45f050a
JF
587
588/*
589 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
590 */
ccb2c380
MP
591#define vFAIL2(m,a1) STMT_START { \
592 if (!SIZE_ONLY) \
a5e7bc51 593 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
594 Simple_vFAIL2(m, a1); \
595} STMT_END
b45f050a
JF
596
597
598/*
599 * Like Simple_vFAIL(), but accepts three arguments.
600 */
ccb2c380 601#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 602 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
603 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
604 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 605} STMT_END
b45f050a
JF
606
607/*
608 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
609 */
ccb2c380
MP
610#define vFAIL3(m,a1,a2) STMT_START { \
611 if (!SIZE_ONLY) \
a5e7bc51 612 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
613 Simple_vFAIL3(m, a1, a2); \
614} STMT_END
b45f050a
JF
615
616/*
617 * Like Simple_vFAIL(), but accepts four arguments.
618 */
ccb2c380 619#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 620 const IV offset = RExC_parse - RExC_precomp; \
c1d900c3
BF
621 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
622 REPORT_LOCATION_ARGS(offset)); \
ccb2c380 623} STMT_END
b45f050a 624
95db3ffa
KW
625#define vFAIL4(m,a1,a2,a3) STMT_START { \
626 if (!SIZE_ONLY) \
627 SAVEFREESV(RExC_rx_sv); \
628 Simple_vFAIL4(m, a1, a2, a3); \
629} STMT_END
630
946095af
BF
631/* A specialized version of vFAIL2 that works with UTF8f */
632#define vFAIL2utf8f(m, a1) STMT_START { \
ef3f731d
BF
633 const IV offset = RExC_parse - RExC_precomp; \
634 if (!SIZE_ONLY) \
635 SAVEFREESV(RExC_rx_sv); \
636 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
637 REPORT_LOCATION_ARGS(offset)); \
946095af
BF
638} STMT_END
639
499333dc
KW
640/* These have asserts in them because of [perl #122671] Many warnings in
641 * regcomp.c can occur twice. If they get output in pass1 and later in that
642 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
643 * would get output again. So they should be output in pass2, and these
644 * asserts make sure new warnings follow that paradigm. */
946095af 645
5e0a247b
KW
646/* m is not necessarily a "literal string", in this macro */
647#define reg_warn_non_literal_string(loc, m) STMT_START { \
648 const IV offset = loc - RExC_precomp; \
499333dc 649 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
c1d900c3 650 m, REPORT_LOCATION_ARGS(offset)); \
5e0a247b
KW
651} STMT_END
652
668c081a 653#define ckWARNreg(loc,m) STMT_START { \
a28509cc 654 const IV offset = loc - RExC_precomp; \
499333dc 655 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 656 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
657} STMT_END
658
b927b7e9
KW
659#define vWARN(loc, m) STMT_START { \
660 const IV offset = loc - RExC_precomp; \
661 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
662 REPORT_LOCATION_ARGS(offset)); \
663} STMT_END
664
0d6106aa
KW
665#define vWARN_dep(loc, m) STMT_START { \
666 const IV offset = loc - RExC_precomp; \
499333dc 667 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
c1d900c3 668 REPORT_LOCATION_ARGS(offset)); \
0d6106aa
KW
669} STMT_END
670
147508a2
KW
671#define ckWARNdep(loc,m) STMT_START { \
672 const IV offset = loc - RExC_precomp; \
499333dc 673 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
147508a2 674 m REPORT_LOCATION, \
c1d900c3 675 REPORT_LOCATION_ARGS(offset)); \
147508a2
KW
676} STMT_END
677
668c081a 678#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 679 const IV offset = loc - RExC_precomp; \
499333dc 680 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18 681 m REPORT_LOCATION, \
c1d900c3 682 REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
683} STMT_END
684
b23eb183 685#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 686 const IV offset = loc - RExC_precomp; \
499333dc 687 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3 688 m REPORT_LOCATION, \
c1d900c3 689 a1, REPORT_LOCATION_ARGS(offset)); \
2335b3d3
KW
690} STMT_END
691
668c081a 692#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 693 const IV offset = loc - RExC_precomp; \
499333dc 694 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 695 a1, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
696} STMT_END
697
698#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 699 const IV offset = loc - RExC_precomp; \
499333dc 700 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 701 a1, a2, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
702} STMT_END
703
668c081a
NC
704#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
705 const IV offset = loc - RExC_precomp; \
499333dc 706 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 707 a1, a2, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
708} STMT_END
709
ccb2c380 710#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 711 const IV offset = loc - RExC_precomp; \
499333dc 712 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 713 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
ccb2c380
MP
714} STMT_END
715
668c081a
NC
716#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
717 const IV offset = loc - RExC_precomp; \
499333dc 718 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 719 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
668c081a
NC
720} STMT_END
721
ccb2c380 722#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 723 const IV offset = loc - RExC_precomp; \
499333dc 724 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
c1d900c3 725 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
ccb2c380 726} STMT_END
9d1d55b5 727
538e84ed 728/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
729 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
730 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
731 * Element 0 holds the number n.
07be1b83 732 * Position is 1 indexed.
fac92740 733 */
7122b237
YO
734#ifndef RE_TRACK_PATTERN_OFFSETS
735#define Set_Node_Offset_To_R(node,byte)
736#define Set_Node_Offset(node,byte)
737#define Set_Cur_Node_Offset
738#define Set_Node_Length_To_R(node,len)
739#define Set_Node_Length(node,len)
6a86c6ad 740#define Set_Node_Cur_Length(node,start)
538e84ed
KW
741#define Node_Offset(n)
742#define Node_Length(n)
7122b237
YO
743#define Set_Node_Offset_Length(node,offset,len)
744#define ProgLen(ri) ri->u.proglen
745#define SetProgLen(ri,x) ri->u.proglen = x
746#else
747#define ProgLen(ri) ri->u.offsets[0]
748#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
749#define Set_Node_Offset_To_R(node,byte) STMT_START { \
750 if (! SIZE_ONLY) { \
751 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 752 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 753 if((node) < 0) { \
538e84ed
KW
754 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
755 (int)(node)); \
ccb2c380
MP
756 } else { \
757 RExC_offsets[2*(node)-1] = (byte); \
758 } \
759 } \
760} STMT_END
761
762#define Set_Node_Offset(node,byte) \
763 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
764#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
765
766#define Set_Node_Length_To_R(node,len) STMT_START { \
767 if (! SIZE_ONLY) { \
768 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 769 __LINE__, (int)(node), (int)(len))); \
ccb2c380 770 if((node) < 0) { \
538e84ed
KW
771 Perl_croak(aTHX_ "value of node is %d in Length macro", \
772 (int)(node)); \
ccb2c380
MP
773 } else { \
774 RExC_offsets[2*(node)] = (len); \
775 } \
776 } \
777} STMT_END
778
779#define Set_Node_Length(node,len) \
780 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
781#define Set_Node_Cur_Length(node, start) \
782 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
783
784/* Get offsets and lengths */
785#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
786#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
787
07be1b83
YO
788#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
789 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
790 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
791} STMT_END
7122b237 792#endif
07be1b83
YO
793
794#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
795#define EXPERIMENTAL_INPLACESCAN
f427392e 796#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 797
9e9ecfdf 798#define DEBUG_RExC_seen() \
538e84ed
KW
799 DEBUG_OPTIMISE_MORE_r({ \
800 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
801 \
e384d5c1
YO
802 if (RExC_seen & REG_ZERO_LEN_SEEN) \
803 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
538e84ed 804 \
e384d5c1
YO
805 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
806 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
538e84ed 807 \
e384d5c1
YO
808 if (RExC_seen & REG_GPOS_SEEN) \
809 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
538e84ed 810 \
e384d5c1
YO
811 if (RExC_seen & REG_RECURSE_SEEN) \
812 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
538e84ed 813 \
e384d5c1
YO
814 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
815 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 816 \
e384d5c1
YO
817 if (RExC_seen & REG_VERBARG_SEEN) \
818 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
538e84ed 819 \
e384d5c1
YO
820 if (RExC_seen & REG_CUTGROUP_SEEN) \
821 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
538e84ed 822 \
e384d5c1
YO
823 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
824 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 825 \
e384d5c1
YO
826 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
827 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 828 \
e384d5c1
YO
829 if (RExC_seen & REG_GOSTART_SEEN) \
830 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
538e84ed 831 \
ee273784
YO
832 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
833 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
834 \
538e84ed 835 PerlIO_printf(Perl_debug_log,"\n"); \
9e9ecfdf
YO
836 });
837
fdfb4f21
YO
838#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
839 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
840
841#define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
842 if ( ( flags ) ) { \
843 PerlIO_printf(Perl_debug_log, "%s", open_str); \
844 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
845 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
846 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
847 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
848 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
849 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
850 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
851 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
852 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
853 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
854 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
855 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
856 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
857 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
858 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
859 PerlIO_printf(Perl_debug_log, "%s", close_str); \
860 }
861
862
304ee84b
YO
863#define DEBUG_STUDYDATA(str,data,depth) \
864DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 865 PerlIO_printf(Perl_debug_log, \
304ee84b 866 "%*s" str "Pos:%"IVdf"/%"IVdf \
fdfb4f21 867 " Flags: 0x%"UVXf, \
1de06328
YO
868 (int)(depth)*2, "", \
869 (IV)((data)->pos_min), \
870 (IV)((data)->pos_delta), \
fdfb4f21
YO
871 (UV)((data)->flags) \
872 ); \
873 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
874 PerlIO_printf(Perl_debug_log, \
875 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328 876 (IV)((data)->whilem_c), \
304ee84b
YO
877 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
878 is_inf ? "INF " : "" \
1de06328
YO
879 ); \
880 if ((data)->last_found) \
881 PerlIO_printf(Perl_debug_log, \
882 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
883 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
884 SvPVX_const((data)->last_found), \
885 (IV)((data)->last_end), \
886 (IV)((data)->last_start_min), \
887 (IV)((data)->last_start_max), \
888 ((data)->longest && \
889 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
890 SvPVX_const((data)->longest_fixed), \
891 (IV)((data)->offset_fixed), \
892 ((data)->longest && \
893 (data)->longest==&((data)->longest_float)) ? "*" : "", \
894 SvPVX_const((data)->longest_float), \
895 (IV)((data)->offset_float_min), \
896 (IV)((data)->offset_float_max) \
897 ); \
898 PerlIO_printf(Perl_debug_log,"\n"); \
899});
900
8e35b056
KW
901/* is c a control character for which we have a mnemonic? */
902#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
903
549b4e78
KW
904STATIC const char *
905S_cntrl_to_mnemonic(const U8 c)
906{
907 /* Returns the mnemonic string that represents character 'c', if one
908 * exists; NULL otherwise. The only ones that exist for the purposes of
909 * this routine are a few control characters */
910
911 switch (c) {
912 case '\a': return "\\a";
913 case '\b': return "\\b";
914 case ESC_NATIVE: return "\\e";
915 case '\f': return "\\f";
916 case '\n': return "\\n";
917 case '\r': return "\\r";
918 case '\t': return "\\t";
919 }
920
921 return NULL;
922}
923
653099ff 924/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 925 Update the longest found anchored substring and the longest found
653099ff
GS
926 floating substrings if needed. */
927
4327152a 928STATIC void
ea3daa5d
FC
929S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
930 SSize_t *minlenp, int is_inf)
c277df42 931{
e1ec3a88
AL
932 const STRLEN l = CHR_SVLEN(data->last_found);
933 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 934 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 935
7918f24d
NC
936 PERL_ARGS_ASSERT_SCAN_COMMIT;
937
c277df42 938 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 939 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
940 if (*data->longest == data->longest_fixed) {
941 data->offset_fixed = l ? data->last_start_min : data->pos_min;
942 if (data->flags & SF_BEFORE_EOL)
b81d288d 943 data->flags
c277df42
IZ
944 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
945 else
946 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 947 data->minlen_fixed=minlenp;
1de06328 948 data->lookbehind_fixed=0;
a0ed51b3 949 }
304ee84b 950 else { /* *data->longest == data->longest_float */
c277df42 951 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d 952 data->offset_float_max = (l
646e8787
DM
953 ? data->last_start_max
954 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
955 ? SSize_t_MAX
956 : data->pos_min + data->pos_delta));
957 if (is_inf
958 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
959 data->offset_float_max = SSize_t_MAX;
c277df42 960 if (data->flags & SF_BEFORE_EOL)
b81d288d 961 data->flags
c277df42
IZ
962 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
963 else
964 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
965 data->minlen_float=minlenp;
966 data->lookbehind_float=0;
c277df42
IZ
967 }
968 }
969 SvCUR_set(data->last_found, 0);
0eda9292 970 {
a28509cc 971 SV * const sv = data->last_found;
097eb12c
AL
972 if (SvUTF8(sv) && SvMAGICAL(sv)) {
973 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
974 if (mg)
975 mg->mg_len = 0;
976 }
0eda9292 977 }
c277df42
IZ
978 data->last_end = -1;
979 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 980 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
981}
982
cdd87c1d
KW
983/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
984 * list that describes which code points it matches */
985
653099ff 986STATIC void
3420edd7 987S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 988{
cdd87c1d
KW
989 /* Set the SSC 'ssc' to match an empty string or any code point */
990
557bd3fb 991 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 992
71068078 993 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 994
cdd87c1d
KW
995 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
996 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
93e92956 997 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
998}
999
653099ff 1000STATIC int
dc3bf405 1001S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1002{
c144baaa
KW
1003 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1004 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1005 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1006 * in any way, so there's no point in using it */
cdd87c1d
KW
1007
1008 UV start, end;
1009 bool ret;
653099ff 1010
557bd3fb 1011 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1012
71068078 1013 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1014
93e92956 1015 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1016 return FALSE;
1017 }
1018
1019 /* See if the list consists solely of the range 0 - Infinity */
1020 invlist_iterinit(ssc->invlist);
1021 ret = invlist_iternext(ssc->invlist, &start, &end)
1022 && start == 0
1023 && end == UV_MAX;
1024
1025 invlist_iterfinish(ssc->invlist);
1026
1027 if (ret) {
1028 return TRUE;
1029 }
1030
1031 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1032 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1033 int i;
1034 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1035 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1036 return TRUE;
1037 }
1038 }
1039 }
1040
1041 return FALSE;
653099ff
GS
1042}
1043
653099ff 1044STATIC void
cdd87c1d 1045S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1046{
cdd87c1d
KW
1047 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1048 * string, any code point, or any posix class under locale */
1049
557bd3fb 1050 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1051
557bd3fb 1052 Zero(ssc, 1, regnode_ssc);
71068078 1053 set_ANYOF_SYNTHETIC(ssc);
93e92956 1054 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1055 ssc_anything(ssc);
cdd87c1d 1056
2f306ab9
KW
1057 /* If any portion of the regex is to operate under locale rules that aren't
1058 * fully known at compile time, initialization includes it. The reason
1059 * this isn't done for all regexes is that the optimizer was written under
1060 * the assumption that locale was all-or-nothing. Given the complexity and
1061 * lack of documentation in the optimizer, and that there are inadequate
1062 * test cases for locale, many parts of it may not work properly, it is
1063 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1064 if (RExC_contains_locale) {
1065 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1066 }
1067 else {
1068 ANYOF_POSIXL_ZERO(ssc);
1069 }
653099ff
GS
1070}
1071
b423522f 1072STATIC int
dc3bf405
BF
1073S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1074 const regnode_ssc *ssc)
b423522f
KW
1075{
1076 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1077 * to the list of code points matched, and locale posix classes; hence does
1078 * not check its flags) */
1079
1080 UV start, end;
1081 bool ret;
1082
1083 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1084
71068078 1085 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1086
1087 invlist_iterinit(ssc->invlist);
1088 ret = invlist_iternext(ssc->invlist, &start, &end)
1089 && start == 0
1090 && end == UV_MAX;
1091
1092 invlist_iterfinish(ssc->invlist);
1093
1094 if (! ret) {
1095 return FALSE;
1096 }
1097
e0e1be5f 1098 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1099 return FALSE;
b423522f
KW
1100 }
1101
1102 return TRUE;
1103}
1104
1105STATIC SV*
1106S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1107 const regnode_charclass* const node)
b423522f
KW
1108{
1109 /* Returns a mortal inversion list defining which code points are matched
1110 * by 'node', which is of type ANYOF. Handles complementing the result if
1111 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1112 * returned list must, and will, contain every code point that is a
1113 * possibility. */
b423522f
KW
1114
1115 SV* invlist = sv_2mortal(_new_invlist(0));
1ee208c4 1116 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1117 unsigned int i;
1118 const U32 n = ARG(node);
31f05a37 1119 bool new_node_has_latin1 = FALSE;
b423522f
KW
1120
1121 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1122
1123 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1124 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1125 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1126 AV * const av = MUTABLE_AV(SvRV(rv));
1127 SV **const ary = AvARRAY(av);
1128 assert(RExC_rxi->data->what[n] == 's');
1129
1130 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1131 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1132 }
1133 else if (ary[0] && ary[0] != &PL_sv_undef) {
1134
1135 /* Here, no compile-time swash, and there are things that won't be
1136 * known until runtime -- we have to assume it could be anything */
1137 return _add_range_to_invlist(invlist, 0, UV_MAX);
1138 }
1ee208c4 1139 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1140
1141 /* Here no compile-time swash, and no run-time only data. Use the
1142 * node's inversion list */
1ee208c4
KW
1143 invlist = sv_2mortal(invlist_clone(ary[3]));
1144 }
1145
1146 /* Get the code points valid only under UTF-8 locales */
1147 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1148 && ary[2] && ary[2] != &PL_sv_undef)
1149 {
1150 only_utf8_locale_invlist = ary[2];
b423522f
KW
1151 }
1152 }
1153
dcb20b36
KW
1154 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1155 * code points, and an inversion list for the others, but if there are code
1156 * points that should match only conditionally on the target string being
1157 * UTF-8, those are placed in the inversion list, and not the bitmap.
1158 * Since there are circumstances under which they could match, they are
1159 * included in the SSC. But if the ANYOF node is to be inverted, we have
1160 * to exclude them here, so that when we invert below, the end result
1161 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1162 * have to do this here before we add the unconditionally matched code
1163 * points */
b423522f
KW
1164 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1165 _invlist_intersection_complement_2nd(invlist,
1166 PL_UpperLatin1,
1167 &invlist);
1168 }
1169
1170 /* Add in the points from the bit map */
dcb20b36 1171 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f
KW
1172 if (ANYOF_BITMAP_TEST(node, i)) {
1173 invlist = add_cp_to_invlist(invlist, i);
31f05a37 1174 new_node_has_latin1 = TRUE;
b423522f
KW
1175 }
1176 }
1177
1178 /* If this can match all upper Latin1 code points, have to add them
1179 * as well */
93e92956 1180 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
b423522f
KW
1181 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1182 }
1183
1184 /* Similarly for these */
93e92956 1185 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1186 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1187 }
1188
1189 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1190 _invlist_invert(invlist);
1191 }
31f05a37
KW
1192 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1193
1194 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1195 * locale. We can skip this if there are no 0-255 at all. */
1196 _invlist_union(invlist, PL_Latin1, &invlist);
1197 }
1198
1ee208c4
KW
1199 /* Similarly add the UTF-8 locale possible matches. These have to be
1200 * deferred until after the non-UTF-8 locale ones are taken care of just
1201 * above, or it leads to wrong results under ANYOF_INVERT */
1202 if (only_utf8_locale_invlist) {
31f05a37 1203 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1204 only_utf8_locale_invlist,
31f05a37
KW
1205 ANYOF_FLAGS(node) & ANYOF_INVERT,
1206 &invlist);
1207 }
b423522f
KW
1208
1209 return invlist;
1210}
1211
1051e1c4 1212/* These two functions currently do the exact same thing */
557bd3fb 1213#define ssc_init_zero ssc_init
653099ff 1214
cdd87c1d
KW
1215#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1216#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1217
557bd3fb 1218/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1219 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1220 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1221
653099ff 1222STATIC void
b423522f 1223S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1224 const regnode_charclass *and_with)
653099ff 1225{
cdd87c1d
KW
1226 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1227 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1228
a0dd4231
KW
1229 SV* anded_cp_list;
1230 U8 anded_flags;
1e6ade67 1231
cdd87c1d 1232 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1233
71068078 1234 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1235
1236 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1237 * the code point inversion list and just the relevant flags */
71068078 1238 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1239 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1240 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1241
1242 /* XXX This is a kludge around what appears to be deficiencies in the
1243 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1244 * there are paths through the optimizer where it doesn't get weeded
1245 * out when it should. And if we don't make some extra provision for
1246 * it like the code just below, it doesn't get added when it should.
1247 * This solution is to add it only when AND'ing, which is here, and
1248 * only when what is being AND'ed is the pristine, original node
1249 * matching anything. Thus it is like adding it to ssc_anything() but
1250 * only when the result is to be AND'ed. Probably the same solution
1251 * could be adopted for the same problem we have with /l matching,
1252 * which is solved differently in S_ssc_init(), and that would lead to
1253 * fewer false positives than that solution has. But if this solution
1254 * creates bugs, the consequences are only that a warning isn't raised
1255 * that should be; while the consequences for having /l bugs is
1256 * incorrect matches */
7dcac5f6 1257 if (ssc_is_anything((regnode_ssc *)and_with)) {
e9b08962
KW
1258 anded_flags |= ANYOF_WARN_SUPER;
1259 }
a0dd4231
KW
1260 }
1261 else {
5c0f85ef 1262 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
eff8b7dc 1263 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1264 }
1265
1266 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1267
1268 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1269 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1270 * 'and_with' may be inverted. When not inverted, we have the situation of
1271 * computing:
1272 * (C1 | P1) & (C2 | P2)
1273 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1274 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1275 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1276 * <= ((C1 & C2) | P1 | P2)
1277 * Alternatively, the last few steps could be:
1278 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1279 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1280 * <= (C1 | C2 | (P1 & P2))
1281 * We favor the second approach if either P1 or P2 is non-empty. This is
1282 * because these components are a barrier to doing optimizations, as what
1283 * they match cannot be known until the moment of matching as they are
1284 * dependent on the current locale, 'AND"ing them likely will reduce or
1285 * eliminate them.
1286 * But we can do better if we know that C1,P1 are in their initial state (a
1287 * frequent occurrence), each matching everything:
1288 * (<everything>) & (C2 | P2) = C2 | P2
1289 * Similarly, if C2,P2 are in their initial state (again a frequent
1290 * occurrence), the result is a no-op
1291 * (C1 | P1) & (<everything>) = C1 | P1
1292 *
1293 * Inverted, we have
1294 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1295 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1296 * <= (C1 & ~C2) | (P1 & ~P2)
1297 * */
1aa99e6b 1298
a0dd4231 1299 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1300 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1301 {
cdd87c1d 1302 unsigned int i;
8951c461 1303
cdd87c1d
KW
1304 ssc_intersection(ssc,
1305 anded_cp_list,
1306 FALSE /* Has already been inverted */
1307 );
c6b76537 1308
cdd87c1d
KW
1309 /* If either P1 or P2 is empty, the intersection will be also; can skip
1310 * the loop */
93e92956 1311 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1312 ANYOF_POSIXL_ZERO(ssc);
1313 }
e0e1be5f 1314 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1315
1316 /* Note that the Posix class component P from 'and_with' actually
1317 * looks like:
1318 * P = Pa | Pb | ... | Pn
1319 * where each component is one posix class, such as in [\w\s].
1320 * Thus
1321 * ~P = ~(Pa | Pb | ... | Pn)
1322 * = ~Pa & ~Pb & ... & ~Pn
1323 * <= ~Pa | ~Pb | ... | ~Pn
1324 * The last is something we can easily calculate, but unfortunately
1325 * is likely to have many false positives. We could do better
1326 * in some (but certainly not all) instances if two classes in
1327 * P have known relationships. For example
1328 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1329 * So
1330 * :lower: & :print: = :lower:
1331 * And similarly for classes that must be disjoint. For example,
1332 * since \s and \w can have no elements in common based on rules in
1333 * the POSIX standard,
1334 * \w & ^\S = nothing
1335 * Unfortunately, some vendor locales do not meet the Posix
1336 * standard, in particular almost everything by Microsoft.
1337 * The loop below just changes e.g., \w into \W and vice versa */
1338
1ee208c4 1339 regnode_charclass_posixl temp;
cdd87c1d
KW
1340 int add = 1; /* To calculate the index of the complement */
1341
1342 ANYOF_POSIXL_ZERO(&temp);
1343 for (i = 0; i < ANYOF_MAX; i++) {
1344 assert(i % 2 != 0
7dcac5f6
KW
1345 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1346 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1347
7dcac5f6 1348 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1349 ANYOF_POSIXL_SET(&temp, i + add);
1350 }
1351 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1352 }
1353 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1354
cdd87c1d
KW
1355 } /* else ssc already has no posixes */
1356 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1357 in its initial state */
71068078 1358 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1359 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1360 {
1361 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1362 * copy it over 'ssc' */
1363 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1364 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1365 StructCopy(and_with, ssc, regnode_ssc);
1366 }
1367 else {
1368 ssc->invlist = anded_cp_list;
1369 ANYOF_POSIXL_ZERO(ssc);
93e92956 1370 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1371 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1372 }
1373 }
1374 }
e0e1be5f 1375 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1376 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1377 {
1378 /* One or the other of P1, P2 is non-empty. */
93e92956 1379 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1380 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1381 }
cdd87c1d
KW
1382 ssc_union(ssc, anded_cp_list, FALSE);
1383 }
1384 else { /* P1 = P2 = empty */
1385 ssc_intersection(ssc, anded_cp_list, FALSE);
1386 }
137165a6 1387 }
653099ff
GS
1388}
1389
653099ff 1390STATIC void
cdd87c1d 1391S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1392 const regnode_charclass *or_with)
653099ff 1393{
cdd87c1d
KW
1394 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1395 * another SSC or a regular ANYOF class. Can create false positives if
1396 * 'or_with' is to be inverted. */
7918f24d 1397
a0dd4231
KW
1398 SV* ored_cp_list;
1399 U8 ored_flags;
c6b76537 1400
cdd87c1d 1401 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1402
71068078 1403 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1404
1405 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1406 * the code point inversion list and just the relevant flags */
71068078 1407 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1408 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1409 ored_flags = ANYOF_FLAGS(or_with);
1410 }
1411 else {
5c0f85ef 1412 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1413 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
a0dd4231
KW
1414 }
1415
1416 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1417
1418 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1419 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1420 * 'or_with' may be inverted. When not inverted, we have the simple
1421 * situation of computing:
1422 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1423 * If P1|P2 yields a situation with both a class and its complement are
1424 * set, like having both \w and \W, this matches all code points, and we
1425 * can delete these from the P component of the ssc going forward. XXX We
1426 * might be able to delete all the P components, but I (khw) am not certain
1427 * about this, and it is better to be safe.
1428 *
1429 * Inverted, we have
1430 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1431 * <= (C1 | P1) | ~C2
1432 * <= (C1 | ~C2) | P1
1433 * (which results in actually simpler code than the non-inverted case)
1434 * */
9826f543 1435
a0dd4231 1436 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1437 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1438 {
cdd87c1d 1439 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1440 } /* else Not inverted */
93e92956 1441 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1442 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1443 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1444 unsigned int i;
1445 for (i = 0; i < ANYOF_MAX; i += 2) {
1446 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1447 {
1448 ssc_match_all_cp(ssc);
1449 ANYOF_POSIXL_CLEAR(ssc, i);
1450 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1451 }
1452 }
1453 }
1aa99e6b 1454 }
cdd87c1d
KW
1455
1456 ssc_union(ssc,
1457 ored_cp_list,
1458 FALSE /* Already has been inverted */
1459 );
653099ff
GS
1460}
1461
b423522f
KW
1462PERL_STATIC_INLINE void
1463S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1464{
1465 PERL_ARGS_ASSERT_SSC_UNION;
1466
71068078 1467 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1468
1469 _invlist_union_maybe_complement_2nd(ssc->invlist,
1470 invlist,
1471 invert2nd,
1472 &ssc->invlist);
1473}
1474
1475PERL_STATIC_INLINE void
1476S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1477 SV* const invlist,
1478 const bool invert2nd)
1479{
1480 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1481
71068078 1482 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1483
1484 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1485 invlist,
1486 invert2nd,
1487 &ssc->invlist);
1488}
1489
1490PERL_STATIC_INLINE void
1491S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1492{
1493 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1494
71068078 1495 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1496
1497 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1498}
1499
1500PERL_STATIC_INLINE void
1501S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1502{
1503 /* AND just the single code point 'cp' into the SSC 'ssc' */
1504
1505 SV* cp_list = _new_invlist(2);
1506
1507 PERL_ARGS_ASSERT_SSC_CP_AND;
1508
71068078 1509 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1510
1511 cp_list = add_cp_to_invlist(cp_list, cp);
1512 ssc_intersection(ssc, cp_list,
1513 FALSE /* Not inverted */
1514 );
1515 SvREFCNT_dec_NN(cp_list);
1516}
1517
1518PERL_STATIC_INLINE void
dc3bf405 1519S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1520{
1521 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1522 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1523
71068078 1524 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1525
1526 ANYOF_POSIXL_ZERO(ssc);
1527 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1528}
1529
b35552de
KW
1530#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1531
1532STATIC bool
1533S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1534{
1535 /* The synthetic start class is used to hopefully quickly winnow down
1536 * places where a pattern could start a match in the target string. If it
1537 * doesn't really narrow things down that much, there isn't much point to
1538 * having the overhead of using it. This function uses some very crude
1539 * heuristics to decide if to use the ssc or not.
1540 *
1541 * It returns TRUE if 'ssc' rules out more than half what it considers to
1542 * be the "likely" possible matches, but of course it doesn't know what the
1543 * actual things being matched are going to be; these are only guesses
1544 *
1545 * For /l matches, it assumes that the only likely matches are going to be
1546 * in the 0-255 range, uniformly distributed, so half of that is 127
1547 * For /a and /d matches, it assumes that the likely matches will be just
1548 * the ASCII range, so half of that is 63
1549 * For /u and there isn't anything matching above the Latin1 range, it
1550 * assumes that that is the only range likely to be matched, and uses
1551 * half that as the cut-off: 127. If anything matches above Latin1,
1552 * it assumes that all of Unicode could match (uniformly), except for
1553 * non-Unicode code points and things in the General Category "Other"
1554 * (unassigned, private use, surrogates, controls and formats). This
1555 * is a much large number. */
1556
1557 const U32 max_match = (LOC)
1558 ? 127
1559 : (! UNI_SEMANTICS)
1560 ? 63
1561 : (invlist_highest(ssc->invlist) < 256)
1562 ? 127
1563 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1564 U32 count = 0; /* Running total of number of code points matched by
1565 'ssc' */
1566 UV start, end; /* Start and end points of current range in inversion
1567 list */
1568
1569 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1570
1571 invlist_iterinit(ssc->invlist);
1572 while (invlist_iternext(ssc->invlist, &start, &end)) {
1573
1574 /* /u is the only thing that we expect to match above 255; so if not /u
1575 * and even if there are matches above 255, ignore them. This catches
1576 * things like \d under /d which does match the digits above 255, but
1577 * since the pattern is /d, it is not likely to be expecting them */
1578 if (! UNI_SEMANTICS) {
1579 if (start > 255) {
1580 break;
1581 }
1582 end = MIN(end, 255);
1583 }
1584 count += end - start + 1;
1585 if (count > max_match) {
1586 invlist_iterfinish(ssc->invlist);
1587 return FALSE;
1588 }
1589 }
1590
1591 return TRUE;
1592}
1593
1594
b423522f
KW
1595STATIC void
1596S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1597{
1598 /* The inversion list in the SSC is marked mortal; now we need a more
1599 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1600 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1601 * map */
b423522f
KW
1602
1603 SV* invlist = invlist_clone(ssc->invlist);
1604
1605 PERL_ARGS_ASSERT_SSC_FINALIZE;
1606
71068078 1607 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 1608
a0dd4231 1609 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
1610 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1611 * by the time we reach here */
eff8b7dc 1612 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
a0dd4231 1613
b423522f
KW
1614 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1615
1ee208c4
KW
1616 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1617 NULL, NULL, NULL, FALSE);
b423522f 1618
85c8e306
KW
1619 /* Make sure is clone-safe */
1620 ssc->invlist = NULL;
1621
e0e1be5f 1622 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 1623 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 1624 }
1462525b
KW
1625
1626 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
1627}
1628
a3621e74
YO
1629#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1630#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1631#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
1632#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1633 ? (TRIE_LIST_CUR( idx ) - 1) \
1634 : 0 )
a3621e74 1635
3dab1dad
YO
1636
1637#ifdef DEBUGGING
07be1b83 1638/*
2b8b4781
NC
1639 dump_trie(trie,widecharmap,revcharmap)
1640 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1641 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1642
1643 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1644 The _interim_ variants are used for debugging the interim
1645 tables that are used to generate the final compressed
1646 representation which is what dump_trie expects.
1647
486ec47a 1648 Part of the reason for their existence is to provide a form
3dab1dad 1649 of documentation as to how the different representations function.
07be1b83
YO
1650
1651*/
3dab1dad
YO
1652
1653/*
3dab1dad
YO
1654 Dumps the final compressed table form of the trie to Perl_debug_log.
1655 Used for debugging make_trie().
1656*/
b9a59e08 1657
3dab1dad 1658STATIC void
2b8b4781
NC
1659S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1660 AV *revcharmap, U32 depth)
3dab1dad
YO
1661{
1662 U32 state;
ab3bbdeb 1663 SV *sv=sv_newmortal();
55eed653 1664 int colwidth= widecharmap ? 6 : 4;
2e64971a 1665 U16 word;
3dab1dad
YO
1666 GET_RE_DEBUG_FLAGS_DECL;
1667
7918f24d 1668 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1669
3dab1dad
YO
1670 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1671 (int)depth * 2 + 2,"",
1672 "Match","Base","Ofs" );
1673
1674 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1675 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1676 if ( tmp ) {
538e84ed 1677 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1678 colwidth,
538e84ed 1679 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1680 PL_colors[0], PL_colors[1],
1681 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1682 PERL_PV_ESCAPE_FIRSTCHAR
1683 )
ab3bbdeb 1684 );
3dab1dad
YO
1685 }
1686 }
1687 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1688 (int)depth * 2 + 2,"");
1689
1690 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1691 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1692 PerlIO_printf( Perl_debug_log, "\n");
1693
1e2e3d02 1694 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1695 const U32 base = trie->states[ state ].trans.base;
3dab1dad 1696
538e84ed
KW
1697 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1698 (int)depth * 2 + 2,"", (UV)state);
3dab1dad
YO
1699
1700 if ( trie->states[ state ].wordnum ) {
538e84ed
KW
1701 PerlIO_printf( Perl_debug_log, " W%4X",
1702 trie->states[ state ].wordnum );
3dab1dad
YO
1703 } else {
1704 PerlIO_printf( Perl_debug_log, "%6s", "" );
1705 }
1706
1707 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1708
1709 if ( base ) {
1710 U32 ofs = 0;
1711
1712 while( ( base + ofs < trie->uniquecharcount ) ||
1713 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
1714 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1715 != state))
3dab1dad
YO
1716 ofs++;
1717
1718 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1719
1720 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
1721 if ( ( base + ofs >= trie->uniquecharcount )
1722 && ( base + ofs - trie->uniquecharcount
1723 < trie->lasttrans )
1724 && trie->trans[ base + ofs
1725 - trie->uniquecharcount ].check == state )
3dab1dad 1726 {
ab3bbdeb
YO
1727 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1728 colwidth,
538e84ed
KW
1729 (UV)trie->trans[ base + ofs
1730 - trie->uniquecharcount ].next );
3dab1dad 1731 } else {
ab3bbdeb 1732 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1733 }
1734 }
1735
1736 PerlIO_printf( Perl_debug_log, "]");
1737
1738 }
1739 PerlIO_printf( Perl_debug_log, "\n" );
1740 }
538e84ed
KW
1741 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1742 (int)depth*2, "");
2e64971a
DM
1743 for (word=1; word <= trie->wordcount; word++) {
1744 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1745 (int)word, (int)(trie->wordinfo[word].prev),
1746 (int)(trie->wordinfo[word].len));
1747 }
1748 PerlIO_printf(Perl_debug_log, "\n" );
538e84ed 1749}
3dab1dad 1750/*
3dab1dad 1751 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 1752 List tries normally only are used for construction when the number of
3dab1dad
YO
1753 possible chars (trie->uniquecharcount) is very high.
1754 Used for debugging make_trie().
1755*/
1756STATIC void
55eed653 1757S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1758 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1759 U32 depth)
3dab1dad
YO
1760{
1761 U32 state;
ab3bbdeb 1762 SV *sv=sv_newmortal();
55eed653 1763 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1764 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1765
1766 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1767
3dab1dad 1768 /* print out the table precompression. */
ab3bbdeb
YO
1769 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1770 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1771 "------:-----+-----------------\n" );
538e84ed 1772
3dab1dad
YO
1773 for( state=1 ; state < next_alloc ; state ++ ) {
1774 U16 charid;
538e84ed 1775
ab3bbdeb 1776 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1777 (int)depth * 2 + 2,"", (UV)state );
1778 if ( ! trie->states[ state ].wordnum ) {
1779 PerlIO_printf( Perl_debug_log, "%5s| ","");
1780 } else {
1781 PerlIO_printf( Perl_debug_log, "W%4x| ",
1782 trie->states[ state ].wordnum
1783 );
1784 }
1785 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
1786 SV ** const tmp = av_fetch( revcharmap,
1787 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1788 if ( tmp ) {
1789 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1790 colwidth,
538e84ed
KW
1791 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1792 colwidth,
1793 PL_colors[0], PL_colors[1],
1794 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1795 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 1796 ) ,
1e2e3d02
YO
1797 TRIE_LIST_ITEM(state,charid).forid,
1798 (UV)TRIE_LIST_ITEM(state,charid).newstate
1799 );
538e84ed 1800 if (!(charid % 10))
664e119d
RGS
1801 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1802 (int)((depth * 2) + 14), "");
1e2e3d02 1803 }
ab3bbdeb
YO
1804 }
1805 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad 1806 }
538e84ed 1807}
3dab1dad
YO
1808
1809/*
3dab1dad 1810 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
1811 This is the normal DFA style state transition table, with a few
1812 twists to facilitate compression later.
3dab1dad
YO
1813 Used for debugging make_trie().
1814*/
1815STATIC void
55eed653 1816S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1817 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1818 U32 depth)
3dab1dad
YO
1819{
1820 U32 state;
1821 U16 charid;
ab3bbdeb 1822 SV *sv=sv_newmortal();
55eed653 1823 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1824 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1825
1826 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 1827
3dab1dad
YO
1828 /*
1829 print out the table precompression so that we can do a visual check
1830 that they are identical.
1831 */
538e84ed 1832
3dab1dad
YO
1833 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1834
1835 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1836 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1837 if ( tmp ) {
538e84ed 1838 PerlIO_printf( Perl_debug_log, "%*s",
ab3bbdeb 1839 colwidth,
538e84ed 1840 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1841 PL_colors[0], PL_colors[1],
1842 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
1843 PERL_PV_ESCAPE_FIRSTCHAR
1844 )
ab3bbdeb 1845 );
3dab1dad
YO
1846 }
1847 }
1848
1849 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1850
1851 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1852 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1853 }
1854
1855 PerlIO_printf( Perl_debug_log, "\n" );
1856
1857 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1858
538e84ed 1859 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
3dab1dad
YO
1860 (int)depth * 2 + 2,"",
1861 (UV)TRIE_NODENUM( state ) );
1862
1863 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1864 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1865 if (v)
1866 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1867 else
1868 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1869 }
1870 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
538e84ed
KW
1871 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1872 (UV)trie->trans[ state ].check );
3dab1dad 1873 } else {
538e84ed
KW
1874 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1875 (UV)trie->trans[ state ].check,
3dab1dad
YO
1876 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1877 }
1878 }
07be1b83 1879}
3dab1dad
YO
1880
1881#endif
1882
2e64971a 1883
786e8c11
YO
1884/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1885 startbranch: the first branch in the whole branch sequence
1886 first : start branch of sequence of branch-exact nodes.
1887 May be the same as startbranch
1888 last : Thing following the last branch.
1889 May be the same as tail.
1890 tail : item following the branch sequence
1891 count : words in the sequence
a4525e78 1892 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 1893 depth : indent depth
3dab1dad 1894
786e8c11 1895Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1896
786e8c11
YO
1897A trie is an N'ary tree where the branches are determined by digital
1898decomposition of the key. IE, at the root node you look up the 1st character and
1899follow that branch repeat until you find the end of the branches. Nodes can be
1900marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1901
786e8c11 1902 /he|she|his|hers/
72f13be8 1903
786e8c11
YO
1904would convert into the following structure. Numbers represent states, letters
1905following numbers represent valid transitions on the letter from that state, if
1906the number is in square brackets it represents an accepting state, otherwise it
1907will be in parenthesis.
07be1b83 1908
786e8c11
YO
1909 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1910 | |
1911 | (2)
1912 | |
1913 (1) +-i->(6)-+-s->[7]
1914 |
1915 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1916
786e8c11
YO
1917 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1918
1919This shows that when matching against the string 'hers' we will begin at state 1
1920read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1921then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1922is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1923single traverse. We store a mapping from accepting to state to which word was
1924matched, and then when we have multiple possibilities we try to complete the
b8fda935 1925rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
1926
1927The only prior NFA like behaviour that would be changed by the TRIE support is
1928the silent ignoring of duplicate alternations which are of the form:
1929
1930 / (DUPE|DUPE) X? (?{ ... }) Y /x
1931
4b714af6 1932Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1933and without the optimisation. With the optimisations dupes will be silently
486ec47a 1934ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1935the following demonstrates:
1936
1937 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1938
1939which prints out 'word' three times, but
1940
1941 'words'=~/(word|word|word)(?{ print $1 })S/
1942
1943which doesnt print it out at all. This is due to other optimisations kicking in.
1944
1945Example of what happens on a structural level:
1946
486ec47a 1947The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1948
1949 1: CURLYM[1] {1,32767}(18)
1950 5: BRANCH(8)
1951 6: EXACT <ac>(16)
1952 8: BRANCH(11)
1953 9: EXACT <ad>(16)
1954 11: BRANCH(14)
1955 12: EXACT <ab>(16)
1956 16: SUCCEED(0)
1957 17: NOTHING(18)
1958 18: END(0)
1959
1960This would be optimizable with startbranch=5, first=5, last=16, tail=16
1961and should turn into:
1962
1963 1: CURLYM[1] {1,32767}(18)
1964 5: TRIE(16)
1965 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1966 <ac>
1967 <ad>
1968 <ab>
1969 16: SUCCEED(0)
1970 17: NOTHING(18)
1971 18: END(0)
1972
1973Cases where tail != last would be like /(?foo|bar)baz/:
1974
1975 1: BRANCH(4)
1976 2: EXACT <foo>(8)
1977 4: BRANCH(7)
1978 5: EXACT <bar>(8)
1979 7: TAIL(8)
1980 8: EXACT <baz>(10)
1981 10: END(0)
1982
1983which would be optimizable with startbranch=1, first=1, last=7, tail=8
1984and would end up looking like:
1985
1986 1: TRIE(8)
1987 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1988 <foo>
1989 <bar>
1990 7: TAIL(8)
1991 8: EXACT <baz>(10)
1992 10: END(0)
1993
c80e42f3 1994 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
1995
1996is the recommended Unicode-aware way of saying
1997
1998 *(d++) = uv;
1999*/
2000
fab2782b 2001#define TRIE_STORE_REVCHAR(val) \
786e8c11 2002 STMT_START { \
73031816 2003 if (UTF) { \
fab2782b 2004 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 2005 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2006 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2007 SvCUR_set(zlopp, kapow - flrbbbbb); \
2008 SvPOK_on(zlopp); \
2009 SvUTF8_on(zlopp); \
2010 av_push(revcharmap, zlopp); \
2011 } else { \
fab2782b 2012 char ooooff = (char)val; \
73031816
NC
2013 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2014 } \
2015 } STMT_END
786e8c11 2016
914a25d5
KW
2017/* This gets the next character from the input, folding it if not already
2018 * folded. */
2019#define TRIE_READ_CHAR STMT_START { \
2020 wordlen++; \
2021 if ( UTF ) { \
2022 /* if it is UTF then it is either already folded, or does not need \
2023 * folding */ \
1c1d615a 2024 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2025 } \
2026 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2027 /* This folder implies Unicode rules, which in the range expressible \
2028 * by not UTF is the lower case, with the two exceptions, one of \
2029 * which should have been taken care of before calling this */ \
2030 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2031 uvc = toLOWER_L1(*uc); \
2032 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2033 len = 1; \
914a25d5
KW
2034 } else { \
2035 /* raw data, will be folded later if needed */ \
2036 uvc = (U32)*uc; \
2037 len = 1; \
2038 } \
786e8c11
YO
2039} STMT_END
2040
2041
2042
2043#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2044 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
2045 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2046 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
2047 } \
2048 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2049 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2050 TRIE_LIST_CUR( state )++; \
2051} STMT_END
07be1b83 2052
786e8c11
YO
2053#define TRIE_LIST_NEW(state) STMT_START { \
2054 Newxz( trie->states[ state ].trans.list, \
2055 4, reg_trie_trans_le ); \
2056 TRIE_LIST_CUR( state ) = 1; \
2057 TRIE_LIST_LEN( state ) = 4; \
2058} STMT_END
07be1b83 2059
786e8c11
YO
2060#define TRIE_HANDLE_WORD(state) STMT_START { \
2061 U16 dupe= trie->states[ state ].wordnum; \
2062 regnode * const noper_next = regnext( noper ); \
2063 \
786e8c11
YO
2064 DEBUG_r({ \
2065 /* store the word for dumping */ \
2066 SV* tmp; \
2067 if (OP(noper) != NOTHING) \
740cce10 2068 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2069 else \
740cce10 2070 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2071 av_push( trie_words, tmp ); \
786e8c11
YO
2072 }); \
2073 \
2074 curword++; \
2e64971a
DM
2075 trie->wordinfo[curword].prev = 0; \
2076 trie->wordinfo[curword].len = wordlen; \
2077 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2078 \
2079 if ( noper_next < tail ) { \
2080 if (!trie->jump) \
538e84ed
KW
2081 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2082 sizeof(U16) ); \
7f69552c 2083 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2084 if (!jumper) \
2085 jumper = noper_next; \
2086 if (!nextbranch) \
2087 nextbranch= regnext(cur); \
2088 } \
2089 \
2090 if ( dupe ) { \
2e64971a
DM
2091 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2092 /* chain, so that when the bits of chain are later */\
2093 /* linked together, the dups appear in the chain */\
2094 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2095 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2096 } else { \
2097 /* we haven't inserted this word yet. */ \
2098 trie->states[ state ].wordnum = curword; \
2099 } \
2100} STMT_END
07be1b83 2101
3dab1dad 2102
786e8c11
YO
2103#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2104 ( ( base + charid >= ucharcount \
2105 && base + charid < ubound \
2106 && state == trie->trans[ base - ucharcount + charid ].check \
2107 && trie->trans[ base - ucharcount + charid ].next ) \
2108 ? trie->trans[ base - ucharcount + charid ].next \
2109 : ( state==1 ? special : 0 ) \
2110 )
3dab1dad 2111
786e8c11
YO
2112#define MADE_TRIE 1
2113#define MADE_JUMP_TRIE 2
2114#define MADE_EXACT_TRIE 4
3dab1dad 2115
a3621e74 2116STATIC I32
538e84ed
KW
2117S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2118 regnode *first, regnode *last, regnode *tail,
2119 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2120{
2121 /* first pass, loop through and scan words */
2122 reg_trie_data *trie;
55eed653 2123 HV *widecharmap = NULL;
2b8b4781 2124 AV *revcharmap = newAV();
a3621e74 2125 regnode *cur;
a3621e74
YO
2126 STRLEN len = 0;
2127 UV uvc = 0;
2128 U16 curword = 0;
2129 U32 next_alloc = 0;
786e8c11
YO
2130 regnode *jumper = NULL;
2131 regnode *nextbranch = NULL;
7f69552c 2132 regnode *convert = NULL;
2e64971a 2133 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2134 /* we just use folder as a flag in utf8 */
1e696034 2135 const U8 * folder = NULL;
a3621e74 2136
2b8b4781 2137#ifdef DEBUGGING
cf78de0b 2138 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2b8b4781
NC
2139 AV *trie_words = NULL;
2140 /* along with revcharmap, this only used during construction but both are
2141 * useful during debugging so we store them in the struct when debugging.
8e11feef 2142 */
2b8b4781 2143#else
cf78de0b 2144 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2145 STRLEN trie_charcount=0;
3dab1dad 2146#endif
2b8b4781 2147 SV *re_trie_maxbuff;
a3621e74 2148 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2149
2150 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2151#ifndef DEBUGGING
2152 PERL_UNUSED_ARG(depth);
2153#endif
a3621e74 2154
1e696034 2155 switch (flags) {
a4525e78 2156 case EXACT: case EXACTL: break;
2f7f8cb1 2157 case EXACTFA:
fab2782b 2158 case EXACTFU_SS:
a4525e78
KW
2159 case EXACTFU:
2160 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2161 case EXACTF: folder = PL_fold; break;
fab2782b 2162 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2163 }
2164
c944940b 2165 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2166 trie->refcount = 1;
3dab1dad 2167 trie->startstate = 1;
786e8c11 2168 trie->wordcount = word_count;
f8fc2ecf 2169 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2170 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2171 if (flags == EXACT || flags == EXACTL)
c944940b 2172 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2173 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2174 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2175
a3621e74 2176 DEBUG_r({
2b8b4781 2177 trie_words = newAV();
a3621e74 2178 });
a3621e74 2179
0111c4fd 2180 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2181 assert(re_trie_maxbuff);
a3621e74 2182 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2183 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2184 }
df826430 2185 DEBUG_TRIE_COMPILE_r({
538e84ed
KW
2186 PerlIO_printf( Perl_debug_log,
2187 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2188 (int)depth * 2 + 2, "",
2189 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2190 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2191 });
538e84ed 2192
7f69552c
YO
2193 /* Find the node we are going to overwrite */
2194 if ( first == startbranch && OP( last ) != BRANCH ) {
2195 /* whole branch chain */
2196 convert = first;
2197 } else {
2198 /* branch sub-chain */
2199 convert = NEXTOPER( first );
2200 }
538e84ed 2201
a3621e74
YO
2202 /* -- First loop and Setup --
2203
2204 We first traverse the branches and scan each word to determine if it
2205 contains widechars, and how many unique chars there are, this is
2206 important as we have to build a table with at least as many columns as we
2207 have unique chars.
2208
2209 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2210 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2211 the native representation of the character value as the key and IV's for
2212 the coded index.
a3621e74
YO
2213
2214 *TODO* If we keep track of how many times each character is used we can
2215 remap the columns so that the table compression later on is more
3b753521 2216 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2217 middle and the least common are on the outside. IMO this would be better
2218 than a most to least common mapping as theres a decent chance the most
2219 common letter will share a node with the least common, meaning the node
486ec47a 2220 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2221 case is when we have the least common nodes twice.
2222
2223 */
2224
a3621e74 2225 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2226 regnode *noper = NEXTOPER( cur );
e1ec3a88 2227 const U8 *uc = (U8*)STRING( noper );
df826430 2228 const U8 *e = uc + STR_LEN( noper );
bc031a7d 2229 int foldlen = 0;
07be1b83 2230 U32 wordlen = 0; /* required init */
bc031a7d
KW
2231 STRLEN minchars = 0;
2232 STRLEN maxchars = 0;
538e84ed
KW
2233 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2234 bitmap?*/
a3621e74 2235
3dab1dad 2236 if (OP(noper) == NOTHING) {
df826430
YO
2237 regnode *noper_next= regnext(noper);
2238 if (noper_next != tail && OP(noper_next) == flags) {
2239 noper = noper_next;
2240 uc= (U8*)STRING(noper);
2241 e= uc + STR_LEN(noper);
2242 trie->minlen= STR_LEN(noper);
2243 } else {
2244 trie->minlen= 0;
2245 continue;
2246 }
3dab1dad 2247 }
df826430 2248
fab2782b 2249 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2250 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2251 regardless of encoding */
fab2782b
YO
2252 if (OP( noper ) == EXACTFU_SS) {
2253 /* false positives are ok, so just set this */
0dc4a61d 2254 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2255 }
2256 }
bc031a7d
KW
2257 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2258 branch */
3dab1dad 2259 TRIE_CHARCOUNT(trie)++;
a3621e74 2260 TRIE_READ_CHAR;
645de4ce 2261
bc031a7d
KW
2262 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2263 * is in effect. Under /i, this character can match itself, or
2264 * anything that folds to it. If not under /i, it can match just
2265 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2266 * all fold to k, and all are single characters. But some folds
2267 * expand to more than one character, so for example LATIN SMALL
2268 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2269 * the string beginning at 'uc' is 'ffi', it could be matched by
2270 * three characters, or just by the one ligature character. (It
2271 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2272 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2273 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2274 * match.) The trie needs to know the minimum and maximum number
2275 * of characters that could match so that it can use size alone to
2276 * quickly reject many match attempts. The max is simple: it is
2277 * the number of folded characters in this branch (since a fold is
2278 * never shorter than what folds to it. */
2279
2280 maxchars++;
2281
2282 /* And the min is equal to the max if not under /i (indicated by
2283 * 'folder' being NULL), or there are no multi-character folds. If
2284 * there is a multi-character fold, the min is incremented just
2285 * once, for the character that folds to the sequence. Each
2286 * character in the sequence needs to be added to the list below of
2287 * characters in the trie, but we count only the first towards the
2288 * min number of characters needed. This is done through the
2289 * variable 'foldlen', which is returned by the macros that look
2290 * for these sequences as the number of bytes the sequence
2291 * occupies. Each time through the loop, we decrement 'foldlen' by
2292 * how many bytes the current char occupies. Only when it reaches
2293 * 0 do we increment 'minchars' or look for another multi-character
2294 * sequence. */
2295 if (folder == NULL) {
2296 minchars++;
2297 }
2298 else if (foldlen > 0) {
2299 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2300 }
2301 else {
bc031a7d
KW
2302 minchars++;
2303
2304 /* See if *uc is the beginning of a multi-character fold. If
2305 * so, we decrement the length remaining to look at, to account
2306 * for the current character this iteration. (We can use 'uc'
2307 * instead of the fold returned by TRIE_READ_CHAR because for
2308 * non-UTF, the latin1_safe macro is smart enough to account
2309 * for all the unfolded characters, and because for UTF, the
2310 * string will already have been folded earlier in the
2311 * compilation process */
2312 if (UTF) {
2313 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2314 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2315 }
2316 }
bc031a7d
KW
2317 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2318 foldlen--;
2319 }
645de4ce 2320 }
bc031a7d
KW
2321
2322 /* The current character (and any potential folds) should be added
2323 * to the possible matching characters for this position in this
2324 * branch */
a3621e74 2325 if ( uvc < 256 ) {
fab2782b
YO
2326 if ( folder ) {
2327 U8 folded= folder[ (U8) uvc ];
2328 if ( !trie->charmap[ folded ] ) {
2329 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2330 TRIE_STORE_REVCHAR( folded );
2331 }
2332 }
a3621e74
YO
2333 if ( !trie->charmap[ uvc ] ) {
2334 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2335 TRIE_STORE_REVCHAR( uvc );
a3621e74 2336 }
02daf0ab 2337 if ( set_bit ) {
62012aee
KW
2338 /* store the codepoint in the bitmap, and its folded
2339 * equivalent. */
fab2782b 2340 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
2341
2342 /* store the folded codepoint */
fab2782b 2343 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
2344
2345 if ( !UTF ) {
2346 /* store first byte of utf8 representation of
acdf4139 2347 variant codepoints */
6f2d5cbc 2348 if (! UVCHR_IS_INVARIANT(uvc)) {
acdf4139 2349 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
2350 }
2351 }
02daf0ab
YO
2352 set_bit = 0; /* We've done our bit :-) */
2353 }
a3621e74 2354 } else {
bc031a7d
KW
2355
2356 /* XXX We could come up with the list of code points that fold
2357 * to this using PL_utf8_foldclosures, except not for
2358 * multi-char folds, as there may be multiple combinations
2359 * there that could work, which needs to wait until runtime to
2360 * resolve (The comment about LIGATURE FFI above is such an
2361 * example */
2362
a3621e74 2363 SV** svpp;
55eed653
NC
2364 if ( !widecharmap )
2365 widecharmap = newHV();
a3621e74 2366
55eed653 2367 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2368
2369 if ( !svpp )
e4584336 2370 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
2371
2372 if ( !SvTRUE( *svpp ) ) {
2373 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2374 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2375 }
2376 }
bc031a7d
KW
2377 } /* end loop through characters in this branch of the trie */
2378
2379 /* We take the min and max for this branch and combine to find the min
2380 * and max for all branches processed so far */
3dab1dad 2381 if( cur == first ) {
bc031a7d
KW
2382 trie->minlen = minchars;
2383 trie->maxlen = maxchars;
2384 } else if (minchars < trie->minlen) {
2385 trie->minlen = minchars;
2386 } else if (maxchars > trie->maxlen) {
2387 trie->maxlen = maxchars;
fab2782b 2388 }
a3621e74
YO
2389 } /* end first pass */
2390 DEBUG_TRIE_COMPILE_r(
538e84ed
KW
2391 PerlIO_printf( Perl_debug_log,
2392 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3dab1dad 2393 (int)depth * 2 + 2,"",
55eed653 2394 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2395 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2396 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2397 );
a3621e74
YO
2398
2399 /*
2400 We now know what we are dealing with in terms of unique chars and
2401 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2402 representation using a flat table will take. If it's over a reasonable
2403 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2404 conservative but potentially much slower representation using an array
2405 of lists.
2406
2407 At the end we convert both representations into the same compressed
2408 form that will be used in regexec.c for matching with. The latter
2409 is a form that cannot be used to construct with but has memory
2410 properties similar to the list form and access properties similar
2411 to the table form making it both suitable for fast searches and
2412 small enough that its feasable to store for the duration of a program.
2413
2414 See the comment in the code where the compressed table is produced
2415 inplace from the flat tabe representation for an explanation of how
2416 the compression works.
2417
2418 */
2419
2420
2e64971a
DM
2421 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2422 prev_states[1] = 0;
2423
538e84ed
KW
2424 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2425 > SvIV(re_trie_maxbuff) )
2426 {
a3621e74
YO
2427 /*
2428 Second Pass -- Array Of Lists Representation
2429
2430 Each state will be represented by a list of charid:state records
2431 (reg_trie_trans_le) the first such element holds the CUR and LEN
2432 points of the allocated array. (See defines above).
2433
2434 We build the initial structure using the lists, and then convert
2435 it into the compressed table form which allows faster lookups
2436 (but cant be modified once converted).
a3621e74
YO
2437 */
2438
a3621e74
YO
2439 STRLEN transcount = 1;
2440
538e84ed 2441 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2442 "%*sCompiling trie using list compiler\n",
2443 (int)depth * 2 + 2, ""));
686b73d4 2444
c944940b
JH
2445 trie->states = (reg_trie_state *)
2446 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2447 sizeof(reg_trie_state) );
a3621e74
YO
2448 TRIE_LIST_NEW(1);
2449 next_alloc = 2;
2450
2451 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2452
df826430 2453 regnode *noper = NEXTOPER( cur );
c445ea15 2454 U8 *uc = (U8*)STRING( noper );
df826430 2455 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
2456 U32 state = 1; /* required init */
2457 U16 charid = 0; /* sanity init */
07be1b83 2458 U32 wordlen = 0; /* required init */
c445ea15 2459
df826430
YO
2460 if (OP(noper) == NOTHING) {
2461 regnode *noper_next= regnext(noper);
2462 if (noper_next != tail && OP(noper_next) == flags) {
2463 noper = noper_next;
2464 uc= (U8*)STRING(noper);
2465 e= uc + STR_LEN(noper);
2466 }
2467 }
2468
3dab1dad 2469 if (OP(noper) != NOTHING) {
786e8c11 2470 for ( ; uc < e ; uc += len ) {
c445ea15 2471
786e8c11 2472 TRIE_READ_CHAR;
c445ea15 2473
786e8c11
YO
2474 if ( uvc < 256 ) {
2475 charid = trie->charmap[ uvc ];
c445ea15 2476 } else {
538e84ed
KW
2477 SV** const svpp = hv_fetch( widecharmap,
2478 (char*)&uvc,
2479 sizeof( UV ),
2480 0);
786e8c11
YO
2481 if ( !svpp ) {
2482 charid = 0;
2483 } else {
2484 charid=(U16)SvIV( *svpp );
2485 }
c445ea15 2486 }
538e84ed
KW
2487 /* charid is now 0 if we dont know the char read, or
2488 * nonzero if we do */
786e8c11 2489 if ( charid ) {
a3621e74 2490
786e8c11
YO
2491 U16 check;
2492 U32 newstate = 0;
a3621e74 2493
786e8c11
YO
2494 charid--;
2495 if ( !trie->states[ state ].trans.list ) {
2496 TRIE_LIST_NEW( state );
c445ea15 2497 }
538e84ed
KW
2498 for ( check = 1;
2499 check <= TRIE_LIST_USED( state );
2500 check++ )
2501 {
2502 if ( TRIE_LIST_ITEM( state, check ).forid
2503 == charid )
2504 {
786e8c11
YO
2505 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2506 break;
2507 }
2508 }
2509 if ( ! newstate ) {
2510 newstate = next_alloc++;
2e64971a 2511 prev_states[newstate] = state;
786e8c11
YO
2512 TRIE_LIST_PUSH( state, charid, newstate );
2513 transcount++;
2514 }
2515 state = newstate;
2516 } else {
2517 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 2518 }
a28509cc 2519 }
c445ea15 2520 }
3dab1dad 2521 TRIE_HANDLE_WORD(state);
a3621e74
YO
2522
2523 } /* end second pass */
2524
1e2e3d02 2525 /* next alloc is the NEXT state to be allocated */
538e84ed 2526 trie->statecount = next_alloc;
c944940b
JH
2527 trie->states = (reg_trie_state *)
2528 PerlMemShared_realloc( trie->states,
2529 next_alloc
2530 * sizeof(reg_trie_state) );
a3621e74 2531
3dab1dad 2532 /* and now dump it out before we compress it */
2b8b4781
NC
2533 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2534 revcharmap, next_alloc,
2535 depth+1)
1e2e3d02 2536 );
a3621e74 2537
c944940b
JH
2538 trie->trans = (reg_trie_trans *)
2539 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2540 {
2541 U32 state;
a3621e74
YO
2542 U32 tp = 0;
2543 U32 zp = 0;
2544
2545
2546 for( state=1 ; state < next_alloc ; state ++ ) {
2547 U32 base=0;
2548
2549 /*
2550 DEBUG_TRIE_COMPILE_MORE_r(
2551 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2552 );
2553 */
2554
2555 if (trie->states[state].trans.list) {
2556 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2557 U16 maxid=minid;
a28509cc 2558 U16 idx;
a3621e74
YO
2559
2560 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2561 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2562 if ( forid < minid ) {
2563 minid=forid;
2564 } else if ( forid > maxid ) {
2565 maxid=forid;
2566 }
a3621e74
YO
2567 }
2568 if ( transcount < tp + maxid - minid + 1) {
2569 transcount *= 2;
c944940b
JH
2570 trie->trans = (reg_trie_trans *)
2571 PerlMemShared_realloc( trie->trans,
446bd890
NC
2572 transcount
2573 * sizeof(reg_trie_trans) );
538e84ed
KW
2574 Zero( trie->trans + (transcount / 2),
2575 transcount / 2,
2576 reg_trie_trans );
a3621e74
YO
2577 }
2578 base = trie->uniquecharcount + tp - minid;
2579 if ( maxid == minid ) {
2580 U32 set = 0;
2581 for ( ; zp < tp ; zp++ ) {
2582 if ( ! trie->trans[ zp ].next ) {
2583 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
2584 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2585 1).newstate;
a3621e74
YO
2586 trie->trans[ zp ].check = state;
2587 set = 1;
2588 break;
2589 }
2590 }
2591 if ( !set ) {
538e84ed
KW
2592 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2593 1).newstate;
a3621e74
YO
2594 trie->trans[ tp ].check = state;
2595 tp++;
2596 zp = tp;
2597 }
2598 } else {
2599 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
2600 const U32 tid = base
2601 - trie->uniquecharcount
2602 + TRIE_LIST_ITEM( state, idx ).forid;
2603 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2604 idx ).newstate;
a3621e74
YO
2605 trie->trans[ tid ].check = state;
2606 }
2607 tp += ( maxid - minid + 1 );
2608 }
2609 Safefree(trie->states[ state ].trans.list);
2610 }
2611 /*
2612 DEBUG_TRIE_COMPILE_MORE_r(
2613 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2614 );
2615 */
2616 trie->states[ state ].trans.base=base;
2617 }
cc601c31 2618 trie->lasttrans = tp + 1;
a3621e74
YO
2619 }
2620 } else {
2621 /*
2622 Second Pass -- Flat Table Representation.
2623
b423522f
KW
2624 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2625 each. We know that we will need Charcount+1 trans at most to store
2626 the data (one row per char at worst case) So we preallocate both
2627 structures assuming worst case.
a3621e74
YO
2628
2629 We then construct the trie using only the .next slots of the entry
2630 structs.
2631
b423522f
KW
2632 We use the .check field of the first entry of the node temporarily
2633 to make compression both faster and easier by keeping track of how
2634 many non zero fields are in the node.
a3621e74
YO
2635
2636 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2637 transition.
2638
b423522f
KW
2639 There are two terms at use here: state as a TRIE_NODEIDX() which is
2640 a number representing the first entry of the node, and state as a
2641 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2642 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2643 if there are 2 entrys per node. eg:
a3621e74
YO
2644
2645 A B A B
2646 1. 2 4 1. 3 7
2647 2. 0 3 3. 0 5
2648 3. 0 0 5. 0 0
2649 4. 0 0 7. 0 0
2650
b423522f
KW
2651 The table is internally in the right hand, idx form. However as we
2652 also have to deal with the states array which is indexed by nodenum
2653 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
2654
2655 */
538e84ed 2656 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1e2e3d02
YO
2657 "%*sCompiling trie using table compiler\n",
2658 (int)depth * 2 + 2, ""));
3dab1dad 2659
c944940b
JH
2660 trie->trans = (reg_trie_trans *)
2661 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2662 * trie->uniquecharcount + 1,
2663 sizeof(reg_trie_trans) );
2664 trie->states = (reg_trie_state *)
2665 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2666 sizeof(reg_trie_state) );
a3621e74
YO
2667 next_alloc = trie->uniquecharcount + 1;
2668
3dab1dad 2669
a3621e74
YO
2670 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2671
df826430 2672 regnode *noper = NEXTOPER( cur );
a28509cc 2673 const U8 *uc = (U8*)STRING( noper );
df826430 2674 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2675
2676 U32 state = 1; /* required init */
2677
2678 U16 charid = 0; /* sanity init */
2679 U32 accept_state = 0; /* sanity init */
a3621e74 2680
07be1b83 2681 U32 wordlen = 0; /* required init */
a3621e74 2682
df826430
YO
2683 if (OP(noper) == NOTHING) {
2684 regnode *noper_next= regnext(noper);
2685 if (noper_next != tail && OP(noper_next) == flags) {
2686 noper = noper_next;
2687 uc= (U8*)STRING(noper);
2688 e= uc + STR_LEN(noper);
2689 }
2690 }
fab2782b 2691
3dab1dad 2692 if ( OP(noper) != NOTHING ) {
786e8c11 2693 for ( ; uc < e ; uc += len ) {
a3621e74 2694
786e8c11 2695 TRIE_READ_CHAR;
a3621e74 2696
786e8c11
YO
2697 if ( uvc < 256 ) {
2698 charid = trie->charmap[ uvc ];
2699 } else {
538e84ed
KW
2700 SV* const * const svpp = hv_fetch( widecharmap,
2701 (char*)&uvc,
2702 sizeof( UV ),
2703 0);
786e8c11 2704 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2705 }
786e8c11
YO
2706 if ( charid ) {
2707 charid--;
2708 if ( !trie->trans[ state + charid ].next ) {
2709 trie->trans[ state + charid ].next = next_alloc;
2710 trie->trans[ state ].check++;
2e64971a
DM
2711 prev_states[TRIE_NODENUM(next_alloc)]
2712 = TRIE_NODENUM(state);
786e8c11
YO
2713 next_alloc += trie->uniquecharcount;
2714 }
2715 state = trie->trans[ state + charid ].next;
2716 } else {
2717 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2718 }
538e84ed
KW
2719 /* charid is now 0 if we dont know the char read, or
2720 * nonzero if we do */
a3621e74 2721 }
a3621e74 2722 }
3dab1dad
YO
2723 accept_state = TRIE_NODENUM( state );
2724 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2725
2726 } /* end second pass */
2727
3dab1dad 2728 /* and now dump it out before we compress it */
2b8b4781
NC
2729 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2730 revcharmap,
2731 next_alloc, depth+1));
a3621e74 2732
a3621e74
YO
2733 {
2734 /*
2735 * Inplace compress the table.*
2736
2737 For sparse data sets the table constructed by the trie algorithm will
2738 be mostly 0/FAIL transitions or to put it another way mostly empty.
2739 (Note that leaf nodes will not contain any transitions.)
2740
2741 This algorithm compresses the tables by eliminating most such
2742 transitions, at the cost of a modest bit of extra work during lookup:
2743
2744 - Each states[] entry contains a .base field which indicates the
2745 index in the state[] array wheres its transition data is stored.
2746
3b753521 2747 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2748
2749 - If .base is nonzero then charid is added to it to find an entry in
2750 the trans array.
2751
2752 -If trans[states[state].base+charid].check!=state then the
2753 transition is taken to be a 0/Fail transition. Thus if there are fail
2754 transitions at the front of the node then the .base offset will point
2755 somewhere inside the previous nodes data (or maybe even into a node
2756 even earlier), but the .check field determines if the transition is
2757 valid.
2758
786e8c11 2759 XXX - wrong maybe?
a3621e74 2760 The following process inplace converts the table to the compressed
3b753521 2761 table: We first do not compress the root node 1,and mark all its
a3621e74 2762 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2763 allows us to do a DFA construction from the compressed table later,
2764 and ensures that any .base pointers we calculate later are greater
2765 than 0.
a3621e74
YO
2766
2767 - We set 'pos' to indicate the first entry of the second node.
2768
2769 - We then iterate over the columns of the node, finding the first and
2770 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2771 and set the .check pointers accordingly, and advance pos
2772 appropriately and repreat for the next node. Note that when we copy
2773 the next pointers we have to convert them from the original
2774 NODEIDX form to NODENUM form as the former is not valid post
2775 compression.
2776
2777 - If a node has no transitions used we mark its base as 0 and do not
2778 advance the pos pointer.
2779
2780 - If a node only has one transition we use a second pointer into the
2781 structure to fill in allocated fail transitions from other states.
2782 This pointer is independent of the main pointer and scans forward
2783 looking for null transitions that are allocated to a state. When it
2784 finds one it writes the single transition into the "hole". If the
786e8c11 2785 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2786
2787 - Once compressed we can Renew/realloc the structures to release the
2788 excess space.
2789
2790 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2791 specifically Fig 3.47 and the associated pseudocode.
2792
2793 demq
2794 */
a3b680e6 2795 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2796 U32 state, charid;
a3621e74 2797 U32 pos = 0, zp=0;
1e2e3d02 2798 trie->statecount = laststate;
a3621e74
YO
2799
2800 for ( state = 1 ; state < laststate ; state++ ) {
2801 U8 flag = 0;
a28509cc
AL
2802 const U32 stateidx = TRIE_NODEIDX( state );
2803 const U32 o_used = trie->trans[ stateidx ].check;
2804 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2805 trie->trans[ stateidx ].check = 0;
2806
538e84ed
KW
2807 for ( charid = 0;
2808 used && charid < trie->uniquecharcount;
2809 charid++ )
2810 {
a3621e74
YO
2811 if ( flag || trie->trans[ stateidx + charid ].next ) {
2812 if ( trie->trans[ stateidx + charid ].next ) {
2813 if (o_used == 1) {
2814 for ( ; zp < pos ; zp++ ) {
2815 if ( ! trie->trans[ zp ].next ) {
2816 break;
2817 }
2818 }
538e84ed
KW
2819 trie->states[ state ].trans.base
2820 = zp
2821 + trie->uniquecharcount
2822 - charid ;
2823 trie->trans[ zp ].next
2824 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2825 + charid ].next );
a3621e74
YO
2826 trie->trans[ zp ].check = state;
2827 if ( ++zp > pos ) pos = zp;
2828 break;
2829 }
2830 used--;
2831 }
2832 if ( !flag ) {
2833 flag = 1;
538e84ed
KW
2834 trie->states[ state ].trans.base
2835 = pos + trie->uniquecharcount - charid ;
a3621e74 2836 }
538e84ed
KW
2837 trie->trans[ pos ].next
2838 = SAFE_TRIE_NODENUM(
2839 trie->trans[ stateidx + charid ].next );
a3621e74
YO
2840 trie->trans[ pos ].check = state;
2841 pos++;
2842 }
2843 }
2844 }
cc601c31 2845 trie->lasttrans = pos + 1;
c944940b
JH
2846 trie->states = (reg_trie_state *)
2847 PerlMemShared_realloc( trie->states, laststate
2848 * sizeof(reg_trie_state) );
a3621e74 2849 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2850 PerlIO_printf( Perl_debug_log,
2851 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2852 (int)depth * 2 + 2,"",
2853 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2854 + 1 ),
2855 (IV)next_alloc,
2856 (IV)pos,
2857 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
2858 );
2859
2860 } /* end table compress */
2861 }
1e2e3d02 2862 DEBUG_TRIE_COMPILE_MORE_r(
538e84ed
KW
2863 PerlIO_printf(Perl_debug_log,
2864 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
1e2e3d02
YO
2865 (int)depth * 2 + 2, "",
2866 (UV)trie->statecount,
2867 (UV)trie->lasttrans)
2868 );
cc601c31 2869 /* resize the trans array to remove unused space */
c944940b
JH
2870 trie->trans = (reg_trie_trans *)
2871 PerlMemShared_realloc( trie->trans, trie->lasttrans
2872 * sizeof(reg_trie_trans) );
a3621e74 2873
538e84ed 2874 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2875 U8 nodetype =(U8)(flags & 0xFF);
2876 char *str=NULL;
538e84ed 2877
07be1b83 2878#ifdef DEBUGGING
e62cc96a 2879 regnode *optimize = NULL;
7122b237
YO
2880#ifdef RE_TRACK_PATTERN_OFFSETS
2881
b57a0404
JH
2882 U32 mjd_offset = 0;
2883 U32 mjd_nodelen = 0;
7122b237
YO
2884#endif /* RE_TRACK_PATTERN_OFFSETS */
2885#endif /* DEBUGGING */
a3621e74 2886 /*
3dab1dad
YO
2887 This means we convert either the first branch or the first Exact,
2888 depending on whether the thing following (in 'last') is a branch
2889 or not and whther first is the startbranch (ie is it a sub part of
2890 the alternation or is it the whole thing.)
3b753521 2891 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2892 the whole branch sequence, including the first.
a3621e74 2893 */
3dab1dad 2894 /* Find the node we are going to overwrite */
7f69552c 2895 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2896 /* branch sub-chain */
3dab1dad 2897 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2898#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2899 DEBUG_r({
2900 mjd_offset= Node_Offset((convert));
2901 mjd_nodelen= Node_Length((convert));
2902 });
7122b237 2903#endif
7f69552c 2904 /* whole branch chain */
7122b237
YO
2905 }
2906#ifdef RE_TRACK_PATTERN_OFFSETS
2907 else {
7f69552c
YO
2908 DEBUG_r({
2909 const regnode *nop = NEXTOPER( convert );
2910 mjd_offset= Node_Offset((nop));
2911 mjd_nodelen= Node_Length((nop));
2912 });
07be1b83
YO
2913 }
2914 DEBUG_OPTIMISE_r(
538e84ed
KW
2915 PerlIO_printf(Perl_debug_log,
2916 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
07be1b83 2917 (int)depth * 2 + 2, "",
786e8c11 2918 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2919 );
7122b237 2920#endif
538e84ed 2921 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
2922 split out as an EXACT and put in front of the TRIE node. */
2923 trie->startstate= 1;
55eed653 2924 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2925 U32 state;
1e2e3d02 2926 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2927 U32 ofs = 0;
8e11feef
RGS
2928 I32 idx = -1;
2929 U32 count = 0;
2930 const U32 base = trie->states[ state ].trans.base;
a3621e74 2931
3dab1dad 2932 if ( trie->states[state].wordnum )
8e11feef 2933 count = 1;
a3621e74 2934
8e11feef 2935 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2936 if ( ( base + ofs >= trie->uniquecharcount ) &&
2937 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2938 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2939 {
3dab1dad 2940 if ( ++count > 1 ) {
2b8b4781 2941 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2942 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2943 if ( state == 1 ) break;
3dab1dad
YO
2944 if ( count == 2 ) {
2945 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2946 DEBUG_OPTIMISE_r(
8e11feef
RGS
2947 PerlIO_printf(Perl_debug_log,
2948 "%*sNew Start State=%"UVuf" Class: [",
2949 (int)depth * 2 + 2, "",
786e8c11 2950 (UV)state));
be8e71aa 2951 if (idx >= 0) {
2b8b4781 2952 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2953 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2954
3dab1dad 2955 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2956 if ( folder )
2957 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2958 DEBUG_OPTIMISE_r(
f1f66076 2959 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2960 );
8e11feef
RGS
2961 }
2962 }
2963 TRIE_BITMAP_SET(trie,*ch);
2964 if ( folder )
2965 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2966 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2967 }
2968 idx = ofs;
2969 }
3dab1dad
YO
2970 }
2971 if ( count == 1 ) {
2b8b4781 2972 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2973 STRLEN len;
2974 char *ch = SvPV( *tmp, len );
de734bd5
A
2975 DEBUG_OPTIMISE_r({
2976 SV *sv=sv_newmortal();
8e11feef
RGS
2977 PerlIO_printf( Perl_debug_log,
2978 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2979 (int)depth * 2 + 2, "",
538e84ed
KW
2980 (UV)state, (UV)idx,
2981 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
2982 PL_colors[0], PL_colors[1],
2983 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 2984 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
2985 )
2986 );
2987 });
3dab1dad
YO
2988 if ( state==1 ) {
2989 OP( convert ) = nodetype;
2990 str=STRING(convert);
2991 STR_LEN(convert)=0;
2992 }
c490c714
YO
2993 STR_LEN(convert) += len;
2994 while (len--)
de734bd5 2995 *str++ = *ch++;
8e11feef 2996 } else {
538e84ed 2997#ifdef DEBUGGING
8e11feef
RGS
2998 if (state>1)
2999 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 3000#endif
8e11feef
RGS
3001 break;
3002 }
3003 }
2e64971a 3004 trie->prefixlen = (state-1);
3dab1dad 3005 if (str) {
8e11feef 3006 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3007 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3008 trie->startstate = state;
07be1b83
YO
3009 trie->minlen -= (state - 1);
3010 trie->maxlen -= (state - 1);
33809eae
JH
3011#ifdef DEBUGGING
3012 /* At least the UNICOS C compiler choked on this
3013 * being argument to DEBUG_r(), so let's just have
3014 * it right here. */
3015 if (
3016#ifdef PERL_EXT_RE_BUILD
3017 1
3018#else
3019 DEBUG_r_TEST
3020#endif
3021 ) {
3022 regnode *fix = convert;
3023 U32 word = trie->wordcount;
3024 mjd_nodelen++;
3025 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3026 while( ++fix < n ) {
3027 Set_Node_Offset_Length(fix, 0, 0);
3028 }
3029 while (word--) {
3030 SV ** const tmp = av_fetch( trie_words, word, 0 );
3031 if (tmp) {
3032 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3033 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3034 else
3035 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3036 }
3037 }
3038 }
3039#endif
8e11feef
RGS
3040 if (trie->maxlen) {
3041 convert = n;
3042 } else {
3dab1dad 3043 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3044 DEBUG_r(optimize= n);
3dab1dad
YO
3045 }
3046 }
3047 }
538e84ed
KW
3048 if (!jumper)
3049 jumper = last;
3dab1dad 3050 if ( trie->maxlen ) {
8e11feef
RGS
3051 NEXT_OFF( convert ) = (U16)(tail - convert);
3052 ARG_SET( convert, data_slot );
538e84ed
KW
3053 /* Store the offset to the first unabsorbed branch in
3054 jump[0], which is otherwise unused by the jump logic.
786e8c11 3055 We use this when dumping a trie and during optimisation. */
538e84ed 3056 if (trie->jump)
7f69552c 3057 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3058
6c48061a
YO
3059 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3060 * and there is a bitmap
3061 * and the first "jump target" node we found leaves enough room
3062 * then convert the TRIE node into a TRIEC node, with the bitmap
3063 * embedded inline in the opcode - this is hypothetically faster.
3064 */
3065 if ( !trie->states[trie->startstate].wordnum
3066 && trie->bitmap
3067 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
3068 {
3069 OP( convert ) = TRIEC;
3070 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 3071 PerlMemShared_free(trie->bitmap);
786e8c11 3072 trie->bitmap= NULL;
538e84ed 3073 } else
786e8c11 3074 OP( convert ) = TRIE;
a3621e74 3075
3dab1dad
YO
3076 /* store the type in the flags */
3077 convert->flags = nodetype;
a5ca303d 3078 DEBUG_r({
538e84ed
KW
3079 optimize = convert
3080 + NODE_STEP_REGNODE
a5ca303d
YO
3081 + regarglen[ OP( convert ) ];
3082 });
538e84ed 3083 /* XXX We really should free up the resource in trie now,
a5ca303d 3084 as we won't use them - (which resources?) dmq */
3dab1dad 3085 }
a3621e74 3086 /* needed for dumping*/
e62cc96a 3087 DEBUG_r(if (optimize) {
07be1b83 3088 regnode *opt = convert;
bcdf7404 3089
e62cc96a 3090 while ( ++opt < optimize) {
07be1b83
YO
3091 Set_Node_Offset_Length(opt,0,0);
3092 }
538e84ed
KW
3093 /*
3094 Try to clean up some of the debris left after the
786e8c11 3095 optimisation.
a3621e74 3096 */
786e8c11 3097 while( optimize < jumper ) {
07be1b83 3098 mjd_nodelen += Node_Length((optimize));
a3621e74 3099 OP( optimize ) = OPTIMIZED;
07be1b83 3100 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
3101 optimize++;
3102 }
07be1b83 3103 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
3104 });
3105 } /* end node insert */
2e64971a
DM
3106
3107 /* Finish populating the prev field of the wordinfo array. Walk back
3108 * from each accept state until we find another accept state, and if
3109 * so, point the first word's .prev field at the second word. If the
3110 * second already has a .prev field set, stop now. This will be the
3111 * case either if we've already processed that word's accept state,
3b753521
FN
3112 * or that state had multiple words, and the overspill words were
3113 * already linked up earlier.
2e64971a
DM
3114 */
3115 {
3116 U16 word;
3117 U32 state;
3118 U16 prev;
3119
3120 for (word=1; word <= trie->wordcount; word++) {
3121 prev = 0;
3122 if (trie->wordinfo[word].prev)
3123 continue;
3124 state = trie->wordinfo[word].accept;
3125 while (state) {
3126 state = prev_states[state];
3127 if (!state)
3128 break;
3129 prev = trie->states[state].wordnum;
3130 if (prev)
3131 break;
3132 }
3133 trie->wordinfo[word].prev = prev;
3134 }
3135 Safefree(prev_states);
3136 }
3137
3138
3139 /* and now dump out the compressed format */
3140 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3141
55eed653 3142 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
3143#ifdef DEBUGGING
3144 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3145 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3146#else
03e70be4 3147 SvREFCNT_dec_NN(revcharmap);
07be1b83 3148#endif
538e84ed
KW
3149 return trie->jump
3150 ? MADE_JUMP_TRIE
3151 : trie->startstate>1
3152 ? MADE_EXACT_TRIE
786e8c11
YO
3153 : MADE_TRIE;
3154}
3155
615a2e7f
YO
3156STATIC regnode *
3157S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
786e8c11 3158{
b423522f
KW
3159/* The Trie is constructed and compressed now so we can build a fail array if
3160 * it's needed
786e8c11 3161
b423522f
KW
3162 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3163 3.32 in the
3164 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3165 Ullman 1985/88
786e8c11
YO
3166 ISBN 0-201-10088-6
3167
b423522f
KW
3168 We find the fail state for each state in the trie, this state is the longest
3169 proper suffix of the current state's 'word' that is also a proper prefix of
3170 another word in our trie. State 1 represents the word '' and is thus the
3171 default fail state. This allows the DFA not to have to restart after its
3172 tried and failed a word at a given point, it simply continues as though it
3173 had been matching the other word in the first place.
786e8c11
YO
3174 Consider
3175 'abcdgu'=~/abcdefg|cdgu/
b423522f
KW
3176 When we get to 'd' we are still matching the first word, we would encounter
3177 'g' which would fail, which would bring us to the state representing 'd' in
3178 the second word where we would try 'g' and succeed, proceeding to match
3179 'cdgu'.
786e8c11
YO
3180 */
3181 /* add a fail transition */
3251b653
NC
3182 const U32 trie_offset = ARG(source);
3183 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
3184 U32 *q;
3185 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 3186 const U32 numstates = trie->statecount;
786e8c11
YO
3187 const U32 ubound = trie->lasttrans + ucharcount;
3188 U32 q_read = 0;
3189 U32 q_write = 0;
3190 U32 charid;
3191 U32 base = trie->states[ 1 ].trans.base;
3192 U32 *fail;
3193 reg_ac_data *aho;
cf78de0b 3194 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
615a2e7f 3195 regnode *stclass;
786e8c11 3196 GET_RE_DEBUG_FLAGS_DECL;
7918f24d 3197
615a2e7f 3198 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
81611534 3199 PERL_UNUSED_CONTEXT;
786e8c11
YO
3200#ifndef DEBUGGING
3201 PERL_UNUSED_ARG(depth);
3202#endif
3203
615a2e7f
YO
3204 if ( OP(source) == TRIE ) {
3205 struct regnode_1 *op = (struct regnode_1 *)
3206 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3207 StructCopy(source,op,struct regnode_1);
3208 stclass = (regnode *)op;
3209 } else {
3210 struct regnode_charclass *op = (struct regnode_charclass *)
3211 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3212 StructCopy(source,op,struct regnode_charclass);
3213 stclass = (regnode *)op;
3214 }
2f306ab9 3215 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
786e8c11
YO
3216
3217 ARG_SET( stclass, data_slot );
c944940b 3218 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 3219 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 3220 aho->trie=trie_offset;
446bd890
NC
3221 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3222 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 3223 Newxz( q, numstates, U32);
c944940b 3224 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
3225 aho->refcount = 1;
3226 fail = aho->fail;
3227 /* initialize fail[0..1] to be 1 so that we always have
3228 a valid final fail state */
3229 fail[ 0 ] = fail[ 1 ] = 1;
3230
3231 for ( charid = 0; charid < ucharcount ; charid++ ) {
3232 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3233 if ( newstate ) {
3234 q[ q_write ] = newstate;
3235 /* set to point at the root */
3236 fail[ q[ q_write++ ] ]=1;
3237 }
3238 }
3239 while ( q_read < q_write) {
3240 const U32 cur = q[ q_read++ % numstates ];
3241 base = trie->states[ cur ].trans.base;
3242
3243 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3244 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3245 if (ch_state) {
3246 U32 fail_state = cur;
3247 U32 fail_base;
3248 do {
3249 fail_state = fail[ fail_state ];
3250 fail_base = aho->states[ fail_state ].trans.base;
3251 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3252
3253 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3254 fail[ ch_state ] = fail_state;
3255 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3256 {
3257 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3258 }
3259 q[ q_write++ % numstates] = ch_state;
3260 }
3261 }
3262 }
3263 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3264 when we fail in state 1, this allows us to use the
3265 charclass scan to find a valid start char. This is based on the principle
3266 that theres a good chance the string being searched contains lots of stuff
3267 that cant be a start char.
3268 */
3269 fail[ 0 ] = fail[ 1 ] = 0;
3270 DEBUG_TRIE_COMPILE_r({
6d99fb9b 3271 PerlIO_printf(Perl_debug_log,
538e84ed 3272 "%*sStclass Failtable (%"UVuf" states): 0",
6d99fb9b 3273 (int)(depth * 2), "", (UV)numstates
1e2e3d02 3274 );
786e8c11
YO
3275 for( q_read=1; q_read<numstates; q_read++ ) {
3276 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3277 }
3278 PerlIO_printf(Perl_debug_log, "\n");
3279 });
3280 Safefree(q);
e384d5c1 3281 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
615a2e7f 3282 return stclass;
a3621e74
YO
3283}
3284
786e8c11 3285
07be1b83 3286#define DEBUG_PEEP(str,scan,depth) \
b515a41d 3287 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83 3288 regnode *Next = regnext(scan); \
8b9781c9 3289 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
fdfb4f21 3290 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
c9f0d54c
YO
3291 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3292 Next ? (REG_NODE_NUM(Next)) : 0 ); \
fdfb4f21
YO
3293 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3294 PerlIO_printf(Perl_debug_log, "\n"); \
b515a41d 3295 }});
07be1b83 3296
bb914485 3297/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
3298 * one. The regop may be changed if the node(s) contain certain sequences that
3299 * require special handling. The joining is only done if:
bb914485
KW
3300 * 1) there is room in the current conglomerated node to entirely contain the
3301 * next one.
3302 * 2) they are the exact same node type
3303 *
87b8b349 3304 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
3305 * these get optimized out
3306 *
0a982f06
KW
3307 * If a node is to match under /i (folded), the number of characters it matches
3308 * can be different than its character length if it contains a multi-character
31f05a37
KW
3309 * fold. *min_subtract is set to the total delta number of characters of the
3310 * input nodes.
bb914485 3311 *
fb29cc72 3312 * And *unfolded_multi_char is set to indicate whether or not the node contains
31f05a37
KW
3313 * an unfolded multi-char fold. This happens when whether the fold is valid or
3314 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3315 * SMALL LETTER SHARP S, as only if the target string being matched against
3316 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3317 * folding rules depend on the locale in force at runtime. (Multi-char folds
3318 * whose components are all above the Latin1 range are not run-time locale
3319 * dependent, and have already been folded by the time this function is
3320 * called.)
f758bddf 3321 *
bb914485 3322 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
3323 * multi-character fold sequences. It's been wrong in Perl for a very long
3324 * time. There are three code points in Unicode whose multi-character folds
3325 * were long ago discovered to mess things up. The previous designs for
3326 * dealing with these involved assigning a special node for them. This
538e84ed 3327 * approach doesn't always work, as evidenced by this example:
a0c4c608 3328 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
538e84ed 3329 * Both sides fold to "sss", but if the pattern is parsed to create a node that
0a982f06 3330 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
3331 * successful match would have to cross the node's boundary. The new approach
3332 * that hopefully generally solves the problem generates an EXACTFU_SS node
538e84ed 3333 * that is "sss" in this case.
bb914485 3334 *
0a982f06 3335 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
3336 * just these three. Now the code is general, for all such cases. The
3337 * approach taken is:
0a982f06 3338 * 1) This routine examines each EXACTFish node that could contain multi-
538e84ed
KW
3339 * character folded sequences. Since a single character can fold into
3340 * such a sequence, the minimum match length for this node is less than
3341 * the number of characters in the node. This routine returns in
31f05a37
KW
3342 * *min_subtract how many characters to subtract from the the actual
3343 * length of the string to get a real minimum match length; it is 0 if
3344 * there are no multi-char foldeds. This delta is used by the caller to
3345 * adjust the min length of the match, and the delta between min and max,
3346 * so that the optimizer doesn't reject these possibilities based on size
3347 * constraints.
cb117658 3348 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
3349 * is used for an EXACTFU node that contains at least one "ss" sequence in
3350 * it. For non-UTF-8 patterns and strings, this is the only case where
3351 * there is a possible fold length change. That means that a regular
3352 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3353 * with length changes, and so can be processed faster. regexec.c takes
3354 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
31f05a37
KW
3355 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3356 * known until runtime). This saves effort in regex matching. However,
3357 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3358 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3359 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3360 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
0a982f06
KW
3361 * possibilities for the non-UTF8 patterns are quite simple, except for
3362 * the sharp s. All the ones that don't involve a UTF-8 target string are
3363 * members of a fold-pair, and arrays are set up for all of them so that
3364 * the other member of the pair can be found quickly. Code elsewhere in
3365 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3366 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3367 * described in the next item.
31f05a37
KW
3368 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3369 * validity of the fold won't be known until runtime, and so must remain
3370 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3371 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3372 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3373 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3374 * The reason this is a problem is that the optimizer part of regexec.c
3375 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3376 * that a character in the pattern corresponds to at most a single
3377 * character in the target string. (And I do mean character, and not byte
3378 * here, unlike other parts of the documentation that have never been
3379 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3380 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3381 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3382 * nodes, violate the assumption, and they are the only instances where it
3383 * is violated. I'm reluctant to try to change the assumption, as the
3384 * code involved is impenetrable to me (khw), so instead the code here
3385 * punts. This routine examines EXACTFL nodes, and (when the pattern
3386 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3387 * boolean indicating whether or not the node contains such a fold. When
3388 * it is true, the caller sets a flag that later causes the optimizer in
3389 * this file to not set values for the floating and fixed string lengths,
3390 * and thus avoids the optimizer code in regexec.c that makes the invalid
1ca267a5 3391 * assumption. Thus, there is no optimization based on string lengths for
31f05a37
KW
3392 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3393 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3394 * assumption is wrong only in these cases is that all other non-UTF-8
3395 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3396 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3397 * EXACTF nodes because we don't know at compile time if it actually
3398 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3399 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3400 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3401 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3402 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3403 * string would require the pattern to be forced into UTF-8, the overhead
3404 * of which we want to avoid. Similarly the unfolded multi-char folds in
3405 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3406 * locale.)
098b07d5
KW
3407 *
3408 * Similarly, the code that generates tries doesn't currently handle
3409 * not-already-folded multi-char folds, and it looks like a pain to change
3410 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3411 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3412 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3413 * using /iaa matching will be doing so almost entirely with ASCII
3414 * strings, so this should rarely be encountered in practice */
1de06328 3415
fb29cc72 3416#define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
07be1b83 3417 if (PL_regkind[OP(scan)] == EXACT) \
fb29cc72 3418 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
07be1b83 3419
be8e71aa 3420STATIC U32
538e84ed 3421S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
fb29cc72 3422 UV *min_subtract, bool *unfolded_multi_char,
538e84ed
KW
3423 U32 flags,regnode *val, U32 depth)
3424{
07be1b83
YO
3425 /* Merge several consecutive EXACTish nodes into one. */
3426 regnode *n = regnext(scan);
3427 U32 stringok = 1;
3428 regnode *next = scan + NODE_SZ_STR(scan);
3429 U32 merged = 0;
3430 U32 stopnow = 0;
3431#ifdef DEBUGGING
3432 regnode *stop = scan;
72f13be8 3433 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 3434#else
d47053eb
RGS
3435 PERL_UNUSED_ARG(depth);
3436#endif
7918f24d
NC
3437
3438 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 3439#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
3440 PERL_UNUSED_ARG(flags);
3441 PERL_UNUSED_ARG(val);
07be1b83 3442#endif
07be1b83 3443 DEBUG_PEEP("join",scan,depth);
bb914485 3444
3f410cf6
KW
3445 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3446 * EXACT ones that are mergeable to the current one. */
3447 while (n
3448 && (PL_regkind[OP(n)] == NOTHING
3449 || (stringok && OP(n) == OP(scan)))
07be1b83 3450 && NEXT_OFF(n)
3f410cf6
KW
3451 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3452 {
538e84ed 3453
07be1b83
YO
3454 if (OP(n) == TAIL || n > next)
3455 stringok = 0;
3456 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
3457 DEBUG_PEEP("skip:",n,depth);
3458 NEXT_OFF(scan) += NEXT_OFF(n);
3459 next = n + NODE_STEP_REGNODE;
3460#ifdef DEBUGGING
3461 if (stringok)
3462 stop = n;
3463#endif
3464 n = regnext(n);
3465 }
3466 else if (stringok) {
786e8c11 3467 const unsigned int oldl = STR_LEN(scan);
07be1b83 3468 regnode * const nnext = regnext(n);
b2230d39 3469
baa60164
KW
3470 /* XXX I (khw) kind of doubt that this works on platforms (should